Lung Cancer Detection Using Deep Learning and

Explainable Methods
Ayah Alomar*,Moayed Alazzam, Hala Mustafa, Ahmad Mustafa
Faculty of Computer and information technology, Jordan university of science and technology, Irbid, Jordan
{afalomar20, maazzam20, htmustafa20}@cit.just.edu.jo, [email protected]
2023 14th International Conference on Information and Communication Systems (ICICS) | 979-8-3503-0786-3/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICICS60529.2023.10330443

Abstract—Lung cancer is one of the most prevalent deadly Explainable artificial intelligence, or XAI, is a branch of
diseases and it can extend to the rest of the human body. One way AI that focuses on creating algorithms and systems that can
to detect it in CT scan images is by using deep learning models. transparently communicate their decision-making processes
Explaining these models by XAI techniques and radiologists make
the results trusted for medical use. In this paper, the deep learning and how they came to a certain result [5]. In order to increase
models inceptionV3 and ResNet50 were used to classify CT scans the accuracy of the results and detect any biases or flaws
of lungs for the presence of cancer. The models were trained on a in the algorithm’s decision-making process, this can be
Kaggle dataset that is pre-processed and augmented. It was able especially helpful for medical image analysis [6].
to accurately detect lung cancer in new patients. Additionally, an
XAI model was used to explain the decision-making process of
the deep learning model, providing insights into which features II. R ELATED WORKS
of the CT scan were most important for the model’s diagnosis.
ResNet50 achieved the highest performance with 100% accuracy Due to the severity and effects it has on people’s lives, a
on testing images, and InceptionV3 got 99.92%. LIME and number of study articles have recently been published that
GRAD-CAM explained the model’s performance by highlighting address the problem of lung cancer [7]. Therefore, publications
the most important features for each model. On the other hand,
radiologists provided insight on whether the deep learning and are examined in order to accomplish this study goal because
XAI models are correct by determining the cancer region in each the concepts for these studies based on image analysis come
image and point out the model’s misclassification. The results from a variety of sources, including transfer learning and deep
show that the radiologist’s diagnosis is important even though learning [8].
the model’s accuracy is high. Patrı́cio et al. [9] aim to detect lung cancer by analyzing
Index Terms—Deep learning, Artificial Intelligence, Explain-
ability, lung cancer, CT scan, radiologist’s diagnosis. pathology images during the widespread of using them in
treatment as a routine procedure, and they sought to give a
general review of the current and future uses of AI techniques
I. I NTRODUCTION
in lung cancer. They began by outlining the potential and
One type of cancer that develops in the lungs is lung cancer problems that exist today for lung cancer pathology image,
[1]. It is more deadly than breast, colon, and prostate cancers then went over recent advances in deep learning that may have
combined, making it the primary cause of cancer deaths an effect on lung cancer digital pathology, authors concluded
globally. Non-small cell lung cancer (NSCLC) and small cell by summarizing the current uses of deep learning models in
lung cancer are the two main subtypes of lung cancer (SCLC). the diagnosis and prognosis of lung cancer.
The risk factors for lung cancer include smoking, exposure to Regarding to the methods and material, they applied deep
certain chemicals and pollutants, and a family history of the learning models especially the supervised learning CNNs were
disease. NSCLC is the more prevalent variety and tends to used because it proved their importance in the pathology
grow and spread more slowly than SCLC. Avoiding tobacco images classification field, lung cancer, head cancer, and
use and limiting exposure to other recognized risk factors are others diseases. The deep learning models have mainly two
the best ways to prevent lung cancer [2]. characteristics, it facilitates feature extraction from images
A CT (computed tomography) scan is a type of imaging because the model does it automatically and it has many
testing that creates finely detailed images of the inside of the layers and kernel which is mean it can deal with any complex
body using specialist X-ray equipment [3]. It can be used function. For training and testing, the dataset they used transfer
to find lung cancer and establish the tumor’s stage (how learning. They applied the deep learning model on a dataset
advanced it is), Since a CT scan of the chest can give precise image with size 300X300 pixels.
images of the lungs and adjacent tissues, it is frequently used Walia et al, [10] stated that the computer-aided diagnosis
to diagnose lung cancer. can be extremely important in the early diagnosis of lung
Overall, the use of deep learning with CT scans has the cancer, which is a fundamental step for the treatment of
potential to significantly improve the accuracy and efficiency lung cancer. Most CAD approaches that have been described
of lung cancer detection, by automating the analysis of classify each lung module separately to diagnose lung cancer.
medical images and identifying patterns that may be missed However, does not correspond with actual practice, in which
by human interpretation [4]. doctors identify patients by comparing a collection of module

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on September 16,2024 at 18:04:14 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
DATASET SPLITTING

Training set Testing set Validation set

Lung cancer 591 84 169
Normal 591 84 169
Total 1182 168 338

first step is to resize the images to a consistent size, this

is to ensure that the model’s input is of the same size and
shape, the second step is to convert all images to the same
file extension. The consistent images format ensures that the
images can be loaded and processed by the deep learning
model without any compatibility issues [14]. The third step
is data augmentation techniques that is applied to the dataset,
this is to artificially increase the size of the dataset by applying
various transformations to the images such as rotation, and
Fig. 1. The difference between the usual workflow of deep learning and our
approach
zooming. These techniques can help the model to generalize
better and improve it is performance, the dataset increased
from 1000 images into 1688 images as the table Ishown.

photos rather than focusing on one at a time. A significant C. Transfer Learning

obstacle to their acceptance is the limited interpretability of ResNet50[15] and InceptionV3[16] are both convolutional
the results these approaches provide, the attention mechanism neural networks (CNNs) that have been trained on the Ima-
offers improved interpretability, the results indicate that the geNet dataset for image classification tasks. ResNet50 is a 50-
approach can produce accuracy of 0.807. layer deep network that uses residual connections to improve
Looking at the strengths and weaknesses of the presented the flow of information through the network and reduce the
papers, our work aims to study the detection of lung cancer vanishing gradient problem. InceptionV3, on the other hand,
using deep learning techniques. Moreover, the previous work uses a modular architecture called Inception, which allows for
focused on the accuracy without considering the experts opin- the network to have a wider receptive field and capture more
ion. We show that the model may focus on wrong part of the information from the input image [17].
image even though it might be achieving high accuracy. So,
D. Explainable Machine Learning
in this research the radiologists defined the important features
in lung cancer images. The predictions of a deep learning model on an image
dataset of lung cancer CT scans were understood and inter-
III. M ETHODOLOGY preted using two well-known explainable machine learning
A. Overview approaches, Grad-CAM [18] and LIME [19].
Grad-CAM (Gradient-weighted Class Activation Mapping)
Using deep learning and XAI to detect lung cancer typi- is a method for producing heat maps that highlight the areas of
cally involves a combination of data collection, preprocessing, an image that are most crucial for the model’s prediction, the
model development, evaluation, and interpretation. The first heat map that was produced reveals which parts of the image
step is collecting a dataset of chest CT scans from Kaggle with were most important for the model’s prediction and can be
and without lung cancer [11], the data is then preprocessed used to determine which attributes the model is employing,
to make it suitable for deep learning, which includes resizing this helps on spot any possible biases or inaccuracies in the
the images, normalizing, and applying data augmentation tech- model’s predictions.
niques [12]. The next step is to develop and train deep learning Another approach called LIME (Local Interpretable Model-
models ResNet50 and InceptionV3 on the preprocessed data agnostic Explanations) produces explanations for a model’s
to classify whether a given CT scan is from a patient with predictions by approximating the model locally around the
lung cancer or not. The model is then evaluated using various input. A simpler model is trained on the affected images by
metrics such as accuracy, precision, recall, and F1 score, and sampling the input space surrounding the input image. The
XAI techniques are used to interpret the model’s decision- explanation for the prediction made by the original model is
making process and identify any potential biases or errors. then produced using the simpler model. LIME can offer a
Finally, the model is given to radiologists to determine the more thorough justification of how the model generates its
detection of lung cancer is accurate or not. prediction.
B. Dataset preprocessing and splitting E. Experiments and evaluation metrics
Preprocessing of images in a lung cancer dataset is a crucial There are many hyper parameters that modified to improve
step in making the data suitable for deep learning [13]. The the performance of the model when experimenting in transfer

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on September 16,2024 at 18:04:14 UTC from IEEE Xplore. Restrictions apply.
learning for images datasets [20], the learning rate decay
is a crucial hyper parameter that used to gradually lower
the learning rate, preventing the model from overfitting by
trying two values .001 or. 0001, the optimizer Adam is also
one of the important hyper parameters and a regularization
method called dropout removes certain neurons by random
during training to avoid overfitting used with values .5 or
.2. The amount of training samples used in each iteration of Fig. 2. The different between radiologist and LIME diagnosis
the training process is controlled by the batch size, the final
factor controlling how many times the model is trained on
the full dataset is the number of epochs and 100 epochs is technique creates a heat map that highlights the crucial areas
applied with using the early stopping technique [21]. In order of the input image that are used to make a prediction by
to evaluate the performance of the model, metrics such as emphasizing the areas of the image that the deep learning
highest validation accuracy or lowest validation loss as well model is concentrating on, it is used to understand how the
as common evaluation metrics for image classification include algorithm makes decisions. LIME and Grad-CAM can give
accuracy, precision, recall, and F1-score is used [22]. more understanding of how the model generates its predic-
In other words, by altering the parameters that are particular to tions, combining the data from the two ways makes it feasible
the task, dataset, and computational resources [23], applying to pinpoint not just the image’s key elements but also the
hyper parameters in transfer learning on image datasets allow precise areas that the model is concentrating on. This can offer
fine-tune the model and make the most of the pre-trained important insights into the model’s underlying mechanics,
model. As a result, the model may perform better and gen- support the creation of improved models for classifying CT
eralize more well to new data by finding the best combination scan images for lung cancer, and improve the comprehension
of hyper parameters that work well for the specific dataset and of the diagnostic procedure.
task [24].
C. Radiologists Diagnosis
IV. R ESULTS A CT scan image for lung cancer is classified using a
A. Transfer Learning deep-learning model called ResNet50. The model is trained
The accuracy of deep learning model is significantly im- to identify characteristics of lung cancer in the CT scan
pacted by the choice of hyper parameters. It is essential to image and classify it as either cancer or normal. by using
compare the results in order to choose the best model and to ResNet50, the classification results are accurate without any
select one model that has the best overall performance and is misclassification, but in this part, it is important to know
most suitable for the task of lung cancer detection. Factors like if the model trained the most important features or not by
the number of layers in the network, the number of neurons taking randomly images and applying the LIME and Grad-
in each layer, and the learning rate can all affect the model’s cam to compare their explanation with radiologists explain,
performance as shown in the table II. in ‘Fig.8’ the original images with square mark determined
by radiologists to focus on the cancer cells on CT images, in
TABLE II LIME image the model focus on different area that does not
M ODELS B EST R ESULTS related to cancer cells.
Training Training Validation Validation Testing Testing ResNet50 and InceptionV3 models were give a high accuracy
Model
Accuracy Loss Accuracy Loss Accuracy Loss in classifying lung cancer images but were found focused on
Inception the wrong features (areas) in some images, indicating that the
99.92 .01 99.1 0.03 99.2 .03
0.001 N model is not properly detecting the characteristics of cancer.
D 0.5 64
This leads to false positives or false negatives, which can have
ResNet
50 0.0001
100 0.0 100 0.0 100 0.0 serious consequences in a clinical setting.
D 0.2 128 In this scenario, it is important to re-evaluate the model’s
performance and identify the reasons for why it is focused on
the wrong features. Additionally, radiologists play a crucial
B. Explainable Machine Learning role in identifying the problem and providing feedback on the
LIME and Grad-CAM combined to provide a more ex- model’s performance and which features are most important
planation of how the model is classifying images as either for accurate diagnosis. By addressing the problem of focusing
cancerous or normal in a ResNet50 and InceptionV3 model for on the wrong features, such as the bad image resolution can
CT scan image analysis of lung cancer. In order to determine be a contributing factor to a deep learning model focusing
which features of the image are most crucial for making the on the wrong features when classifying lung cancer images.
diagnosis, LIME creates heat maps that highlight the areas When an image has low resolution, it can be more difficult
of the image that have the most influence on the model’s for the model to distinguish important features such as tumors
prediction. On the other hand, the Grad-CAM visualization or nodules, which can lead to the model focusing on other,

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on September 16,2024 at 18:04:14 UTC from IEEE Xplore. Restrictions apply.
less relevant features instead. also, low resolution images can [5] J. U. Lim, H. S. Kang, A. Y. Shin, C. D. Yeo, C. K. Park, S. H. Lee, S. J.
result in a loss of important details and textures, which can Kim, and K. C. C. R. Korean Association for Lung Cancer, “Association
between clinical outcomes and local treatment in stage iv non-small
further complicate the model’s ability to accurately identify cell lung cancer patients with single extrathoracic metastasis,” Thoracic
and classify cancerous regions. Cancer, vol. 13, no. 9, pp. 1349–1360, 2022.
[6] W. Jiang, G. Zeng, S. Wang, X. Wu, C. Xu, et al., “Application of
V. D ISCUSSION deep learning in lung cancer imaging diagnosis,” Journal of Healthcare
Engineering, vol. 2022, 2022.
Using deep learning models such as ResNet50 and In- [7] D. S. Jeon, J. W. Kim, S. G. Kim, H. R. Kim, S. Y. Song, J. C. Lee,
ceptionV3 to classify lung cancer images, along with inter- W. Ji, C.-M. Choi, H. C. Kim, and S. K. C. C. R. Korean Association for
Lung Cancer, “Sex differences in the characteristics and survival of
pretation methods LIME and Grad-CAM, helps to identify patients with non-small-cell lung cancer: A retrospective analytical study
important features in the images that are indicative of cancer. based on real-world clinical data of the korean population,” Thoracic
However, it is important to evaluate the model’s performance Cancer, vol. 13, no. 18, pp. 2584–2591, 2022.
[8] Q. Wang, Y. Zhou, J. Huang, Z. Liu, W. Zhang, Q. Liu, and J.-Z. Cheng,
by radiologists to ensure that it is focusing on the correct “Cascaded-recalibrated multiple instance deep model for pathologic-
features and not being influenced by irrelevant or confounding level lung cancer prediction in ct images,” Computational Intelligence
factors. By involving radiologists in the evaluation process, and Neuroscience, vol. 2022, 2022.
[9] C. Patrı́cio, J. C. Neves, and L. F. Teixeira, “Explainable deep
they provide valuable insights and feedback on the model’s learning methods in medical diagnosis: a survey,” arXiv preprint
performance and help to identify any areas where it may be arXiv:2205.04766, 2022.
lacking. Additionally, interpretation methods such as LIME [10] S. Walia, K. Kumar, S. Agarwal, and H. Kim, “Using xai for deep
learning-based image manipulation detection with shapley additive ex-
and Grad-CAM can be useful in understanding the model’s planation,” Symmetry, vol. 14, no. 8, p. 1611, 2022.
decision-making process and identifying which features are [11] Z. Hussain, F. Gimenez, D. Yi, and D. Rubin, “Differential data
most important in the classification of lung cancer images. augmentation techniques for medical imaging classification tasks,” in
AMIA annual symposium proceedings, vol. 2017. American Medical
This can help to improve the model’s overall performance Informatics Association, 2017, p. 979.
and ensure that it is focused on the correct features for [12] D. Ray, O. Pinti, and A. A. Oberai, “Deep learning and computational
accurate diagnosis, more examples are available on our GitHub physics (lecture notes),” arXiv preprint arXiv:2301.00942, 2023.
[13] P. Balaji, R. Aluvalu, and K. Sagar, “Residual attention network based
repository [25]. hybrid convolution network model for lung cancer detection,” Intelligent
Decision Technologies, no. Preprint, pp. 1–14.
VI. C ONCLUSION AND FUTURE WORKS [14] D. Kaul, H. Raju, and B. Tripathy, “Deep learning in healthcare,”
Deep Learning in Data Analytics: Recent Techniques, Practices and
Deep learning models ResNet50 and Inceptionv3 are em- Applications, pp. 97–115, 2022.
ployed to classify lung CT images. ResNet50 achieved 100% [15] R. Zhang, Y. Zhu, Z. Ge, H. Mu, D. Qi, and H. Ni, “Transfer learning
for leaf small dataset using improved resnet50 network with mixed
testing accuracy and InceptionV3 achieved 99.2%. The out- activation functions,” Forests, vol. 13, no. 12, p. 2072, 2022.
comes of these deep learning models are also interpreted using [16] N. S. Shadin, S. Sanjana, and N. J. Lisa, “Covid-19 diagnosis from chest
XAI models like Grad-CAM and LIME, which explained the x-ray images using convolutional neural network (cnn) and inceptionv3,”
in 2021 International Conference on Information Technology (ICIT).
model’s prediction. These techniques can be used to improve IEEE, 2021, pp. 799–804.
the understanding and interpretability of the deep learning [17] A. Odeh, A. Alomar, and S. Aljawarneh, “Detection of covid-19 using
models used to classify lung CT images, also the physicians deep learning on x-ray lung images,” PeerJ Computer Science, vol. 8,
p. e1082, 2022.
provide feedback on whether the model’s classifications are in [18] M. Lerma and M. Lucas, “Grad-cam++ is equivalent to grad-cam with
line with their own observations and clinical experience and positive gradients,” arXiv preprint arXiv:2205.10838, 2022.
identifying which regions of the image the model is focusing [19] J. Dieber and S. Kirrane, “Why model why? assessing the strengths and
limitations of lime,” arXiv preprint arXiv:2012.00093, 2020.
on when making it is predictions. Overall, the collaboration [20] J. Plested and T. Gedeon, “Deep transfer learning for image classifica-
between physicians, deep learning, and XAI models can lead tion: a survey,” arXiv preprint arXiv:2205.09904, 2022.
to more accurate and reliable predictions, and ultimately, better [21] R. Mohakud and R. Dash, “Skin cancer image segmentation utilizing a
novel en-gwo based hyper-parameter optimized fcedn,” Journal of King
patient care. For future work we plan to use more XAI models Saud University-Computer and Information Sciences, vol. 34, no. 10,
to interpret images classification to be more trusted for medical pp. 9889–9904, 2022.
fields. [22] H. Wang, T. Li, Z. Zhuang, T. Chen, H. Liang, and J. Sun, “Early
stopping for deep image prior,” arXiv preprint arXiv:2112.06074, 2021.
[23] G. Shao, L. Tang, and H. Zhang, “Introducing image classification
R EFERENCES efficacies,” IEEE Access, vol. 9, pp. 134 809–134 816, 2021.
[1] K. L. Kohsasih and B. H. Hayadi, “Classification sars-cov-2 disease [24] C. Chen, P. Zhang, H. Zhang, J. Dai, Y. Yi, H. Zhang, and Y. Zhang,
based on ct-scan image using convolutional neural network,” Scientific “Deep learning on computational-resource-limited platforms: a survey,”
Journal of Informatics, vol. 9, no. 2, pp. 197–204, 2022. Mobile Information Systems, vol. 2020, pp. 1–19, 2020.
[2] K. Ramana, M. R. Kumar, K. Sreenivasulu, T. R. Gadekallu, S. Bha- [25] ayah989, “Ayah989/-lung-cancer-detection-using-deep-learning-and-
tia, P. Agarwal, and S. M. Idrees, “Early prediction of lung cancers explainable-methods.” [Online]. Available: https://github.com/ayah989/
using deep saliency capsule and pre-trained deep learning frameworks,” -Lung-Cancer-Detection-Using-Deep-Learning-and-Explainable-Methods.
Frontiers in oncology, vol. 12, p. 886739, 2022.
[3] G. Ren, B. Li, S.-k. Lam, H. Xiao, Y.-H. Huang, A. L.-y. Cheung, Y. Lu,
R. Mao, H. Ge, F.-M. S. Kong, et al., “A transfer learning framework for
deep learning-based ct-to-perfusion mapping on lung cancer patients,”
Frontiers in Oncology, vol. 12, p. 883516, 2022.
[4] S. Wang, D. M. Yang, R. Rong, X. Zhan, J. Fujimoto, H. Liu, J. Minna,
I. I. Wistuba, Y. Xie, and G. Xiao, “Artificial intelligence in lung cancer
pathology image analysis,” Cancers, vol. 11, no. 11, p. 1673, 2019.

Authorized licensed use limited to: VIT-Amaravathi campus. Downloaded on September 16,2024 at 18:04:14 UTC from IEEE Xplore. Restrictions apply.