technologies

Article
A Collaborative Federated Learning Framework for Lung and
Colon Cancer Classifications
Md. Munawar Hossain 1,* , Md. Robiul Islam 1, Md. Faysal Ahamed 1 , Mominul Ahsan 2 and Julfikar Haider 3,*

1 Department of Electrical & Computer Engineering, Rajshahi University of Engineering & Technology,
Rajshahi 6204, Bangladesh; [email protected] (M.R.I.); [email protected] (M.F.A.)
2 Department of Computer Science, University of York, Deramore Lane, Heslington YO10 5GH, UK;
[email protected]
3 Department of Engineering, Manchester Metropolitan University, Chester Street, Manchester M1 5GD, UK
* Correspondence: [email protected] (M.M.H.); [email protected] (J.H.)

Abstract: Lung and colon cancers are common types of cancer with significant fatality rates. Early
identification considerably improves the odds of survival for those suffering from these diseases.
Histopathological image analysis is crucial for detecting cancer by identifying morphological anoma-
lies in tissue samples. Regulations such as the HIPAA and GDPR impose considerable restrictions
on the sharing of sensitive patient data, mostly because of privacy concerns. Federated learning
(FL) is a promising technique that allows the training of strong models while maintaining data
privacy. The use of a federated learning strategy has been suggested in this study to address privacy
concerns in cancer categorization. To classify histopathological images of lung and colon cancers, this
methodology uses local models with an Inception-V3 backbone. The global model is then updated
on the basis of the local weights. The images were obtained from the LC25000 dataset, which consists
of five separate classes. Separate analyses were performed for lung cancer, colon cancer, and their
combined classification. The implemented model successfully classified lung cancer images into three
separate classes with a classification accuracy of 99.867%. The classification of colon cancer images
was achieved with 100% accuracy. More significantly, for the lung and colon cancers combined, the
accuracy reached an impressive 99.720%. Compared with other current approaches, the proposed
framework showed an improved performance. A heatmap, visual saliency map, and GradCAM were
Citation: Hossain, M.M.; Islam, M.R.; generated to pinpoint the crucial areas in the histopathology pictures of the test set where the models
Ahamed, M.F.; Ahsan, M.; Haider, J. A focused in particular during cancer class predictions. This approach demonstrates the potential of
Collaborative Federated Learning federated learning to enhance collaborative efforts in automated disease diagnosis through medical
Framework for Lung and Colon image analysis while ensuring patient data privacy.
Cancer Classifications. Technologies
2024, 12, 151. https://doi.org/ Keywords: lung cancer; colon cancer; histopathological image analysis; image classification; decentralized
10.3390/technologies12090151 machine learning; federated learning; privacy preservation; explainability
Academic Editor: Sikha Bagui

Received: 30 July 2024

Revised: 27 August 2024
1. Introduction
Accepted: 2 September 2024
Published: 4 September 2024
Cancer ranks as the second most prevalent cause of mortality globally, following
cardiovascular ailments. In 2022, there were 9.74 million cancer-related deaths and almost
20 million new cases reported globally [1]. Lung cancer is responsible for the highest
mortality rate and the second highest number of incidents worldwide. On the basis of
Copyright: © 2024 by the authors. the latest data from the World Health Organization (WHO), 2.20 million new cases and
Licensee MDPI, Basel, Switzerland. 1.79 million deaths were reported in 2020. Figure 1 shows that lung cancer has the highest
This article is an open access article number of cases, accounting for 12.4%, narrowly surpassing breast cancer at 11.6% [2].
distributed under the terms and Nevertheless, it is clear that lung cancer is the leading cause of death among all cancer-
conditions of the Creative Commons related deaths, with a mortality rate of 18%. There is no other cancer that even comes close
Attribution (CC BY) license (https:// to having this enormous number of fatalities.
creativecommons.org/licenses/by/
4.0/).

Technologies 2024, 12, 151. https://doi.org/10.3390/technologies12090151 https://www.mdpi.com/journal/technologies

 Technologies 2024, 12, x FOR PEER REVIEW 2 of 31
Technologies 2024, 12, 151 2 of 28

(a) (b)
Figure1. 1.
Figure Total
Total number
number andand proportion
proportion of cancer
of cancer (a) incidences
(a) incidences and
and (b) (b) mortalities
mortalities in 2022in 2022taken
(data (data
taken from [1]).
from [1]).

Onthe
On thebasis
basisofofthetheanalysis,
analysis,ititcan canbebeconcluded
concludedthat thatcolorectal
colorectalcancercancerranks
ranksasasthe the
secondleading
second leadingcause causeofofmortality
mortalityand andthe thethird
thirdhighest
highestinintermstermsofofincidence
incidenceglobally.
globally.InIn
2022,it it
2022, ranked
ranked second
second in terms
in terms of newof newcases,cases, accounting
accounting for 9.6% for of
9.6% of alljust
all cases, cases,
afterjust after
breast
breast cancer. Additionally, it constituted 9.3% of the total
cancer. Additionally, it constituted 9.3% of the total number of deaths, ranking second number of deaths, ranking sec-
ond only to lung cancer [1]. The global impact of this phenomenon
only to lung cancer [1]. The global impact of this phenomenon resulted in approximately resulted in approxi-
mately
1.93 million1.93newmillioncasesnewand cases and 935,000
935,000 deaths, deaths, representing
representing approximately
approximately 10% of10% of all
all new
new cancer
cancer cases and cases and fatalities
fatalities worldwide.worldwide. Withprevalence
With these these prevalence rates sharply
rates rising rising sharply
annually,an-
itnually, it is projected
is projected that by
that by 2040, there2040,
willthere
be 3.2will be 3.2new
million million new instances
instances of colorectal
of colorectal cancer
cancerposing
(CRC), (CRC),aposingseriousathreat
serious to threat
global to global
public public
health [3].health
Hence, [3].itHence,
is crucialit is
tocrucial
employto
employ expeditious
expeditious and efficientandprotocols
efficient protocols
for diagnosticfor diagnostic decisiontomaking
decision making formulate to formulate
a tailored a
tailored treatment
treatment strategy that strategy that maximizes
maximizes patientrates
patient survival survival ratesindividual
in every in every individual
case. case.
Multiple
Multiplefactors factorscontribute
contributetotothe thedevelopment
developmentofofcancer.cancer.TheseTheseinclude
includeexposure
exposuretoto
physical
physicalcarcinogens,
carcinogens,such suchasasradiation
radiationand andultraviolet
ultravioletrays,rays,and andcertain
certainbehaviors,
behaviors,such such
asashaving
havinga ahigh highbody
bodymassmassindexindexand andusing
usingalcohol
alcoholand andtobacco,
tobacco,ininaddition
additiontotospecific
specific
biological
biologicaland andgenetic
genetic factors [4]. [4]. While
Whilesymptoms
symptoms maymay varyvary
amongamong individuals
individuals andandeven
even different types of cancers, none of these signs are exclusive
different types of cancers, none of these signs are exclusive to cancer, and not all patients to cancer, and not all
patients will encounter them. In light of this, cancer detection
will encounter them. In light of this, cancer detection may be difficult in the absence of may be difficult in the
absence of specialized
specialized diagnostic diagnostic
tools such tools
as such as ultrasound,
ultrasound, positron positron
emission emission tomography
tomography (PET),
(PET), computed
computed tomography
tomography (CT), or
(CT), MRI, MRI, or biopsy.
biopsy. Prompt Prompt identification
identification is crucial
is crucial for theforde-
the detection
tection of both of lung
both and
lungcolonand colon
cancers.cancers.
In the In theoffield
field of clinical
clinical medicine, medicine,
symptomssymptoms
of these
ofparticular
these particular
types oftypescancersof cancers
typically typically
manifest manifest
during during the advanced
the advanced stages ofstages of the
the disease.
disease. Physicians face difficulties in the early diagnosis of lung
Physicians face difficulties in the early diagnosis of lung cancer through exclusive reliance cancer through exclusive
reliance
on visual on assessment
visual assessment of CT images.
of CT images. The application
The application of computer-aided
of computer-aided diagnosisdiagnosis
(CAD)
(CAD)
in the examination of histopathological images continues to be a prominent areaarea
in the examination of histopathological images continues to be a prominent of
of em-
emphasis
phasis inin the
the field
field ofof cancerdetection.
cancer detection.
Privacy
Privacy is a significant concernfor
is a significant concern forhealthcare
healthcarefacilities.
facilities.As Asa aresult,
result,regulations
regulationssuch such
asasHIPAA
HIPAA [5] and GDPRs [6] pose significant challenges for institutionswith
[5] and GDPRs [6] pose significant challenges for institutions withrespect
respecttoto
disclosing
disclosingpatient
patientdata,
data,including
includinganonymous
anonymousinformation.
information.To Toovercome
overcomethese thesedifficulties,
difficulties,
the
the adoption of a decentralized method for machine learning called federatedlearning
adoption of a decentralized method for machine learning called federated learninghas has
been suggested. Currently, several deep learning techniques have shown excellent accuracy
been suggested. Currently, several deep learning techniques have shown excellent accu-
when predicting the lung cell class. Nevertheless, none of these entities have adopted a
racy when predicting the lung cell class. Nevertheless, none of these entities have adopted
thorough strategy to address data privacy regulations.
a thorough strategy to address data privacy regulations.
Technologies 2024, 12, 151 3 of 28

It is challenging for clinicians to diagnose cancer at an early stage by relying merely

on a visual assessment of CT images, as evidenced by the fact that symptoms of cancer
often manifest in the advanced stages of the disease. By pooling information, healthcare
providers can better pinpoint areas for improvement and perhaps diagnose lung cancer
at an earlier, more treatable stage. By sharing data, organizations can increase efficiency
and maximize resources, leading to greater productivity and cost savings. Medical data
sharing is a powerful tool that can drive significant advancements in the field of medicine.
Hospitals in the healthcare industry or individual devices are examples of the many parties
or devices used across federated learning methodologies in healthcare.
Figure 2 illustrates a federated learning model where multiple healthcare providers
collaborate to train a shared machine learning model without sharing patient data directly.
Each healthcare provider trains its local model using its own data, and only the model
updates are sent to a central server. The central server aggregates these updates to create
a global model, which is then shared back with all the healthcare providers, who then
continue to refine the global model by performing further training with their local data,
and this iterative process continues until the model is fully trained. By collaborating and
sharing data, organizations can unlock the potential for groundbreaking advancements in
our current understanding of diseases and treatments.
This study aims to design and implement a distributed federated learning (FL) archi-
tecture that allows for adaptive thresholding and the customization of local neural networks
(NNs) in the context of medical image classification, with a specific focus on lung and
colon cancer diagnoses. The goal is to harness the power of federated learning to facilitate
collaboration among healthcare organizations while maintaining a decentralized approach.
Through this decentralized approach, each institution can train a local neural network on
its own data while contributing to a shared global model that benefits from the collective
knowledge of all participants. By doing so, individual institutions can contribute to a
shared global model without directly sharing sensitive patient data, ensuring compliance
with stringent healthcare privacy and security regulations. This study not only seeks to
increase the accuracy of cancer diagnoses but also aims to create a scalable and secure
solution that can be applied to a wide range of medical imaging tasks. To summarize, this
study makes the following significant contributions:
(1) A robust federated learning architecture was designed and developed, particularly for
medical image classification, which is demonstrated through lung and colon cancer
classifications. The framework seamlessly consolidated data from many healthcare
organizations while upholding data privacy and security regulations.
(2) The federated learning workflow was streamlined for smooth global model updates
after each communication round, with local model weights adjusted to align with the
global model. A comprehensive evaluation process was also applied, assessing each
client’s model performance after every training epoch, enhancing transparency and
identifying performance variations or underperforming clients.
(3) Explainable AI techniques were integrated to provide visual and quantitative insights
into the model’s decision-making process and provide further interpretability.
(4) The performance of the proposed federated learning (FL) model is evaluated against
well-known transfer learning (TL) models and other current state-of-the-art
(SOTA) approaches.
The main novelty of this work lies in comparing the performance of federated and
centralized machine learning models for the LC-25000 dataset, which is specifically applied
to lung and colon cancer classifications. This research is the first to demonstrate the
efficiency of a unique combination of federated learning in this context, highlighting the
collaborative nature of the framework, introducing technical innovations, and emphasizing
the potential impact on clinical practices.
The subsequent sections of the article are structured as follows: Section 2 provides a
concise overview of prior research that is similar to this study. Section 3 presents a detailed
representation of the dataset used in this study and the methodology proposed for the
Technologies 2024, 12, x FOR PEER REVIEW 4 of 31
Technologies 2024, 12, 151 4 of 28

representation of the dataset used in this study and the methodology proposed for the
federated learning
federated learningmodel.
model. TheThe experimental
experimental conditions
conditions underunder
which which the proposed
the proposed method
method
is is compared
compared are presented
are presented in4.
in Section Section 4. 5Section
Section 5 provides
provides a comprehensive
a comprehensive analysisanal-
and
evaluation of the experiment’s
ysis and evaluation performance,
of the experiment’s highlighting
performance, its findings.
highlighting itsSection 6 concludes
findings. Section 6
the proposed
concludes theresearch
proposedand provides
research recommendations
and for further enhancement.
provides recommendations Finally,
for further enhance-
the article concludes by providing a concise overview of the results.
ment. Finally, the article concludes by providing a concise overview of the results.

Figure 2.
Figure 2. Federated learning
learning conceptual
conceptual framework
framework applicable
applicable in
in healthcare
healthcare sector.
sector.

2.
2. Literature
Literature Survey
Survey
2.1. Lung and Colon Cancer Diagnoses
2.1. Lung and Colon Cancer Diagnoses
Lung and colon cancers are prevalent and highly fatal diseases that have a significant
Lung and colon cancers are prevalent and highly fatal diseases that have a significant
global impact, affecting many individuals annually. Despite their organ-specific origins,
global impact, affecting many individuals annually. Despite their organ-specific origins,
there are several shared features and important distinctions between the two in terms
there are several shared features and important distinctions between the two in terms of
of prognosis, diagnostic criteria, and therapeutic options. Numerous research groups
prognosis, diagnostic criteria, and therapeutic options. Numerous research groups have
have achieved substantial advancements in the detection of lung and colon cancers in
achieved substantial advancements in the detection of lung and colon cancers in recent
recent years. These developments include the application of deep learning methods on the
years.ofThese
basis developmentsimage
histopathological include the application
analysis. The worksof deep learning methods
are organized into threeon the basis
categories
of histopathological
for image(adenocarcinoma,
classifying lung cancer analysis. The works are organized
squamous into three
cell carcinoma, andcategories for
benign), two
Technologies 2024, 12, 151 5 of 28

categories for classifying colon cancer subtypes (adenocarcinoma and benign), and five
categories for classifying both lung and colon cancer categories.
In a previous study [7], a deep learning-based supervised learning technique was
developed to classify lung and colon cancer tissues into five distinct categories. The ap-
proach implemented utilized two methods for feature extraction: 2D Fourier features and
2D wavelet features. The final accuracy of the work was 96.33%. Another study [8] utilized
feature extraction from histopathology images and various machine learning classifiers,
such as random forest (RF) and XGBoost, to classify lung and colon cancers. The study
achieved impressive accuracies of 99%. A CNN pretrained diagnostic network was specifi-
cally designed for the detection of lung and colon cancers [9]. The network demonstrated
a high level of accuracy in diagnosing colon and lung cancers, achieving accuracies of
96% and 97%, respectively. Convolutional neural networks (CNNs) using the VGG16
model and contrast-limited adaptive histogram equalization (CLAHE) were used by other
researchers [10] to classify 25,000 histopathology images. Transformers have advanced
medical image analysis but struggle with feature capture, information loss, and segmen-
tation accuracy; CASTformer addresses these issues with multi-scale representations, a
class-aware transformer module, and adversarial training [11]. Furthermore, incremental
transfer learning (ITL) offers an efficient solution for multi-site medical image segmentation
by sequentially training a model across datasets, mitigating catastrophic forgetting, and
achieving superior generalization to new domains with minimal computational resources
and domain-specific expertise [12].
One study [13] discussed the use of histogram equalization as a preprocessing step,
followed by the application of pretrained AlexNet, to improve the classification of lung
and colon cancers. Toğaçar et al. [14] utilized a pretrained DarkNet-19 in conjunction with
support vector machine classifiers to attain a 99.69% accuracy rate in their study. Using
DenseNet-121 and RF classifiers, Kumar et al. [15] achieved a 98.6% accuracy rate in their
classification. Another study utilized feature extraction and ensemble learning techniques,
along with the incorporation of high-performance filtering, to attain an impressive accuracy
rate of 99.3% when using LC25000 data [16]. The use of artificial neural networks (ANNs)
with merged features from the VGG-19 and GoogLeNet models was covered in [17]. The
ANN achieved an accuracy of 99.64% when the fusion features of VGG-19 and the hand-
crafted features were combined. In a separate study, the authors employed a convolutional
neural network (CNN) with a SoftMax classifier, which they named AdenoCanNet [18].
The accuracy of the entire LC25000 dataset was 99.00%.
In addition to the previously discussed methods, recent studies have made significant
advancements in learning-based methods for medical image segmentation. Contrastive
learning and distillation techniques have shown promise in addressing the challenges of
limited labeled data and segmentation accuracy in medical image analysis, with methods
like contrastive voxel-wise representation learning (CVRL) [19] and SimCVD [20] advanc-
ing state-of-the-art voxel-wise representation learning by capturing 3D spatial context,
leveraging bi-level contrastive optimization, and utilizing simple dropout-based augmenta-
tion to achieve competitive performance even with less labeled data. Additionally, ACTION
(Anatomical-aware ConTrastive dIstillatiON) [21] tackles multi-class label imbalance by
using soft negative labeling and anatomical contrast, improving segmentation accuracy and
outperforming state-of-the-art semi-supervised methods on benchmark datasets. Finally,
ARCO enhances semi-supervised medical image segmentation by introducing stratified
group theory and variance-reduction techniques, addressing tail-class misclassification and
model collapse, and demonstrating superior performance across eight benchmarks [22].
When the disadvantages of the current models used in lung and colon cancer classifi-
cations are analyzed, several challenges and limitations can be identified:
(1) Data Privacy Concerns: Many existing models require centralized data collection,
where medical images from different institutions are pooled together in a single
repository. This raises serious privacy concerns, especially in healthcare, where patient
Technologies 2024, 12, 151 6 of 28

data are highly sensitive. Centralized models can be susceptible to data breaches and
may not comply with regulations such as HIPAA or GDPR.
(2) Limited Generalization: Centralized models are often trained on data from a limited
number of sources or geographic locations, which can result in poor generalizability
to other patient populations. This lack of diversity in the training data can lead to
biases and reduced effectiveness when applied to new datasets, limiting the model’s
ability to handle variations in medical imaging from different institutions or regions.
(3) Computational Requirements: Modern models for cancer classification, such as
deep convolutional neural networks (CNNs), demand significant computational
resources. This can be a barrier for smaller institutions with limited access to high-
performance computing infrastructure. Moreover, training large-scale models can be
time-consuming and energy-intensive.
(4) Imbalance in Class Distribution: Medical datasets, including lung and colon cancer
imaging datasets, often suffer from class imbalance, where the number of images of
cancerous tissues is much lower than that of non-cancerous ones. This imbalance can
bias the model, making it more likely to misclassify cancer cases, which is especially
problematic in clinical settings where false negatives can be life-threatening. Work
reported by You et al. [23] introduced adaptive anatomical contrast with a dynamic
contrastive loss, which better handles class imbalances in long-tail distributions.
(5) Difficulty in Handling Heterogeneous Data: Medical imaging data can be highly
heterogeneous due to differences in imaging equipment, protocols, and settings across
institutions. Current models may struggle to handle this heterogeneity, leading to
reduced performance when applied to data from sources other than the training data.
2.2. Federated Learning Applications
The integration of massive amounts of data can benefit machine learning models,
as stated previously. Access to data in the medical field is highly limited because of
the strict considerations of user privacy and data security. In this context, decentralized
collaborative machine learning algorithms that protect privacy are appropriate for creating
intelligent medical diagnostic systems. The notion of federated learning, which was initially
introduced by Google in 2016 [24,25], has since been expanded to encompass scenarios
involving knowledge integration and collaborative learning between organizations.
A client server-based method called federated averaging (FedAvg) was used for breast
density classification in [26]. This method incorporates local stochastic gradient descent
(SGD) on each client with a server that performs model averaging. In their publication [27],
the authors proposed a federated learning approach utilizing pretrained deep learning
models for the purpose of COVID-19 detection. The clients aimed to collaborate to achieve
a global model without the need to share individual samples from the dataset. Another
federated learning framework [28] for lung cancer classification utilizing histopathological
images demonstrated 99.867% accuracy, while imposing significant limitations on data
sharing between institutions. Zhang et al. [29] introduced a dynamic fusion-based approach
for COVID-19 detection. An image fusion technique was employed to diagnose patients
with COVID-19 on the basis of their medical data. The evaluation parameters yielded
favorable outcomes. However, the lack of consideration for patient data privacy was a
significant oversight in the proposed medical image analysis.
In the healthcare industry 5.0 domain, researchers have proposed that the Google
net deep machine learning model is utilized for precise disease prediction in the smart
healthcare industry 5.0 [30]. The proposed methodology for secure IoMT-based transfer
learning achieved a 98.8% accuracy rate, surpassing previous state-of-the-art methodologies
used in cancer disease prediction within the smart healthcare industry 5.0 on the LC25000
dataset. In a parallel investigation of Society 5.0 [31], researchers presented data as a service
(DaaS) along with a suggested framework that uses the blockchain network to provide
safe and decentralized transmission and distribution of data and machine learning systems
on the cloud. The main contributions and shortcomings of previous federated learning
research can be found in Table 1.
Technologies 2024, 12, 151 7 of 28

Table 1. Summary of the main contributions and shortcomings of previous FL research.

Previous Study Main Contribution of the Research Limitations of the Work

Dynamic fusion-based approach for CT scan image analysis Concerns regarding the appropriateness of controls for
Zhang et al. [29]
to diagnose COVID-19. patient data privacy and authenticity.
In a real-world collaborative setting, the author employed FL The proposed model is overly simplistic and requires
Roth et al. [26]
to develop medical imaging classification models. additional simulations.
Focusing more on the application of IoMT devices,
Khan et al. [30] Proposed a secure IoMT-based transfer learning methodology intended for industry 5.0 application.
There is chance of data corruption through IoMT devices.
A federated learning (FL) system was implemented for
Doubts about the suitability of safeguards for maintaining
Florescu et al. [27] COVID-19 detection using CT images, with clients deployed
the confidentiality and integrity of patient data.
locally on a single machine.
Proposed blockchain-based DCIaaS framework enhances
data and computational intelligence quality, equality, and Potential complexity and computational overhead
Peyvandi et al. [31] privacy for machine learning, demonstrating improved introduced by using blockchain technology, which could
accuracy in biomedical image classification and hazardous affect the efficiency and scalability of the system.
litter management.

The major gaps in the literature concerning lung and colon cancer classifications that
inspired the current study framework are briefly summarized below.
• There is a noticeable absence of sufficient measures to guarantee the privacy and
security of patient data.
• There are instances where the computational cost becomes considerably higher owing
to the substantial increase in the data scale, making it challenging to maintain efficiency
and performance.

3. Data and Methodology

This section provides insights into the dataset and various approaches to implement-
ing federated learning. Figure 3 illustrates the overall study path. Initially, the preprocessed
dataset was divided into training, testing, and validation sets. These datasets are dis-
tributed across multiple healthcare institutions, each training their local neural networks
independently using their own data, without sharing any sensitive patient information.
Each institution trains its local neural network independently, ensuring that patient data
remain private. The locally trained models are then aggregated in a central server to
update a global model, which is shared back with the institutions for further refinement.
Explainable AI techniques are applied to enhance model interpretability by visualizing
the features that drive predictions. The process culminates in performance evaluation,
ensuring accuracy and transparency in predictions, while fostering collaboration across
healthcare organizations.
The data that are subsequently distributed among clients via an independent and
identically distributed (IID) approach. Local models are first developed through training
on the data, and then the clients send the model parameters to the server. After training the
local models, the results are aggregated in a secure, centralized server to update a global
model, which represents the combined knowledge of all institutions. This process involves
training data on individual client devices and subsequently merging the local models on
a central server. The workflow for federated learning via Inception-V3 is illustrated in
Figure 4.

3.1. Dataset, Preprocessing, and Splitting

The dataset LC25000, published in 2020 by A. Borkowski and colleagues [32], was
utilized in this study. The collection contains images of lung and colon tissues, which are
categorized into five distinct classes. There are three distinct types of lung tissue images:
adenocarcinoma, squamous cell carcinoma, and benign. Some sample images of the classes
can be seen in Figure 5. The production of this content was made possible through the
Technologies 2024, 12, 151 8 of 28

Technologies 2024, 12, x FOR PEER REVIEW

provisionof resources and utilization of facilities at James A. Haley Veterans’ Hospital.8Itofis31
collected from patients through keen observation by physiologists.

Figure
Figure 3. 3.Diagram
Diagramillustrating
illustratingthe
theentire
entirepath
pathofofthe
thefederated
federatedlearning
learningstudy,
study, including
including steps
steps from
from
data collection to model evaluation and explainable AI
data collection to model evaluation and explainable AI integration.integration.

TheLC25000
The data that are subsequently
dataset distributed
consisted primarily among
of 1250 clientsslide
pathology via images
an independent
of lung and and
identically
colon tissues.distributed
Borkowski (IID) approach.
et al. [32] usedLocal models are library
an Augmentor first developed
to applythrough training
preprocessing
on the data,
techniques andimages
to the then the clients
and send the size
increased model parameters
of our dataset to athe server.
total Afterimages.
of 25,000 training
This
thewas achieved
local models,through the implementation
the results are aggregatedofinvarious augmentations,
a secure, including
centralized server left and a
to update
right rotations
global model,with
which a maximum
representsangle of 25 degrees
the combined and a probability
knowledge of 1.0. Additionally,
of all institutions. This process
horizontal
involves training data on individual client devices and subsequently merging the the
and vertical flips were applied with a probability of 0.5. Consequently, local
dataset
modelswas on aexpanded to a total
central server. of 25,000 images,
The workflow whichlearning
for federated were further categorizedisinto
via Inception-V3 illus-
five distinct
trated categories.
in Figure 4. Each category contained 5000 images. The images were resized to
dimensions of 768 × 768 prior to the application of augmentation techniques. To guarantee
Technologies 2024, 12, 151 9 of 28

privacy and unrestricted usage, these images underwent validation and adhered to the
regulations set forth by the Health Insurance Portability and Accountability Act (HIPAA).
Table 2 displays the designated names and IDs assigned to each class of images within
the dataset and an overview of the dataset split. To reduce computational complexity, we
downsized the images in our training and test directories from our pre-existing dataset,
which had 100 × 100 pixels. The utilization of the training and test directories is justified
by the fact that the test directory’s images are utilized to test the global model, whereas the
training directory’s images are disseminated to end devices/clients for local data training.
Technologies 2024, 12, x FOR PEER REVIEW 9 of 31
The dataset containing 25,000 lung and colon cancer images is organized into training,
testing, and validation sets at an 80:10:10 ratio.

Federated
Learning

Split the data for each

local client

Set model parameters and

distribute the model on each local
client.

Train Model
Locally

Send
Send Parameters
Parameters
Back to the
Back to the Server
Server Communication
Rounds

Server Creates Improvise

Improved Global Model
Updates to Local Through Local
Clients Model Updates

After 50 communication
rounds, final evaluation is
performed

Figure
Figure 4.
4. Federated
Federated learning
learning with
with Inception-V3
Inception-V3 methodology.
methodology.

Table
3.1. 2. Description
Dataset, of the employed
Preprocessing, dataset.
and Splitting
The dataset
Image LC25000, publishedFolder in 2020 Total
by A. Borkowski
TrainingandTesting
colleagues [32], was
Validation
utilized in this
Type study. The collection
Title contains images
Images of lung and
Set colon tissues,
Set which
Set are
categorized into five distinct classes.
Lung Adenocarcinoma lung_aca
There are5000
three distinct types of 500
4000
lung tissue 500
images:
adenocarcinoma,
Lung Benign squamous cell carcinoma, and
lung_bnt benign. Some
5000 4000 sample500
images of the
500 clas-
ses canLung
be seen in Figure 5. The production of this content was made possible through the
Squamous
lung_scc 5000 4000 500 500
Cell of
provision Carcinoma
resources and utilization of facilities at James A. Haley Veterans’ Hospital. It
Colon Adenocarcinoma
is collected from patients throughcolon_aca 5000 by physiologists.
keen observation 4000 500 500
Colon Benign colon_bnt 5000 4000 500 500
Technologies 2024,12,
Technologies2024, 12,151
x FOR PEER REVIEW 10 of
10 of 28
31

Figure 5.
Figure 5. Histopathological images from LC25000 dataset where (a) lung adenocarcinoma; (b) lung
benign; (c)
benign; (c) lung
lung squamous
squamous cell
cell carcinoma;
carcinoma; (d)
(d) colon
colon adenocarcinoma;
adenocarcinoma;and
and(e)
(e)colon
colonbenign.
benign.

3.2. Description
The LC25000 of the Classes
dataset consisted primarily of 1250 pathology slide images of lung and
3.2.1. Lung Adenocarcinoma
colon tissues. Borkowski et al. [32] used an Augmentor library to apply preprocessing
techniques to the images and
Lung adenocarcinoma increased
represents thethe size of our
prevailing formdataset to a total
of primary lungofcancer
25,000observed
images.
This was
inside theachieved throughThis
United States. the particular
implementation of various
condition augmentations,
is classified within theincluding
categoryleft
of
and right rotations
non-small with a (NSCLC)
cell lung cancer maximumand angle of 25 degrees
is closely linked toand a probability
a history of 1.0.
of tobacco Addi-
smoking.
tionally, horizontal
Although and vertical
there has been a decreaseflipsinwere applied
incidence andwith a probability
mortality of 0.5.continues
rates, cancer Consequently,
to be
the primary
the dataset wascauseexpanded
of deathto a totaltoofthis
related 25,000 images,
disease in thewhich
United were further
States. categorized into
Adenocarcinoma of
the
fivelung typically
distinct arisesEach
categories. fromcategory
the mucosal glands5000
contained and images.
accountsThe forimages
approximately 40% of
were resized to
the total cases
dimensions of of
768lung cancer
× 768 prior[33].
to the application of augmentation techniques. To guarantee
privacy and unrestricted usage, these images underwent validation and adhered to the
3.2.2. Lung Benign
regulations set forth by the Health Insurance Portability and Accountability Act (HIPAA).
TableThe lung and
2 displays the bronchus
designatedencompass
names andaIDs diverse collection
assigned to eachof benign
class tumors,
of images which
within the
typically
dataset and manifest as single,
an overview ofperipheral
the datasetlung nodules
split. or, less
To reduce frequently, ascomplexity,
computational endobronchialwe
lesions
downsizedthat the
result in obstructive
images symptoms.
in our training These
and test tumors from
directories commonly occur without
our pre-existing any
dataset,
noticeable
which hadsymptoms. Surgical
100 × 100 pixels. Theremoval of all
utilization endobronchial
of the training andlesions is recommended
test directories to
is justified
ease symptoms and prevent potential damage to distal lung tissue.
by the fact that the test directory’s images are utilized to test the global model, whereas
the training directory’s images are disseminated to end devices/clients for local data train-
3.2.3. Lung Squamous Cell Carcinoma
ing. The dataset containing 25,000 lung and colon cancer images is organized into training,
Squamous
testing, cell carcinoma
and validation sets at an(SCC) of theratio.
80:10:10 lung, alternatively referred to as squamous cell
lung cancer, represents a subtype of non-small cell lung cancer (NSCLC). Squamous cell
lung cancers frequently manifest in the central region of the lung or the primary airway,
specifically the left or right bronchus. Tobacco smoke is well recognized as the primary
causal agent responsible for cellular change. The prevalence of smoking-related lung cancer
is estimated to be approximately 80% in males and 90% in females [34].
Technologies 2024, 12, 151 11 of 28

3.2.4. Colon Adenocarcinoma

Colon adenocarcinoma is a type of colorectal cancer that originates from glandular cells
that line the inner surface of the colon. It generally arises from pre-existing adenomatous
polyps, which can gradually become cancerous. Colon adenocarcinoma is characterized by
alterations in bowel patterns, bleeding from the rectum, and discomfort in the abdomen.
Screening techniques, such as colonoscopies, are vital for identifying and preventing cancer
at an early stage by enabling the elimination of precancerous polyps before they develop
into cancerous growths.

3.2.5. Colon Benign

Benign conditions in the colon are characterized by the presence of non-malignant
growths or anomalies that do not pose any risk to one’s health. Examples include non-
malignant polyps, diverticulosis, and inflammatory bowel illnesses such as ulcerative
colitis. Although some disorders may not be malignant, they can nonetheless induce
symptoms and may necessitate medical attention and treatment.

3.3. Federated Learning (FL)

The introduction of federated learning technology facilitates the training of a model
by incorporating a central server while simultaneously maintaining decentralized training
data on distributed clients. The objective is to leverage FL technology to ensure the
confidentiality of the user data and facilitate data expansion. In a concurrent manner, this
approach enables participants to collectively train a global model without the need to
share their individual private data, as illustrated in Figure 6. This diagram illustrates a
federated learning architecture where multiple decentralized data sources (represented
by servers) locally train machine learning models on their own data. The locally trained
models are then encrypted and sent to a central server, which aggregates them into a
global model without accessing the raw data. This decentralized approach preserves data
privacy while improving the global model through collaborative learning. The regional data
need to undergo preprocessing for each contributor. This involves making modifications,
digitizing, and standardizing the data to transform it into a standardized format while
ensuring privacy. The distribution of images from our dataset among various clients is
achieved by dividing the total size of the image by the number of clients. The dataset is
divided uniformly, resulting in the generation of independent and identically distributed
(IID) data.
After receiving the model parameters from the clients, the server will summarize the
information on the basis of the structure of the central server. It updates the parameters of
the existing model and stores it for the subsequent round of training parameter upload and
collection from the participants before it is redistributed. The initial FL iterative procedure
is subsequently followed by the iterative process of our comprehensive model. In this
process, a CNN was integrated by tailoring for cancer tissue data samples and modifying
the model to enable continuous iterations. In the context of dispersed machine learning,
the prevailing approach for aggregating models involves ensuring that all participants
possess an equal number of training samples. In the context of federated learning, it is
common for participating members to have varying quantities of data at their disposal.
The aggregation of local models is achieved by applying a weighting scheme based on
the quantity of training samples available for each model. Consequently, the models that
possess a greater quantity of samples are given preference over those with a limited number.
The aforementioned method has a straightforward nature, yet it has demonstrated a notable
level of prevalence and effectiveness in scenarios involving federated learning. The global
model parameter can be defined as

k
nk k
ω t +1 ← ∑ ω
n t +1
(1)
k =1
 proach enables participants to collectively train a global model without the need to share
their individual private data, as illustrated in Figure 6. This diagram illustrates a federated
learning architecture where multiple decentralized data sources (represented by servers)
locally train machine learning models on their own data. The locally trained models are
then encrypted and sent to a central server, which aggregates them into a global model
Technologies 2024, 12, 151 12 of 28
without accessing the raw data. This decentralized approach preserves data privacy while
improving the global model through collaborative learning. The regional data need to un-
dergo preprocessing for each contributor. This involves making modifications, digitizing,
where
and standardizing the data to transform it into a standardized format while ensuring pri-
k = participants;
vacy. The distribution of images from our dataset among various clients is achieved by
nk = samples of participants k;
dividing the total size of the image by the number of clients. The dataset is divided uni-
n = samples of all participants;
formly, resulting in the generation of independent and identically distributed (IID) data.
wt+1 = local parameter of k participants.

Figure Workflowofoffederated
Figure 6. Workflow federated learning
learning within
within a federated
a federated server server to client.
to a local a local client.

A federated
After round
receiving is the name
the model given to
parameters eachthe
from iteration
clients,ofthe
this process,
server willand it consists
summarize the of
concurrent training, update aggregation, and parameter distribution. The following
information on the basis of the structure of the central server. It updates the parameters are the
primary
of control
the existing parameters
model thatitare
and stores forutilized in the process
the subsequent roundof ofcomputing FL:
training parameter upload
C = customers or contributors who took part in an update cycle;
and collection from the participants before it is redistributed. The initial FL iterative pro-
cedureE is
= number of local
subsequently epochs by
followed thatthe
each contributor
iterative process has
ofbeen responsible for; model. In
our comprehensive
B = smallest batch size that can be utilized for a local update.
this process, a CNN was integrated by tailoring for cancer tissue data samples and modi-
1 and
fying βthe β2 are
model toconsidered hyperparameters.
enable continuous iterations. In the context of dispersed machine
During the regional model optimization step, the clients execute a specific number of
local epochs. The Adam optimizer is utilized for the first- and second-order moments to
overcome local minima [35]:

1 − βn2
p
m
ωi,t ← ωi,t − η × √ i,n (2)
1 − βn1 vi,n + σ

3.4. Inception-V3 Model

Inception-V3 is an advanced deep convolutional neural network (CNN) architecture
that has been meticulously crafted by Google as an integral component of the Incep-
tion project [36]. The algorithm has been specifically developed to cater to a range of
computer vision applications, with a particular focus on image classification and object
recognition by enabling both efficiency and accuracy. The exceptional performance and
efficiency of Inception-V3 are attributed to its utilization of a distinctive inception module,
which integrates multiple filter sizes into a single layer, enabling the capture of features at
various scales.
The primary technological advancement of Inception-V3 lies within its inception
modules, which encompass various building blocks, such as convolutions, average pooling,
max pooling, concatenations, dropouts, and fully connected layers. Figure 7 represents the
architecture of Inception-V3. The architecture is composed of multiple modules, including
Technologies 2024, 12, 151 13 of 28

convolutional layers, pooling layers, and inception modules that combine various filter
sizes (e.g., 1 × 1, 3 × 3, and 5 × 5) to capture different types of image features. The
model extensively utilizes batch normalization, which is applied to activation inputs.
The computation of loss involves the utilization of Softmax. The Inception-V3 model
also includes the integration of global average pooling, a technique that replaces the
conventional fully connected layers located at the final stage of the network. By reducing
Technologies 2024, 12, x FOR PEER REVIEW 14 of 31
overfitting and parameterizing the model size, the efficiency of the system is enhanced and
its adaptability to new tasks is improved.

Figure 7. Inception-V3 architecture [36].

Figure 7. Inception-V3 architecture [36].
3.5. Proposed Workflow for FL
3.5. Proposed Workflow
This section for FLan understanding of various implementations of federated
provides
ThisThis
learning. section provides the
encompasses an procedure
understanding of various
of training implementations
data on individual clientofdevices
federated
and
learning. Thisaggregating
subsequently encompasses thethe procedure
local of atraining
models on central data on individual client devices
server.
and subsequently aggregating the local models on a central server.
3.5.1. Local Device (Client) Creation
3.5.1.Benign
Local Device (Client)
conditions Creation
in the colon are characterized by the presence of non-malignant
growths or anomalies
Benign conditionsthat do colon
in the not pose
are any risk to one’s
characterized health.
by the Examples
presence include non-
of non-malignant
malignant
growths or polyps, and diverticulosis.
anomalies that do not pose any risk to one’s health. Examples include non-
Data distribution
malignant polyps, and among end devices is achieved through a technique called data
diverticulosis.
sharding.
DataThis process involves
distribution among end the devices
distribution of smaller
is achieved datasets,
through referred to
a technique as logical
called data
shards or chunks, from a larger dataset. In the context of a real-world application,
sharding. This process involves the distribution of smaller datasets, referred to as logical it is
important
shards or to note that
chunks, fromeach clientdataset.
a larger may possess
In the datasets
context ofofavarying sizes
real-world and variations.
application, it is
However,
importantfor developing
to note that eachtheclient
federated learning
may possess (FL) prototype,
datasets of varyingthe shards
sizes among the
and variations.
clients
However,in the
formodel were evenly
developing distributed.
the federated learningThe
(FL) size of each shard
prototype, was determined
the shards among the cli-via
the following
ents formula:
in the model were evenly distributed. The size of each shard was determined via the
following formula: Total No. o f Images
Shard Size = (3)
No. o f Clients
.
Shard Size = (3)
.
The data shards subsequently underwent processing and were allocated to the clients
via theThebatching procedure.
data shards Uponunderwent
subsequently conclusionprocessing
of the process, eachallocated
and were client obtained their
to the clients
respective local datasets
via the batching and prepared
procedure. them for of
Upon conclusion training purposes.
the process, each client obtained their
respective local datasets and prepared them for training purposes.
3.5.2. Integration of Inception-V3 and Its Configuration
3.5.2.As
Integration
previously of stated,
Inception-V3 anddatasets
the local Its Configuration
underwent training on local devices. The
Inception-V3 architectural model was implemented
As previously stated, the local datasets underwent on each clientontolocal
training facilitate the The
devices. training
In-
ofception-V3
their local architectural
data. Additionally,
modelitwaswasimplemented
utilized in theon
global
eachmodel forfacilitate
client to testing purposes. Ad-
the training
ditionally, three
of their local additional
data. convolutional
Additionally, neural in
it was utilized network (CNN)
the global layers
model for were incorporated
testing purposes.
Additionally, three additional convolutional neural network (CNN) layers were incorpo-
rated into the implementation. These layers included the dense layer, flatten layer, and
dropout layer. The ImageNet dataset was utilized as the default weight in the model, sim-
ilar to the implementation observed in the CNN. The Adam optimizer with a learning rate
of 0.00001 was used to optimize the accuracy of the Inception-V3 model. Additionally, the
Technologies 2024, 12, 151 14 of 28

into the implementation. These layers included the dense layer, flatten layer, and dropout
layer. The ImageNet dataset was utilized as the default weight in the model, similar to the
implementation observed in the CNN. The Adam optimizer with a learning rate of 0.00001
was used to optimize the accuracy of the Inception-V3 model. Additionally, the categorical
cross-entropy method was implemented to calculate the loss function.

3.5.3. Communication Rounds

The communication round was executed a total of 50 times to achieve a specific level
of accuracy for our global model. The communication cycle involved the clients, which
hold the datasets previously generated and acquire their weights on the basis of the global
model’s weight. Each of the five clients then proceeded to train their respective local
data and generate the accuracy of their individual local models. The clients subsequently
proceeded to transmit their trained models to the central server, which is commonly referred
to as the global model, for the purpose of aggregation. This aggregation process involves
performing an averaging operation on the federated learning (FL) model. After the initial
aggregation, the global model generated a new weight. In the subsequent round, the clients
adjusted their weights on the basis of the updated global weight. The weight scaling factor
was incorporated into the model to facilitate this particular operation.

3.6. Experimental Setup and Hyperparameter Settings

The experiments are performed on Google Colab, and the hyperparameters of the
experiments are shown in Table 3. The proposed federated learning model was run on an
NVIDIA A100 40 GB GPU and 85 GB RAM provided by Google Colab services. The images
for the original dataset had dimensions of 100 × 100 and were trained in this environment.
The hyperparameter settings were set to achieve the maximum computational capacity for
the best possible performance on the whole dataset. The hyperparameters were optimized
through trial and error, as shown in Table 4.

Table 3. Description of the system configuration of Google Colab.

System Specification
Processor Intel Xeon CPU
CPU ~2.30 GHz
RAM 85 GB
GPU NVIDIA A100
GPU RAM 40 GB
Hard Disk 80 GB

Table 4. Description of system hyperparameters.

Hyperparameter Value
Optimizer Adam
Loss Categorical Crossentropy
Batch Size 16
Image Size 100 × 100
No. of Epochs 50
No. of Clients 5

3.7. Evaluation Metrics

Several evaluation criteria were used to evaluate the performance of our proposed
model. They are the accuracy, precision, recall, F1 score, specificity, and confusion matrix.
The metric commonly used to evaluate the performance of machine learning models is
accuracy. The metric reflects the frequency at which the model accurately predicts the
positive class. Precision is the proportion of accurate forecasts. It is the ratio of the number
of accurate positive forecasts to the entire number of positive predictions. Recall indicates
Technologies 2024, 12, 151 15 of 28

the proportion of accurate predictions relative to the ground truth. The F1 score takes
the harmonic mean of both precision and recall to create a single metric. The respective
formulas are presented here.

TP + TN
Accuracy = (4)
TP + TN + FP + FN
TP
Precision = (5)
TP + FP
TP
Recall = (6)
TP + FN
2 × ( Precision × Recall )
F1_score = (7)
Precisin + Recall
TN
Speci f icity = (8)
TN + FP
where TP = True Positive, TN = True Negative, FP = False Positive, and FN = False Negative.

4. Experimental Results
The suggested technique is evaluated via histopathology images obtained from the
LC25000 dataset [32]. This section presents a concise overview of the outcomes obtained
by the proposed federated learning model when provided with Inception-V3 for the
categorization of histological images from the lung and colon cancer datasets. The results
were categorized into three sections: lung, colon, and combined lung–colon outcomes for
the model.
Every experiment involved the training of five distinct transfer learning models,
namely, Inception-V3, VGG16, ResNet-50, ResNeXt50, and Xception. Initially, transfer
learning models were utilized to train the initial lung cancer images to evaluate the
models’ performance on the dataset. A comparative analysis of the outcomes of the
models was conducted to determine which model was more suitable for the federated
learning approach.

4.1. Lung Cancer

Table 5 and Figure 8 present a concise overview of the performance of the federated
learning model in comparison to various base models when it is applied to the lung cancer
images. The Inception-V3, VGG16 [37], ResNet50 [38], ResNeXt50 [39], and Xception [40]
models achieved average classification accuracies of 99.16%, 98.33%, 99.20%, 99.20%, and
99.27%, respectively. The Inception-V3 and Xception models exhibited the best accuracy,
with a precision and recall of 99.27%. Conversely, the VGG16 model had the lowest
performance. Inception-V3 displayed a commendable performance, approaching the
outcomes achieved by Xception. Upon integrating the federated learning model with
Inception-V3, it became clear that it significantly outperformed all the other findings. The
model attained 99.87% accuracy, with a precision and recall of 1.0.

Table 5. Performance analysis of the lung cancer images compared to the base models.

Classification Model Precision Recall Accuracy

“Federated Learning with Inception-V3” 1.0 1.0 99.87%
Inception-V3 0.9916 0.9916 99.16%
VGG16 0.9833 0.9833 98.33%
ResNet-50 0.9926 0.992 99.20%
ResNeXt50 0.992 0.9927 99.20%
Xception 0.9927 0.9927 99.27%
Technologies 2024, 12, x FOR PEER REVIEW 17 of 31

Technologies 2024, 12, x FOR PEER REVIEW 17 of 31

Technologies 2024, 12, 151 16 of 28

Figure 8. Performance comparisons of federated learning with Inception-V3 with other centralized
transfer8.learning
Figure 8. models,
Performance i.e., Inception-V3,
comparisons federatedVGG-16,
of federated ResNet-50,
learning with ResNeXt-50,
with Inception-V3
Inception-V3 andcentralized
with other
other Xception, on lung
Figure Performance comparisons of learning with centralized
cancer
transferimages.
learning models, i.e., Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on lung
transfer learning models, i.e., Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on lung
cancer images.
cancer images.
Figure 9 displays the confusion matrix generated by the proposed technique for lung
Figure
cancer. The99data
Figure displays
displays the
theconfusion
in Table 6 clearlymatrix
confusion generated
indicate
matrix byproposed
that the
generated the the
by proposed technique
model
proposed achievedfor lung
technique 100%
for accu
cancer.
lung The
cancer. data
The in Table
data in 6 clearly
Table 6 indicate
clearly that
indicatethe
thatproposed
the model
proposed achieved
model 100%
achieved
racy in detecting squamous cell carcinoma (lung_scc) and benign lung (lung_bnt) images accu-
100%
racy in detecting
accuracy squamous
in detecting cell carcinoma
squamous (lung_scc)
cell carcinoma and benign benign
(lung_scc) lung (lung_bnt) images.
Additionally, the model demonstrated 99.60% accuracyandin correctlylung (lung_bnt)
identifying lung ade
Additionally,
images. the model
Additionally, demonstrated
the model 99.60% accuracy
demonstrated 99.60% in correctly
accuracy in identifying
correctly lung ade-
identifying
nocarcinoma (lung_aca) images.
nocarcinoma (lung_aca) images.
lung adenocarcinoma (lung_aca) images.

Figure 9. Confusion matrix for federated learning with Inception-V3 for lung cancer classification.

Figure9.9.Confusion
Figure Confusionmatrix
matrixfor for
federated learning
federated with Inception-V3
learning for lungfor
with Inception-V3 cancer
lungclassification.
cancer classification.
Technologies 2024, 12, x FOR PEER REVIEW 18 of 31

Table 6. Class-wise performance analysis of the lung cancer images using federated learning with
Technologies 2024, 12, 151 17 of 28
Inception-V3.

Type of Class Precision Recall F1 Score Specificity Accuracy

Table 6. Class-wise performance analysis of the lung cancer images using federated learning with
lung_aca
Inception-V3. 99.60% 100.00% 99.80% 99.80% 99.60%
lung_bnt 100.00% 100.00% 100.00% 100.00% 100.00%
Type of Class Precision Recall F1 Score Specificity Accuracy
lung_scc 100.00% 100.00% 100.00% 100.00% 100.00%
lung_aca 99.60% 100.00% 99.80% 99.80% 99.60%
Macro Average
lung_bnt 99.87%
100.00% 100.00%
100.00% 99.93%
100.00% 99.93%
100.00% 99.87%
100.00%
lung_scc 100.00% 100.00% 100.00% 100.00% 100.00%
Macro Average
4.2. Colon Cancer 99.87% 100.00% 99.93% 99.93% 99.87%

Table 7 and Figure 10 provide a comprehensive overview of the performance of the

4.2. Colon learning
federated Cancer model in comparison to other base models when analyzing colon can-
cer images. The Figure
Table 7 and Inception-V3,
10 provide VGG16, ResNet50, ResNeXt50,
a comprehensive andperformance
overview of the Xception models
of the
federatedaverage
achieved learningclassification
model in comparison
accuracies to other base99.60%,
of 100%, models99.70%,
when analyzing
100%, andcolon cancer
100%, re-
images. The
spectively. Inception-V3,
The Xception and VGG16,
ResNeXtResNet50,
modelsResNeXt50,
achieved theand Xception
highest modelsprecision,
accuracy, achieved
average
and recallclassification
rates of 100%,accuracies
1.0, andof 100%,
1.0, 99.60%, 99.70%,
respectively, whereas100%, and 100%,
the VGG16 modelrespectively.
achieved
Thelowest
the Xception and ResNeXt
performance. models achieved
Inception-V3 the highest
demonstrated accuracy,
strong precision,
performance, and recall
comparable to
rates
the of 100%,
results 1.0, and
achieved by 1.0, respectively,
Xception. whereas our
After applying the VGG16 model
federated achieved
learning model the lowest
with In-
performance.
ception-V3, Inception-V3
it became demonstrated
clear that strong
it significantly performance,allcomparable
outperformed to the results
the other methods. The
achieved results include an accuracy rate of 100%, precision score of 1.0, and recall scoreit
achieved by Xception. After applying our federated learning model with Inception-V3,
became
of 1.0. clear that it significantly outperformed all the other methods. The achieved results
include an accuracy rate of 100%, precision score of 1.0, and recall score of 1.0.

Figure
Figure 10.
10. Performance
Performancecomparisons
comparisonsof offederated
federatedlearning
learningwith
withInception-V3
Inception-V3 with
with other
othercentralized
centralized
transfer learning models, i.e., Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on co-
transfer learning models, i.e., Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on colon
lon cancer images.
cancer images.

Figure 11 displays the confusion matrix of the proposed method for colon cancer.
On the basis of the data presented in Table 8, the proposed model clearly exhibited a
remarkable ability to accurately detect colon adenocarcinoma (colon_aca) and colon benign
(colon_bnt) images, achieving a 100% accuracy rate.
 Table 7. Performance analysis of colon cancer images compared to the base models.

Classification Model Precision Recall Accuracy

“Federated Learning with Inception-V3” 1.0 1.0 100.00%
Technologies 2024, 12, 151 Inception-V3 0.996 0.996 18 of 99.60%
28
VGG16 0.994 0.994 99.40%
ResNet-50 0.997 0.997 99.70%
Table 7. Performance analysis of
ResNeXt50colon cancer images compared to the
1.0 base models. 1.0 100.00%
Xception
Classification Model Precision 1.0Recall 1.0Accuracy 100.00%
“Federated Learning with Inception-V3” 1.0 1.0 100.00%
Figure 11Inception-V3
displays the confusion matrix0.996
of the proposed method for
0.996 colon cancer. On
99.60%
the basis of the data
VGG16presented in Table 8, the proposed model
0.994 0.994 clearly exhibited
99.40% a remark
ResNet-50 0.997 0.997 99.70%
able ability to accurately detect colon adenocarcinoma (colon_aca) and colon benign (co
ResNeXt50 1.0 1.0 100.00%
lon_bnt) images, achieving a 100% accuracy 1.0
Xception rate. 1.0 100.00%

Figure 11.Confusion
Figure11. Confusionmatrix for for
matrix federated learning
federated with Inception-V3
learning for colon
with Inception-V3 cancer
for colonclassification.
cancer classification
Table 8. Class-wise performance analysis of the colon cancer images using federated learning with
Table 8. Class-wise performance analysis of the colon cancer images using federated learning with
Inception-V3.
Inception-V3.
Type of Class Precision Recall F1 Score Specificity Accuracy
Type of Class Precision Recall F1 Score Specificity Accuracy
colon_aca 100% 100% 100% 100% 100%
colon_aca
colon_bnt 100%
100% 100%
100% 100% 100% 100% 100% 100% 100%
colon_bnt
Macro Average 100%
100% 100%
100% 100% 100% 100% 100% 100% 100%
Micro Average
Macro Average 100%
100% 100%
100% 100% 100% 100% 100% 100% 100%
Micro Average 100% 100% 100% 100% 100%
4.3. Lung and Colon Cancers
4.3. Table 9 and
Lung and Figure
Colon 12 provide a comprehensive overview of the performance of
Cancers
the federated learning model in comparison to other base models on lung and colon
cancerTable 9 and
images. TheFigure 12 provide
Inception-V3, a comprehensive
VGG16, overview
ResNet50, ResNeXt50, andofXception
the performance
models of th
federated learning model in comparison to other base models on lung and
achieved average classification accuracies of 98.96%, 98.36%, 98.88%, 98.88%, and 99.10%, colon cance
images. TheThe
respectively. Inception-V3,
Xception modelVGG16, ResNet50,
demonstrated the ResNeXt50, andprecision,
highest accuracy, Xceptionand
models
recall,achieved
average
all classification
at 99.10%. Conversely,accuracies
the VGG16of 98.96%,
model 98.36%,
exhibited 98.88%,
the lowest 98.88%, and
performance. 99.10%, respec
Inception-
V3 demonstrated
tively. a strong
The Xception performance,
model comparable
demonstrated to theaccuracy,
the highest results achieved by Xception.
precision, and recall, all a
After implementing the federated learning model with Inception-V3,
99.10%. Conversely, the VGG16 model exhibited the lowest performance. it became clear that it
Inception-V
significantly outperformed all the other methods. The achieved accuracy, precision, and
demonstrated a strong performance, comparable to the results achieved by Xception. Af
recall were all 99.72%.
ter implementing the federated learning model with Inception-V3, it became clear that i
 significantly outperformed all the other methods. The achieved accuracy, precision, and
recall were all 99.72%.
Technologies 2024, 12, 151 19 of 28
Table 9. Performance analysis of lung and colon cancer images compared to the base models.

Classification Model Precision Recall Accuracy

Table 9. Performance
“Federated analysis
Learning of lung and colon cancer images compared to the base models.
with Incep-
99.72% 99.72% 99.72%
tion-V3”
Classification Model Precision Recall Accuracy
Inception-V3 98.96% 98.96% 98.96%
“Federated Learning with Inception-V3” 99.72% 99.72% 99.72%
VGG16
Inception-V3 98.36% 98.96% 98.36%98.96% 98.36%
98.96%
ResNet-50
VGG16 98.96% 98.36% 98.84%98.36% 98.88%
98.36%
ResNet-50
ResNeXt50 98.88% 98.96% 98.88%98.84% 98.88%
98.88%
ResNeXt50 98.88%
Xception 99.10% 99.10%98.88% 98.88%
99.10%
Xception 99.10% 99.10% 99.10%

Figure 12. Performance comparisons of federated learning with Inception-V3 with other centralized
Figure 12. Performance comparisons of federated learning with Inception-V3 with other centralized
transfer learning models, i.e., Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on com-
transfer
bined lunglearning models,
and colon i.e.,
cancer Inception-V3, VGG-16, ResNet-50, ResNeXt-50, and Xception, on
images.
combined lung and colon cancer images.
Figure 13 displays the confusion matrix of the proposed method for lung and colon
Figure 13 displays the confusion matrix of the proposed method for lung and colon
cancers. On the basis of the data presented in Table 10, the proposed model exhibited a
cancers. On the basis of the data presented in Table 10, the proposed model exhibited a
high level of accuracy in detecting benign lung (lung_bnt) images, with a 100% accuracy
high level of accuracy in detecting benign lung (lung_bnt) images, with a 100% accuracy
rate. Similarly, the model demonstrated 99.72% accuracy in correctly identifying lung ad-
rate. Similarly, the model demonstrated 99.72% accuracy in correctly identifying lung
enocarcinoma (lung_aca) and squamous cell carcinoma (lung_ssc) images. For the colon
adenocarcinoma (lung_aca) and squamous cell carcinoma (lung_ssc) images. For the colon
cancer
cancer images,
images, 100%
100% accuracy
accuracy was
was achieved
achieved for
for both
both classes.
classes.

Table 10. Class-wise performance analysis of the lung and colon cancer images using federated
learning with Inception-V3.

Type of Class Precision Recall F1 Score Specificity Accuracy

colon_aca 100% 100% 1.000 100% 100%
colon_bnt 100% 100% 1.000 100% 100%
lung_aca 98.80% 99.80% 0.999 99.70% 99.72%
lung_bnt 100% 100% 1.000 100% 100%
lung_scc 99.80% 98.812% 0.993 99.95% 99.72%
Macro Average 99.72% 99.72% 0.9984 99.93% 99.88%
Micro Average 99.72% 99.72% 99.72% - 99.72%
 lung_bnt 100% 100% 1.000 100% 100%
lung_scc 99.80% 98.812% 0.993 99.95% 99.72%
Macro Average 99.72% 99.72% 0.9984 99.93% 99.88%
Technologies 2024, 12, 151 20 of 28
Micro Average 99.72% 99.72% 99.72% - 99.72%

13.Confusion
Figure 13.
Figure Confusion matrix
matrixforfor
federated learning
federated with with
learning Inception-V3 for combined
Inception-V3 lung andlung
for combined colonand colon
cancer classifications.
cancer classifications.
4.4. Client-Wise Results
4.4. Client-Wise Results
Figure 14 illustrates the iterative process of updating and optimizing the global model
Figure
following each14communication
illustrates the iterative
event and how process of updating
the clients’ individualand optimizing
accuracies the global
vary from
model
the globalfollowing
accuracy. each
After communication event and
a total of 50 communication how the
rounds, theaccuracy
clients’reached
individual accuracies
99.867%
vary from
for lung theclassification,
cancer global accuracy.100% forAfter
colona total
cancerofclassification,
50 communicationand 99.72% rounds,
for lungthe
andaccuracy
colon cancer
reached classifications.
99.867% for lungIncancer
the local context, each 100%
classification, communication
for colonround cancerencompassed
classification, and
comprehensive
99.72% for lungtraining sessions
and colon cancerfor all clients. To facilitate
classifications. In thethe simulation,
local context, aeach
single epoch
communication
was conducted for the purpose of local communication rounds. Following
round encompassed comprehensive training sessions for all clients. To facilitate the sim- the completion
of the simulation, accuracy, loss, and categorical accuracy metrics for all of the esteemed
ulation, a single epoch was conducted for the purpose of local communication rounds.
clients were successfully acquired. Despite not being immediately evident, the findings
Following the completion
indicate a positive upward trend of the simulation,
(thicker red lines).accuracy,
This impliedloss,that
and categorical
there accuracy met-
was a possibility
rics for all of the esteemed clients were successfully acquired.
of an increase in the accuracy of the client, although it cannot be guaranteed. Despite not being immedi-
atelyItevident,
was evidentthethat,
findings indicatewith
in accordance a positive upward
the initial trend
predictions, the(thicker red lines). This im-
clients demonstrated
plied thataccuracy
increased there was withaeach
possibility
successiveof communication
an increase in round.
the accuracy of the client,
In the client-wise although it
accuracy
measure,
cannot bethere was a steady increase in accuracy, indicating that performance improved
guaranteed.
with each communication round.

4.5. Explainable AI (XAI)

XAI methods [41], such as Grad-CAM, heatmaps, and saliency maps, were imple-
mented to explore their usefulness in the classification of lung and colon cancers. Grad-
CAM [42] was utilized in deep learning to visualize the key regions of an input image that
had the greatest impact on the output of a convolutional neural network (CNN) model.
The heatmap demonstrated the correlation between the image’s features and the model’s
prediction, highlighting the significant areas that played a role in determining the final
prediction. As saturation increases, the weight assigned to those pixels also increases.
Figure 15 presents a clear visualization of the significant features (pixels) in the image that
the model deemed crucial for its prediction. The saliency map was generated by calculating
the gradients of the predicted class score with respect to the input image pixels. It is a
method similar to the heatmap. The algorithm identifies the pixels in the input image that
have the greatest impact on the predicted class score.
Technologies 2024, 12, 151 21 of 28
Technologies 2024, 12, x FOR PEER REVIEW 22 of 31

Figure 14. Global accuracy vs. local accuracy for the combination of the lung, colon and lung–colon methods.
Technologies
Technologies 2024,
2024, 12,
12, x151
FOR PEER REVIEW 24 of 31
22 of 28

Figure 15. Attention visualization of images employing GradCAM, heatmaps, and saliency maps.
Figure 15. Attention visualization of images employing GradCAM, heatmaps, and saliency maps.
5. Discussion
Grad-CAM determines the gradients of the target class in relation to the last convolu-
5.1. Comparative
tional layer of theAnalysis
model. Grad-CAM uses gradients to identify the influential regions of an
imageTothat contributethe
demonstrate to the classification
effectiveness of decision. The heatmap
the proposed method,illustrates the significance
a comparative analysis
of each
was pixel in with
conducted contributing
previous to studies
the model’s
that classification decision. dataset.
utilized the LC25000 The colorThere
red represents
are very
regions
few thatthat
studies havearegreater significance
exclusively concernedfor the
withclassification,
lung and colon whereas cooler colors
malignancies, and thesuch as
ma-
blue indicate regions that are less important. Saliency maps are generated
jority of them employed CNN-based centralized deep learning techniques. The method- by calculating
the gradient
ologies of thehere
discussed predicted
do not class
ensurescore
dataofprivacy,
the model with necessitate
as they respect to the pixels
access of the
to all input
training
image. and
images The their
purpose of this technique
corresponding is to identify
classification labels.the regions
The in themethodology
proposed input image that have
not only
the greatest impact
outperforms the mostonadvanced
the model’s output.
methods forThe images
detecting were
lung andchosen
colon randomly from the
cancers separately
dataset
and andcomparable
shows may behave differently from
performance one image
for detecting to and
lung the other.
colon cancers combinedly but
By examining the GRAD-CAM output and heatmaps,
also maintains the privacy element of patient data (Table 11). it was observed that the classi-
fier distinguishes lung adenocarcinoma by focusing on cell near-white regions, which are
indicative of mucosa or connective tissue. For lung squamous cell carcinoma, the classifi-
cation is based on the fish-scale appearance of the cells under the microscope. In healthy
lung tissue (lung_benign), the classifier identifies red blood cells as a key distinguishing
feature. For colon adenocarcinoma, the classifier relies on areas with irregular glandular
structures, atypical cell shapes, and disrupted tissue architecture, highlighting the presence
of desmoplastic stroma and inflammatory cell infiltrates. In contrast, benign colon tissue is
characterized by regular, well-organized glandular structures and uniform cell shapes.
Technologies 2024, 12, 151 23 of 28

5. Discussion
5.1. Comparative Analysis
To demonstrate the effectiveness of the proposed method, a comparative analysis was
conducted with previous studies that utilized the LC25000 dataset. There are very few
studies that are exclusively concerned with lung and colon malignancies, and the majority
of them employed CNN-based centralized deep learning techniques. The methodologies
discussed here do not ensure data privacy, as they necessitate access to all training images
and their corresponding classification labels. The proposed methodology not only outper-
forms the most advanced methods for detecting lung and colon cancers separately and
shows comparable performance for detecting lung and colon cancers combinedly but also
maintains the privacy element of patient data (Table 11).

Table 11. Comparative analysis of previous studies on LC25000.

Performance
Previous Studies Year Approaches
Colon Lung Lung and Colon
Mangal et al. [9] 2020 Deep learning approach using CNN Accuracy: 96.00% Accuracy: 97.89% -
Tasnim et al. [43] 2021 CNN with max pooling Accuracy: 99.67% - -
Talukder et al. [16] 2021 Deep feature extraction and ensemble learning Accuracy: 100% Accuracy: 99.05% Accuracy: 99.30%
Shandilya et al. [44] 2021 Pretrained CNN - Accuracy: 98.67% -
Hadiyoso et al. [10] 2022 VGG-19 architecture and CLAHE framework Accuracy: 98.96% - -
Karim et al. [45] 2022 Extreme learning machine (ELM)-based DL - Accuracy: 98.07% -
Accuracy: 98.97%
Raju et al. [46] 2022 Extreme learning machine (ELM)-based DL Precision: 98.87% - -
F1 Score: 98.84%
Accuracy: 99.00% Accuracy: 99.53%
Chehade et al. [8] 2022 XGBoost Precision: 98.6% Precision: 99.33% Accuracy: 99%
F1 Score: 98.8% F1 Score: 99.33%
Accuracy: 99.84%
Ren et al. [47] 2022 Deep convolutional GAN (LCGAN) - Precision: 99.84% -
F1 Score: 99.84%
Transfer learning
Mehmood et al. [13] 2022 - - Accuracy: 98.4%
with class selective image processing
Transfer learning with a secure
Khan et al. [30] 2023 - - Accuracy: 98.80%
IoMT-based approach
Toğaçar et al. [14] 2022 DarkNet-19 model and SVM classifier - - Accuracy: 99.69%
Attallah et al. [48] 2022 CNN features with transformation methods - - Accuracy: 99.6%
Accuracy: 96.33%
Deep learning (DL) and digital image processing
Masud et al. [7] 2022 - - Precision: 96.39%
(DIP) techniques
F1 Score: 96.38%
Accuracy: 99.64%
Al-Jabbar et al. [17] 2023 Fusion of GoogleNet and VGG-19 - -
Precision: 100%
Ananthakrishnan et al. [18] 2023 CNN with an SVM classifier Accuracy: 99.8% Accuracy: 98.77% Accuracy: 100%
Accuracy: 100% Accuracy: 99.87% Accuracy: 99.72%
Proposed Model 2024 Federated learning with Inception-V3 Precision: 100% Precision: 99.87% Precision: 99.72%
F1 Score: 100% F1 Score: 99.87% F1 Score: 99.72%
Note: Bold numerical values indicate best results.

5.2. Strengths of the Proposed Model

The initial approach involves the use of the Inception-V3 model to process local models
on individual client devices, leveraging the data specific to each device. However, it is
important to note that individual devices retain ownership of their respective local data. The
second process involves aggregating local models to generate a global model with enhanced
accuracy. The global server receives only the parameters and weights. Both centralized
training and decentralized training are conducted with identical hyperparameter settings.
In the centralized case, the models are trained directly on the overall training data with
50 epochs. During the FL experiment, the training processes of five clients are stuck with
their local data, each with 50 communication rounds. This component is an integral part
of the federated learning system and is designed specifically to address and mitigate
privacy concerns.
Both lung cancer and colon cancers are prevalent types of cancers with high mortality
rates. Identifying a condition at an early stage significantly improves the chances of
survival. Physicians may face challenges in precisely identifying lung cancer with CT
images. Sharing patient data for research purposes among health organizations is a growing
concern because of the numerous restrictions in controlling patient data. Adopting suitable
data-sharing techniques can increase the likelihood of detecting cancer through CT scans.
Technologies 2024, 12, 151 24 of 28

In this context, we introduce the proposed federated learning methodology, which can
effectively guarantee adherence to regulations while addressing patient data and precisely
categorizing cancer cells with great accuracy.
The federated averaging technique aggregates model updates from multiple clients
to compute a global model. This aggregation process combines knowledge from diverse
sources while mitigating biases and noise present in individual client datasets. As a result,
the global model benefits from the collective intelligence of all participating clients, leading
to enhanced performance in cancer classification tasks. The framework used here hosts five
clients to demonstrate the federated learning process, but it is possible to scale efficiently
to a large number of devices. Each device performs local computations, and only model
updates are aggregated centrally, reducing the burden on central servers.
The utilization of federated learning techniques for lung and colon cancer classifica-
tions represents a significant advancement with important clinical implications. Through
the utilization of decentralized data sources from various healthcare institutions, federated
learning models present unique possibilities to improve diagnostic accuracy and customize
treatment strategies. By engaging in collaborative data sharing while prioritizing patient
privacy, these models have the ability to identify complex patterns within different types
of cancers. This allows for personalized interventions and well-informed clinical decision
making. Real-time decision support systems, powered by federated learning algorithms,
enable healthcare providers to gain timely insights, facilitating proactive management and
enhancing patient outcomes. In addition, the iterative process of federated learning allows
for the ongoing improvement of models, ensuring their ability to adapt to changing clinical
practices and contributing to advancements in precision oncology and population health
management. Federated learning is a revolutionary method in cancer care that promotes
cooperation, creativity, and advancement in the pursuit of better patient care and results.

5.3. Challenges in Federated Learning for Medical Diagnostics

Federated learning presents unique challenges in medical diagnostics, particularly
when applied to lung and colon cancer classifications. One of the foremost difficulties is
data heterogeneity across healthcare institutions. Medical images can vary significantly on
the basis of equipment used, imaging protocols, and patient demographics. This variation
can cause discrepancies in local model performance, making it difficult to achieve consistent
convergence of the global model. Ensuring that the federated model can be generalized ef-
fectively across diverse imaging environments while maintaining high diagnostic accuracy
remains a critical challenge.
Another challenge is the increased communication overhead inherent in federated
learning. Unlike traditional centralized models, where all data are aggregated in one place,
federated learning requires frequent communication between institutions to share model
updates without sharing raw data. This decentralized nature can lead to slower global
model updates, especially when dealing with large datasets typical of medical imaging.
Additionally, network connectivity issues or differences in computational resources across
institutions can further delay the training process and affect the overall performance of
the model.
An additional challenge in federated learning for medical diagnostics is the reluctance
of healthcare institutions to share data, even in a decentralized framework. Despite the
promise of privacy-preserving methods, many clinics are hesitant to participate because of
concerns about data ownership, security, and compliance with stringent regulations such
as the HIPAA or GDPR. Institutions often worry about losing control over their data or
exposing themselves to potential breaches or misuse, which makes collaboration difficult.
Overcoming this challenge requires building trust among participants through robust legal
agreements, transparent protocols, and guarantees that ensure data sovereignty while still
allowing for effective collaboration in training the global model.
Technologies 2024, 12, 151 25 of 28

5.4. Limitations and Future Work

One major limitation of the model is that it cannot handle heterogeneous data as
IID data were chosen for this study. However, user data can vary significantly from
institution to institution. This means that the data distribution varies significantly across
different devices, leading to challenges in model convergence and performance consistency.
The non-independent and identically distributed nature of data across clients can hinder
model convergence and lead to suboptimal global model performance. This needs to be
implemented to test the capability of the proposed model.
This study implemented federated learning to classify lung and colon cancer images on
IID data. The entire LC25000 dataset was divided among clients. However, in implementing
federated learning across multiple organizations, there is an enormous challenge due to
variations in data distribution and quality across participating institutions. Addressing
data heterogeneity is essential for improving the robustness of the federated learning model.
The size of the federated learning dataset might have been constrained by the number
of participating institutions. Expanding the dataset or exploring methods for effective
knowledge transfer with a limited sample size could enhance model generalizability.
Despite efforts to ensure privacy through federated learning, concerns regarding the
security of patient data transmission persist. Individuals can employ various techniques
to reconstruct the original data from these updates, such as model inversion or gradient
leakage. Further exploration of advanced privacy-preserving techniques, such as secure
aggregation or differential privacy, is warranted. With that in mind, diverse and hetero-
geneous data from various sources, such as multi-modal imaging (CT and PET), genetic
data, and clinical records, could be incorporated to increase generalizability. Advanced
privacy-preserving techniques, including homomorphic encryption or federated learning
with differential privacy, are researched and implemented to strengthen the security and
confidentiality of patient data during the federated learning process. The feasibility of real-
time implementation of federated learning models for cancer classification is particularly
valuable for a timely diagnosis and treatment planning in clinical settings.

6. Conclusions
The proposed federated learning approach with the Inception-V3 model to classify
lung and colon histopathological images yielded a significant outcome in accurately distin-
guishing between three subtypes of lung cancer and two subtypes of colon cancer from
histopathological images. The model demonstrated a remarkable accuracy of 99.720%,
as well as a recall, a precision, and an F1 score of 99.720%. The detection of colon cancer
achieved 100% accuracy in classifying both classes. The accuracy of lung cancer classifica-
tion for the three classes was 99.867%. This proves that the proposed federated learning
methodology can effectively guarantee adherence to regulations while dealing with patient
data and precisely categorizing cancer cells with great accuracy.

Author Contributions: All authors had an equal contribution in preparing and finalizing the
manuscript. Conceptualization: M.M.H., M.R.I., M.F.A., M.A. and J.H.; methodology, M.M.H.,
M.R.I., M.F.A., M.A. and J.H.; validation: M.M.H., M.R.I., M.F.A., M.A. and J.H.; formal analysis:
M.M.H., M.R.I., M.F.A., M.A. and J.H.; investigation: M.M.H., M.R.I., M.F.A., M.A. and J.H.; data
curation: M.M.H., M.R.I. and M.F.A.; writing—original draft preparation: M.M.H., M.R.I. and M.F.A.;
writing—review and editing: M.M.H., M.R.I., M.F.A., M.A. and J.H.; supervision: M.R.I., M.A. and
J.H. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data is contained within the article.
Conflicts of Interest: The authors declare no conflicts of interest.
Technologies 2024, 12, 151 26 of 28

Abbreviations
The following table alphabetically lists all the acronyms along with their definitions.
Acronym Stands for
ANN Artificial Neural Network
CAD Computer-aided Diagnosis
CNN Convolutional Neural Network
CT Computed Tomography
colon_aca Colon Adenocarcinoma
colon_bnt Colon Benign
DaaS Data as a Service
DL Deep Learning
DT Decision Tree
ELM Extreme Learning Machine
FL Federated Learning
FedAvg Federated Averaging
GDPR General Data Protection Regulation
HIPAA Health Insurance Portability and Accountability Act
IID Independent and Identically Distributed
IoMT The Internet of Medical Things
LCC Large Cell Carcinoma
lung_aca Lung Adenocarcinoma
lung_bnt Lung Benign
lung_scc Lung Squamous Cell Carcinoma
MRI Magnetic Resonance Imaging
NSCLC Non-small Cell Lung Cancer
RF Random Forest
SCC Squamous Cell Carcinoma
SGD Stochastic Gradient Descent
TL Transfer Learning
WHO World Health Organization
XAI Explainable Artificial Intelligence
XGBoost Extreme Gradient Boosting

References
1. Bray, F.; Laversanne, M.; Sung, H.; Ferlay, J.; Siegel, R.L.; Soerjomataram, I.; Jemal, A. Global cancer statistics 2022: GLOBOCAN
estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA A Cancer J. Clin. 2024, 74, 229–263. [CrossRef]
[PubMed]
2. Cancer Today. Available online: https://gco.iarc.fr/today/online-analysis-pie?v=2020&mode=cancer&mode_population=
continents&population=900&populations=900&key=total&sex=0&cancer=39&type=0&statistic=5&prevalence=0&population_
group=0&ages_group[]=0&ages_group[]=17&nb_items=7&group_cancer=1&include_nmsc=1&include_nmsc_other=1&half_
pie=0&donut=0 (accessed on 13 January 2024).
3. Xi, Y.; Xu, P. Global colorectal cancer burden in 2020 and projections to 2040. Transl. Oncol. 2021, 14, 101174. [CrossRef] [PubMed]
4. Cancer. Available online: https://www.who.int/news-room/fact-sheets/detail/cancer (accessed on 13 January 2024).
5. N.A. and R.A. Office of the Federal Register. Public Law 104-191-Health Insurance Portability and Accountability Act of
1996. govinfo.gov, August 1996. Available online: https://www.govinfo.gov/app/details/PLAW-104publ191 (accessed on 13
January 2024).
6. I (Legislative Acts) Regulations Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016
on the Protection of Natural Persons with Regard to the Processing of Personal Data and on the Free Movement of Such
Data, and Repealing Directive 95/46/EC (General Data Protection Regulation) (Text with EEA Relevance)’. Available online:
https://eur-lex.europa.eu/eli/reg/2016/679/oj (accessed on 10 January 2024).
7. Masud, M.; Sikder, N.; Nahid, A.-A.; Bairagi, A.K.; Al Zain, M.A. A Machine Learning Approach to Diagnosing Lung and Colon
Cancer Using a Deep Learning-Based Classification Framework. Sensors 2021, 21, 748. [CrossRef]
8. Chehade, A.H.; Abdallah, N.; Marion, J.-M.; Oueidat, M.; Chauvet, P. Lung and colon cancer classification using medical imaging:
A feature engineering approach. Phys. Eng. Sci. Med. 2022, 45, 729–746. [CrossRef]
9. Mangal Engineerbabu, S.; Chaurasia Engineerbabu, A.; Khajanchi, A. Convolution Neural Networks for Diagnosing Colon and
Lung Cancer Histopathological Images. September 2020. Available online: https://arxiv.org/abs/2009.03878v1 (accessed on 1
February 2024).
Technologies 2024, 12, 151 27 of 28

10. Hadiyoso, S.; Aulia, S.; Irawati, I.D. Diagnosis of lung and colon cancer based on clinical pathology images using convolutional
neural network and CLAHE framework. Int. J. Appl. Sci. Eng. 2023, 20, 1–7. [CrossRef]
11. You, C.; Zhao, R.; Liu, F.; Dong, S.; Chinchali, S.; Topcu, U.; Staib, L.; Duncan, J. Class-aware adversarial transformers for medical
imagesegmentation. Adv. Neural Inf. Process. Syst. 2022, 35, 29582–29596. [PubMed]
12. You, C.; Xiang, J.; Su, K.; Zhang, X.; Dong, S.; Onofrey, J.; Staib, L.; Duncan, J.S. Incremental Learning Meets Transfer Learning:
Application to Multi-Site Prostate MRI Segmentation; Springer: Cham, Switzerland, 2022; Volume 13573, pp. 3–16. [CrossRef]
13. Mehmood, S.; Ghazal, T.M.; Khan, M.A.; Zubair, M.; Naseem, M.T.; Faiz, T.; Ahmad, M. Malignancy Detection in Lung and
Colon Histopathology Images Using Transfer Learning with Class Selective Image Processing. IEEE Access 2022, 10, 25657–25668.
[CrossRef]
14. Toğaçar, M. Disease type detection in lung and colon cancer images using the complement approach of inefficient sets. Comput.
Biol. Med. 2021, 137, 104827. [CrossRef]
15. Kumar, N.; Sharma, M.; Singh, V.P.; Madan, C.; Mehandia, S. An empirical study of handcrafted and dense feature extraction
techniques for lung and colon cancer classification from histopathological images. Biomed. Signal Process. Control 2022, 75, 103596.
[CrossRef]
16. Talukder, A.; Islam, M.; Uddin, A.; Akhter, A.; Hasan, K.F.; Moni, M.A. Machine learning-based lung and colon cancer detection
using deep feature extraction and ensemble learning. Expert Syst. Appl. 2022, 205, 117695. [CrossRef]
17. Al-Jabbar, M.; Alshahrani, M.; Senan, E.M.; Ahmed, I.A. Histopathological Analysis for Detecting Lung and Colon Cancer
Malignancies Using Hybrid Systems with Fused Features. Bioengineering 2023, 10, 383. [CrossRef]
18. Ananthakrishnan, B.; Shaik, A.; Chakrabarti, S.; Shukla, V.; Paul, D.; Kavitha, M.S. Smart Diagnosis of Adenocarcinoma Using
Convolution Neural Networks and Support Vector Machines. Sustainability 2023, 15, 1399. [CrossRef]
19. You, C.; Zhao, R.; Staib, L.H.; Duncan, J.S. Momentum Contrastive Voxel-Wise Representation Learning for Semi-Supervised Volumetric
Medical Image Segmentation; Springer: Cham, Switzerland, 2022; Volume 13434, pp. 639–652. [CrossRef]
20. You, C.; Zhou, Y.; Zhao, R.; Staib, L.; Duncan, J.S. SimCVD: Simple Contrastive Voxel-Wise Representation Distillation for
Semi-Supervised Medical Image Segmentation. IEEE Trans. Med. Imaging 2022, 41, 2228–2237. [CrossRef] [PubMed]
21. You, C.; Dai, W.; Min, Y.; Staib, L.; Duncan, J.S. Bootstrapping Semi-Supervised Medical Image Segmentation with Anatomical-Aware
Contrastive Distillation; Springer: Cham, Switzerland, 2022; Volume 13939, pp. 641–653. [CrossRef]
22. You, C.; Dai, W.; Min, Y.; Liu, F.; Clifton, D.A.; Zhou, S.K.; Staib, L.; Duncan, J.S. Rethinking Semi-Supervised Medical Image
Segmentation: A Variance-Reduction Perspective. Adv. Neural Inf. Process. Syst. 2023, 36, 9984–10021.
23. You, C.; Dai, W.; Min, Y.; Staib, L.; Sekhon, J.; Duncan, J.S. ACTION++: Improving Semi-Supervised Medical Image Segmentation with
Adaptive Anatomical Contrast; Springer: Cham, Switzerland, 2023; Volume 14223, pp. 194–205. [CrossRef]
24. Konečn, J.K.; Brendan, H.; Google, M.; Google, D.R.; Richtárik, P. Federated Optimization: Distributed Machine Learning for
On-Device Intelligence. October 2016. Available online: https://arxiv.org/abs/1610.02527v1 (accessed on 1 February 2024).
25. Konečn, J.; McMahan, H.B.; Yu, F.X.; Suresh, A.T.; Google, D.B.; Richtárik, P. Federated Learning: Strategies for Improving
Communication Efficiency. October 2016. Available online: https://arxiv.org/abs/1610.05492v2 (accessed on 1 February 2024).
26. Roth, H.R.; Chang, K.; Singh, P.; Neumark, N.; Li, W.; Gupta, V.; Gupta, S.; Qu, L.; Ihsani, A.; Bizzo, B.C.; et al. Federated Learning
for Breast Density Classification: A Real-World Implementation; Springer: Cham, Switzerland, 2020; Volume 12444, pp. 181–191.
[CrossRef]
27. Florescu, L.M.; Streba, C.T.; Şerbănescu, M.-S.; Mămuleanu, M.; Florescu, D.N.; Teică, R.V.; Nica, R.E.; Gheonea, I.A.; Florescu,
L.M.; Streba, C.T.; et al. Federated Learning Approach with Pre-Trained Deep Learning Models for COVID-19 Detection from
Unsegmented CT images. Life 2022, 12, 958. [CrossRef]
28. Hossain, M.; Ahamed, F.; Islam, R.; Imam, R. Privacy Preserving Federated Learning for Lung Cancer Classification. In
Proceedings of the 2023 26th International Conference on Computer and Information Technology, ICCIT 2023, Cox’s Bazar,
Bangladesh, 13–15 December 2023. [CrossRef]
29. Zhang, W.; Zhou, T.; Lu, Q.; Wang, X.; Zhu, C.; Sun, H.; Wang, Z.; Lo, S.K.; Wang, F.-Y. Dynamic-Fusion-Based Federated Learning
for COVID-19 Detection. IEEE Internet Things J. 2021, 8, 15884–15891. [CrossRef] [PubMed]
30. Khan, T.A.; Fatima, A.; Shahzad, T.; Rahman, A.U.; Alissa, K.; Ghazal, T.M.; Al-Sakhnini, M.M.; Abbas, S.; Khan, M.A.; Ahmed, A.
Secure IoMT for Disease Prediction Empowered with Transfer Learning in Healthcare 5.0, the Concept and Case Study. IEEE
Access 2023, 11, 39418–39430. [CrossRef]
31. Peyvandi, A.; Majidi, B.; Peyvandi, S.; Patra, J.C. Privacy-preserving federated learning for scalable and high data quality
computational-intelligence-as-a-service in Society 5.0. Multimed. Tools Appl. 2022, 81, 25029–25050. [CrossRef] [PubMed]
32. Borkowski, A.A.; Bui, M.M.; Thomas, L.B.; Wilson, C.P.; DeLand, L.A.; Mastorides, S.M. Lung and Colon Cancer Histopathological
Image Dataset (LC25000). December 2019. Available online: https://arxiv.org/abs/1912.12142v1 (accessed on 1 February 2024).
33. Bhimji, S.S.; Wallen, J.M. Lung Adenocarcinoma. StatPearls, June 2023. Available online: https://www.ncbi.nlm.nih.gov/books/
NBK519578/ (accessed on 1 February 2024).
34. Walser, T.; Cui, X.; Yanagawa, J.; Lee, J.M.; Heinrich, E.; Lee, G.; Sharma, S.; Dubinett, S.M. Smoking and Lung Cancer: The Role
of Inflammation. Proc. Am. Thorac. Soc. 2008, 5, 811–815. [CrossRef]
35. Ma, Z.; Zhang, M.; Liu, J.; Yang, A.; Li, H.; Wang, J.; Hua, D.; Li, M. An Assisted Diagnosis Model for Cancer Patients Based on
Federated Learning. Front. Oncol. 2022, 12, 860532. [CrossRef]
Technologies 2024, 12, 151 28 of 28

36. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the Inception Architecture for Computer Vision. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016;
pp. 2818–2826. [CrossRef]
37. Simonyan, K.; Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. 2015. Available online:
http://www.robots.ox.ac.uk/ (accessed on 17 February 2024).
38. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. Available online: http://image-net.org/
challenges/LSVRC/2015/ (accessed on 17 February 2024).
39. Xie, S.; Girshick, R.; Dollár, P.; Tu, Z.; He, K. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings
of the 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, 21–26 July 2017;
pp. 5987–5995. [CrossRef]
40. Chollet, F. Xception: Deep Learning with Depthwise Separable Convolutions. In Proceedings of the 30th IEEE Conference on
Computer Vision and Pattern Recognition, CVPR, Honolulu, HI, USA, 21–26 July 2017; pp. 1800–1807. [CrossRef]
41. Ribeiro, M.; Singh, S.; Guestrin, C. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In Proceedings of the
2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, San Diego,
CA, USA, 12–17 June 2016; pp. 97–101. [CrossRef]
42. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks
via Gradient-based Localization. Int. J. Comput. Vis. 2016, 128, 336–359. [CrossRef]
43. Tasnim, Z.; Chakraborty, S.; Shamrat, F.M.J.M.; Chowdhury, A.N.; Alam Nuha, H.; Karim, A.; Zahir, S.B.; Billah, M. Deep Learning
Predictive Model for Colon Cancer Patient using CNN-based Classification. Int. J. Adv. Comput. Sci. Appl. 2021, 12. [CrossRef]
44. Shandilya, S.; Nayak, S.R. Analysis of Lung Cancer by Using Deep Neural Network; Springer: Singapore, 2022; Volume 814,
pp. 427–436. [CrossRef]
45. Karim, D.Z.; Bushra, T.A. Detecting Lung Cancer from Histopathological Images using Convolution Neural Network. In
Proceedings of the IEEE Region 10 Annual International Conference, Proceedings/TENCON, Auckland, New Zealand, 7–10
December 2021; pp. 626–631. [CrossRef]
46. Raju, M.S.N.; Rao, B.S. Lung and colon cancer classification using hybrid principle component analysis network-extreme learning
machine. Concurr. Comput. Pr. Exp. 2022, 35, e7361. [CrossRef]
47. Ren, Z.; Zhang, Y.; Wang, S. A Hybrid Framework for Lung Cancer Classification. Electronics 2022, 11, 1614. [CrossRef]
48. Attallah, O.; Aslan, M.F.; Sabanci, K. A Framework for Lung and Colon Cancer Diagnosis via Lightweight Deep Learning Models
and Transformation Methods. Diagnostics 2022, 12, 2926. [CrossRef] [PubMed]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.