This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 1

An Open IoHT-based Deep Learning Framework

for Online Medical Image Recognition
Carlos M. J. M. Dourado Jr, Suane Pires P. da Silva, Raul Victor M. da Nóbrega, Pedro P. Rebouças Filho, Khan
Muhammad, Member, IEEE, Victor Hugo C. de Albuquerque, Senior Member, IEEE .
Abstract—Systems developed to work with computational Internet, which when combined with the development of new
intelligence have become very efficient, and in some cases obtain software techniques allowed a much wider range of usages, not
more accurate results than evaluations by humans. Hence, this only for entertainment, but also for the treatment and
work proposes a new online approach based on deep learning tools
according to the concept of transfer learning to generate a monitoring of health needs, for surveillance of environments,
computational intelligence framework for use with the Internet of and for home automation, among others [3].
Health Things (IoHT) devices. This framework allows the user to Another facility added to this scenario is cloud computing,
add their images and perform platform training almost as easily which provides computing power and storage availability from
as creating folders and placing files in regular cloud storage any location with an Internet connection. This increases the
services. The trials carried out with the tool showed that even
people with no programming and image processing knowledge variability of available resources, depending on the need for an
were able to set up projects in a few minutes. The proposed application at relatively low prices and proportional to use,
approach is validated using three medical databases, which which makes these features even more accessible [4].
include cerebral vascular accident images for stroke type An exploratory research in the literature found various
classification, lung nodule images for malignant classification, and studies, as well as the development of applications specifically
skin images for the classification of melanocytic lesions. The
results show the efficiency and reliability of the framework, which for the medical domain, such as patient monitoring. Among the
reached 91.6% Accuracy in the stroke images and lung nodules studies found, in [5], the detection and classification of stroke
databases, and 92% Accuracy in the skin images databases. This is performed with the aid of a feature extraction method entitled
prove the immense contribution that this work can bring to assist Analysis of Brain Tissue Density, which is based on brain
medical professionals in analyzing complex examinations quickly radiological density patterns. Regarding the Smart IoT medical
and accurately, allowing a large medical examination database
through a consolidated collaborative IoT platform. systems, in [6], an intelligent IoT system is provided that
provides, interprets, and stores patient gait pattern data, using
Index Terms—IoT, imaging processing, deep learning, machine Inertial Measurement Unit, to monitor people with diabetes, for
learning, computational intelligence.
example, a disease that causes a decline in motor activity.
While at [7], a cardiac monitoring system using Raspberry Pi
I. I NTRODUCTION and IoT is developed. An Arduino- based microcontroller is
Technological development has enabled the construction of used for communication with sensors. All data analysis is
portable computing devices at low cost that are accessible to performed at Raspberry Pi. After signal processing, the results
users. Cellular devices have gained more technological are sent to the web server. In [8], a system based on cloud and
capabilities and increased processing power, and by integrating fog computing along with IoT and Big Data platforms is
them to mobile data networks, they have become a source for presented, with the aim of assisting in resolving sleep apnea-
reference, interaction and sharing. Nowadays, they have related limitations. In [9], a cloud- based IoT system to aid the
become almost indispensable in the day-to-day life of our medical diagnosis of Chronic Kidney Diseases is developed
culture. Surveys have shown that as of 2008 there are more employing Deep Neural Network. A real-time epilepsy
interconnected devices than the number of people on the planet detection system is proposed by [10], considering
and the expectation is for continued exponential growth [1], [2]. electroencephalogram signals and processed using Discrete
Technological breakthroughs have begun to link physical Wavelets Transform, Hjorth Parameters, Statistical
devices to the Internet. In 1999, the term Internet of Things (IoT) Characteristics and Deep Neural Network. And study that uses
was coined to define the use of devices and sensors that could machine learning, big data and IoT can be found on [11].
be accessed and/or controlled remotely [1]. The use of IoT Whether they are independent devices or are solutions that
started to fill the gap connecting physical devices to the monitor health variables via sensors (heart pressure, oxygena-
tion etc.) all acquired data can be integrated into a single
Corresponding authors: Khan Muhammad and V. H. C. de Albuquerque platform where their goal is to facilitate the prevention and
CMJMDJ and VHCA are with the University of Fortaleza, Brazil. e-mail:
[email protected], [email protected]. diagnosis of diseases. Also there are diagnostic assistance op-
SPS, RVMN and PPRF are with Instituto Federal do Ceará, Brazil. e- tions, some of which have been developed by large companies,
mail: [email protected], [email protected], pe- such as IBM with its high-cost Watson Health application, and
[email protected]
KM is with department of Software, Sejong University, Seoul, Republic of there is the use of social networking solutions for team and
Korea. e-mail: [email protected] patient monitoring [12]–[14] or analysis of images in social
media [15]. Finally, there are commercial solutions for the

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 2

Convolutional Neural Networks (CNN) were conceived by [20]

scanning of exams in order to enable remote diagnoses. and later consolidated by [21], more specifically, in the realization
In view of the current state of the art, the present study stands of ImageNet 2012 Large Scale Visual Recognition Challenge
out as it uses image processing with computational intelligence (ILSVRC-2012), when CNN achieved convincing results. The
to identify anomalies and thus returns diagnostic data in an goal of ILSVRC-2012 was to solve a classification task, which
environment that does not require specific knowledge of the consisted of 1000 classes and over 1 million samples [21].
technologies involved. This study is also innovative because it Depending on the purpose, it is impractical to collect a
integrates an image processing application that can be shared database to successfully train a CNN [22]. Thus, the concept of
through a platform similar to social media, aiming to intensify Transfer Learning appeared as an alternative to solve this
the use of more voluminous databases and consequently problem, being widely used in different applications [23]. Deep
improve the accuracy of the algorithms and methods used. This learning concepts are also applied and are very important in this
is an important step in the current cultural context of knowledge type of study [24], [25].
management, regulated by sharing, and allowing the exchange
of experiences and information among professionals. The III. T HE PROPOSED FRAMEWORK
proposal here gathers together several computational methods,
The tool developed by the team was named LINDA, which
with the differential of guaranteeing excellent usability and
is an acronym for Lapisco Image iNterface for the
promoting the intuitive use by lay users of informatics who
Development of Applications, and is made up of two
have no knowledge of image processing techniques.
subsystems. Lapisco is the name of the group responsibly by
Another indirect contribution is the support to create larger
development and is the portuguese acronym of Laboratory of
image databanks, within a scientific platform, with shared data
Image Processing, Signals and Applied Computing. The first,
of users that work in the same area of study. This will enable
developed in Java, implements a web service, which is
the improvement of the framework, by applying its algorithms
responsible for the exchange of information between the device
to much broader bases. Thus, it is expected that the platform,
and the cloud platform. The device is integrated to a free
specifically in the health domain, will contribute to imaging
relational database, Postgre SQL, which stores information
diagnosis, and provide health professionals with more accurate
about the available extractors and classifiers, images used for
diagnoses in less time.
system training, and a selection of the extractor and classifier
There is a noticeable challenge to implement this type of
pairs that present the best result for image processing as well
application, since it requires the combination of three elements
as the processing statistics and results. Also it has a friendly and
that are not usually integrated: computer knowledge, especially
responsive web interface that allows users to make
in image processing, medical knowledge and high
adjustments.
computational power [16]. Various research projects in medical
diagnostic imaging can be found that apply various algorithms The second subsystem of the device was developed in
and have very precise solutions, but without an interface to be Python and makes use of libraries such as Tensor Flow and
used by ordinary health professionals [17]–[19]. Keras to carry out the processing of the images sent via the
Given this, the proposed framework intends to overcome Convolutional Neural Networks (CNN). The use of these
these challenges by proposing a platform accessible to a user platforms ensures a constant implementation of the algorithms
with a basic knowledge of computer science but without Figure 1 shows the data flow of the LINDA system, which
specific knowledge in the diagnostic area. Then, as a user has two different ways. Consider that the image bases have
collaborates and shares images on this platform, the already been configured and the algorithms have been selected,
performance of the interface will improve and consequently the then the prediction API will be ready for use. The information
work of users can be optimized. The main contributions of this flow will begin by sending the image from a device (indicated
work are:
• allows the use of a Graphic Process Unit (GPU) for
IoHT-based image processing and analysis;
• develops a large database of medical exams using a
collaborative platform based on an IoT environment;
• provides a collaborative social network of IoHT-
based artificial intelligence techniques;
• improve medical diagnostics and try to avoid
mistakes;
• disseminate machine learning techniques, allowing
them to be used by people without specific knowledge;
• bring expert-level technology to geographically Fig. 1: LINDA system flow: 1) the same software can be used
distant, resource-poor locations. to send new images for training or prediction; 2) web
application allows users to configure system parameters; 3)
II. OVERVIEW OF C ONVOLUTIONAL N EURAL N ETWORK results of training are shown in graphs and tables, and, finally,4)
API prediction is responsible for image processing according to
the parameters configured on the web application.

2 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 3

Analytics, which is composed of the API prediction that per-

in Figure 1 by the number 1). Then it will suffice to send the forms attribute extraction, image classification and prediction,
image to be evaluated and the security hash code to the system. depending on the case. The third part is the Persistence, based
This will make the system call the prediction API, which will on the Postgre SQL database, which stores the images obtained,
select, according to the security hash, the algorithms to be the system configuration data and the processing projects, as
applied. Such algorithms will have been previously defined in well as their results. The fourth and final part is the
the platform configuration. Then, an adjustment will be made Interoperability, which ensures system integration with other
to the image to suit the processing standards, which gives the systems or devices. The safety features are incorporated in all
system flexibility to work with various types of cameras and parts of the system; however, the interoperability part has no
forms of image acquisition, without the need for additional interface.
work for the user or adjustments of systems that are integrated
via web services. The necessary models will be loaded into the
memory. Moreover, if a call for the same model has been made
recently, even if by another user or another project, the model
does not need to be reloaded; thus a gain of up to 60% of
performance maybe acquired in some cases. If the settings of
the system have not been done and some changes are desired
then the (indicated in Figure 1 by the number 2) web
application path is used, which allows the uploading and
categorization of images. The user may create as many classes
as he or she deems necessary to classify the images; for
example - 2 classes, one class for healthy lung images and
Fig. 2: The layers of the LINDA system: a) image acquisition
another class of images with nodules. An accurate analysis of
layer; b) analyzes images according to the defined parameters;
the images requires at least 2 classes with 10 images each. The
c) saves images and results in a safe database, and d) controls
platform has 10 algorithms for the extraction of attributes and
connections with external systems.
8 algorithms for the classification steps, which after processing
the registered image database provide information such as
accuracy, time of the processing in the form of graphs and V. N ON -F UNCTIONAL S YSTEM R EQUIREMENTS
tables to allow the user to choose the extractor and classifier set In addition to the safety concerns of the LINDA system,
he needs. Initially the idea was to make the selection of the best which were considered during development as the main non-
pair automatic; however it was decided to leave the choice up functional requirement, some other aspects in this category
to the user in order to make academic research and scientific were considered and they are listed below.
studies feasible. On the other hand for possible future (a) Scalability: scalability is a fundamental characteristic, as
implementations the user will be free to accept the automatic one of the goals of the system was to make the system available
selection or make his or her own choice. As already mentioned, online to anyone who wished to foster research. The
the web module was developed in Java language and uses the modularized development of LINDA makes this possible, as
Java Server Faces (JSF), Java Persistence API (JPA) and the web visualization subsystem is totally decoupled from the
Hibernate frameworks and is already available for access image processing part and carried out using different equipment.
through http://linda.lapisco.ifce.edu.br. The prediction API was Data on the consumption of resources for processing will be
specifically designed to work in conjunction with the proposed presented later.
system as well as other systems and devices. Therefore, to (b) Availability: the care to make the platform available at
ensure good interoperability, the API is able to communicate all times, especially when updating is very important due to the
through the exchange of JSON files, which are trafficked under type of application of LINDA, and the requirements of different
a secure environment under the HTTPS protocol (port 443). research environments and especially for the use with medical
During the development process XML files were considered, applications. Once again the modularization of the system
but as the JSON set up already suited the need, XML was not makes it possible to use several nodes of a network to take
implemented. Future versions may include the use of XML care of the availability of each of the necessary subsystems. In
which would allow comparisons of performance to be made. the case of total interruption in the image processing subsystem,
However, when compressed and sent in a test with a file that the user requests are placed in a queue and are performed as
was generated manually both were the same size, therefore soon as the module is available. As for the web and persistence
implying no change in bandwidth usage. part, there are several forms of redundancy of web servers and
databases that are simple to apply to this type of solution.
IV. S YSTEM D ESIGN (c) Interoperability: Considering the heterogeneity of the
environments, such as the Internet, the equipment that will be
The system was divided into four parts at conception. The used for access is unknown; however the use of the web
first part is represented in Figure 2 as Resource-and-Events that platform already solves this question. On the other hand,
is responsible to obtain images, either for prediction or as a
complement, where the system has elements of security and
user interface. The second part is the Context- and-

3 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 4

can be seen in the way people relate. Today, technology- driven

research environments, besides the Internet, require operations relationships are very common, where geographical and
together with or associated to other applications or equipment. temporal barriers are overcome by the omnipresence generated
External systems with direct access to databases or any other by ICTs.
elements of a system can create multiple points of code failures, This intermediation can be said to create virtual communities,
and could add serious security problems, which was one of the that is, virtual communication networks that allow the
main concerns during the development of the platform. interaction of its participants in an organized way around com-
Therefore to enable interoperability a module with a JSON file mon objectives, which are materialized through instruments
exchange format associated to a hash validation key was and documents shared by a group of people.
implemented. The goal here was to ensure a simple connection Some works in the literature use social networks in the
for a wide variety of existing or future platforms. treatment of health; however, they are based on the monitoring
of patients and their health through these social networks [29],
VI. S ECURITY ASPECTS [30]. No articles were found that used a specific social network
Security is essential in any application, and this aspect for the collaboration of data, although health professionals that
becomes more critical when a system is available online, as it were known to each other would share their results. However,
becomes open to attack from anywhere in the world. The lack for image processing there was only direct communication be-
of adequate security criteria compromises other characteristics tween the parties without any collaboration through a system.
of the system, such as availability, as well as affecting the Considering this reality, the LINDA system was designed,
credibility of the application [26], [27]. Our research into as of its conception, in such a way as to create a social
published works, especially those related to IoT, came across network associated with it in order to stimulate the sharing
various projects that had not given the important aspect of of knowledge and to motivate discussions of the results of the
safety any priority; perhaps it was a way to make a more agile projects developed within the system [31], [32].
market launch of the product, but this is not good practice [1], Although there are other social networks and collaboration
[26], [28]. platforms, which have a different focus, we do not think that
In light of these facts, some special care was taken in terms they will hinder the growth of the proposed system, because the
of the LINDA system security: a hash code is generated objectives of these other platforms have a target audience and
whenever a new project is created, which prevents the passing purpose very different from those intended in this work. On the
of parameters such as the extractor algorithm, the project ID other hand, this does not exclude the possibility of integrating
(Identity) or other elements that facilitate the analysis of data these systems in the future, but this will depend on the needs
and possible reverse engineering attempts, preventing possible that arise.
fraud. In addition, an IoT device able to communicate with the Thus, the proposed application allows the base images
system, must first be registered on the platform by the user, developed in the system to be shared among various users;
adding to the security. however, the original user base is always maintained, while at
Still with regard to the security of the framework, as LINDA the same time other users can delete or include new images.
allows access to a prediction page, a request analysis is Besides, the system has a cropping tool that allow users to limit
implemented in the firewall of the host machine, in order to the space of the image to be considered by the processing
limit the number of requests to the Prediction API and thus to algorithms, so users can focus on a specific area of an existing
mitigate denial of service attacks. The system also stores image without the need of an additional image editing
information such as IP and MAC addresses, in order to be able application.
to analyze the behavior of the devices and block access of Based on the resources present in the developed approach,
equipment with strange behavior. Although the Java no similar work was found in a search of articles and periodicals
frameworks used had already addressed the issue of SQL that combined a simple way to perform image processing and
injection, several tests were still performed in order to at the same time allow the generated knowledge to be shared
guarantee protection against this type of attack. and adjusted to the needs of each user. And if only the health
Further security tests are being conducted by the team using area is considered, the projects are limited to team training or
traffic assessment tools to mitigate any possible failures, since patient follow-up.
not only the Java language but any language may have security Figure 3 shows an overview of how our approach works.
holes which are constantly being solved with security updates. Initially, the images are obtained from the IoT device and
However, it is necessary to evaluate the impacts and uploaded to the framework via the Internet by the specialists.
possibilities so that if any failure occurs they do not interfere After that, to perform its actions, LINDA makes a system call
severely in the system, regardless of the release of the security to the prediction web service, sending in addition to the image
updates by the company that maintains the language. provided by the user, a hash code. The Prediction API is then
triggered, which will perform a preprocessing of the image,
VII. S OCIAL N ETWORK FOR O NLINE IMAGE P ROCESSING which consists of adjusting the size and, if necessary, color
AND ANALYSIS
conversion and bit depth, as a way to ensure proper operation
regardless of the capture device used. After which, CNN and
Simultaneously with technological advances and increased machine learning techniques are applied to the images and then
portability of computer equipment there has been clear changes the diagnosis is returned to the specialist and can be viewed on
in the behavior of people. Part of these changes

4 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 5

The users, in this trial, answered a simple questionnaire with

a computer or smartphone, for example. In the system, the IoT 5 questions, which were related to ease-of-use; if the system
Management function is to coordinate data transfers, regulate met its expectations; would the user recommend the system to
access to the system, ensure security aspects and take care of another person; and if the lack of documentation would hinder
integration with other applications or devices. the use of the system. The result was a mean positive evaluation
of 87%, from the average scores given to each of the questions,
all with equal weights.
The questions asked to user were: 1) How difficult is the
system to use? 2) Do you think the time taken to understand and
use the system was? 3) Did the layout of the menus and the
icons used allow you to locate the desired features? 4) Do you
consider a tutorial to explain how it works as fundamental to
system use? 5) To what extent do the features provided meet
your usage needs?
The equipment used in the development of the work has the
following configurations: Intel Xeon server with 16 processors
and 32 GB RAM, Linux operating system Ubuntu 16.04 64-
bit with version 1.8.0 of Java and Python version 3 installed.
Fig. 3: LINDA framework overview: 1) IoT devices send The tests used 3 databases of medical images, for the detection
images for processing or training; 2) IoT management is re- of stroke, lung nodules and skin lesions. All databases are
sponsible for system security, interoperability and persistence. described below. Examples of the images present in each
It controls the access for API Prediction; 3) API prediction database are shown in Figure 4.
processes the images and 4) human interaction with system.
A. Stroke Images Database
VIII. IMAGE P ROCESSING AND ANALYSIS
The first monitored test and performance-focused test was
In spite of the improvements in location, segmentation and carried out on 25 computerized CT scan images of cerebral
object detection performance provided by modern CNN vascular accidents (CVA), also called brain strokes. In total,
architectures, which have been tested using very large datasets there were 420 CT images in the database, 140 of healthy brains
such as ILSVRC, COCO and JFT, these techniques usually and 280 of brains with stroke, of which 140 were hemorrhagic
have thousands of parameters and, if applied to image data with and 140 were ischemic. The database was divided into 3 classes,
few samples, will probably lead to overfitting. as presented in [34].
Therefore, the use of Transfer Learning, which has already
been used in several works, is a good way to overcome such B. Lung Nodules Database
limitations. The main idea of transfer learning is to transfer
the ”knowledge” learned to solve a task [33], such as the The second monitored test classified 1536 lung nodules, of
classification of an object, to another task such as the which 1146 were non-malignant and 390 malignant. The set of
identification of a stroke on CT scans. samples used to represent the nodule malignancy classification
In order to be able to use transfer learning with CNN, a few problem in this work were obtained from the LIDC/IDRI
steps, such as building and training a model, like VGG or database, which contains 1010 chest CT scans and their
ResNet networks with sufficiently large datasets, similar to respective diagnoses, made by 4 experienced radiologists. The
ImageNet, need to be carried out and then the layers closest to database used (LIDC/IDRI) is available for free download at
the output must be removed. Then, the remaining architecture the National Biomedical Image Archive (NBIA) website
is applied to each of the images in the dataset to be used, and http://ncia.nci.nih.gov/ [35].
saves each of them, along with their labels, as a new dataset.
Then a new classifier is trained such as Support Vector Machine
or Multilayer Perceptron, using this new dataset. C. Skin Images Database
The third monitored test was done using the PH² Database, a
IX. P ROPOSED S YSTEM T EST
dermoscopic image database for research and benchmarking.
The system was tested for 2 months using projects of our The images that compose this database were acquired from the
research group where the system was created and with the Dermatology Service of Hospital Pedro Hispano (Matosinhos,
collaboration of partner institutions. The primary aim was to Portugal). In all, the database consists of 120 dermoscopic
verify the achievement of the initial objective, which was to images of melanocytic lesions which are divided into 80
create a simple platform that did not require advanced samples of common nevi and 40 samples of melanomas. This
knowledge for its use and verify its integration with a social database is public and available for download at the website
network developed by our team, in way that a common user http://www.fc.up.pt/addi/ph2%20database.html [36].
was be able to work with the complete solution without
problems of usability.

5 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 6

TABLE II: Results of accuracy (Acc), F1-Score (F1S), training

and test times for lung nodules database.
Extractor Acc (%) F1S (%) Training Time (s) Test Time (s)
VGG16 76.6/86.7/84.2/85.5/83.4/85.4 51.3/81.0/78.8/78.7/79.4/79.6 0.02/16.96/0.05/9.38/4.29/5.82 0.03/0.11/7.09/4.63/1.79/3.11
VGG19 72.9/99.8/100.0/100.0/99.8/100.0 58.3/99.7/100.0/100.0/99.7/100.0 0.1/34.2/0.1/30.7/1.0/1.4 0.2/0.5/1.2/33.0/0.4/0.5
MobileNet 91.6/100.0/100.0/100.0/100.0/100.0 90.7/100.0/100.0/100.0/100.0/100.0 0.1/117.6/0.3/33.8/2.3/2.5 0.2/3.3/3.5/33.6/0.9/1.0
ResNet50 71.7/100.0/100.0/99.3/100.0/100.0 55.5/100.0/100.0/99.1/100.0/100.0 0.2/384.5/0.2/44.0/3.1/2.4 0.2/3.7/3.1/31.4/1.0/0.8
InceptionV3 71.7/71.7/100.0/100.0/100.0/100.0 55.6/55.5/100.0/100.0/100.0/100.0 0.1/0.1/368.5/0.2/84.0/2.1 0.2/0.5/3.1/2.5/32.5/0.9
Xception 71.7/100.0/100.0/99.8/100.0/100.0 55.5/100.0/100.0/99.7/100.0/100.0 0.1/675.3/0.4/38.2/2.3/2.4 0.2/4.7/2.5/32.2/0.7/0.7
InceptionResNetV2 68.5/97.3/100.0/99.8/100.0/100.0 53.3/97.0/100.0/99.7/100.0/100.0 0.1/497.4/0.2/81.2/1.6/1.9 0.2/2.5/2.0/32.8/0.6/0.7
DenseNet121 71.7/97.8/100.0/99.8/100.0/100.0 55.5/97.7/100.0/99.7/100.0/100.0 0.1/403.1/0.2/48.5/1.6/1.4 0.2/1.8/1.7/31.5/0.7/0.6
DenseNet201 71.7/98.1/100.0/99.8/100.0/100.0 55.5/98.0/100.0/99.7/100.0/100.0 0.1/674.1/0.2/35.8/2.7/3.1 0.2/2.9/3.4/33.8/0.9/1.1
DenseNet169 71.7/98.8/100.0/100.0/100.0/100.0 55.4/98.8/100.0/100.0/100.0/100.0 0.1/795.3/0.1/30.4/1.7/2.0 0.2/2.9/2.5/34.0/0.7/0.8
NASNetLarge 71.7/99.8/100.0/100.0/100.0/100.0 55.6/99.7/100.0/100.0/100.0/100.0 0.2/385.5/0.4/29.9/8.2/6.5 0.2/3.7/4.5/35.2/1.5/1.3
NASNetMobile 91.6/100.0/100.0/100.0/100.0/100.0 90.7/100.0/100.0/100.0/100.0/100.0 0.1/117.6/0.3/33.8/2.3/2.5 0.2/3.3/3.5/33.6/0.9/1.0

TABLE III: Results of accuracy (Acc), F1-Score (F1S), train-

ing and test times for skin database.
Extractor Acc (%) F1S (%) Training Time (s) Test Time (s)
VGG16 62.0/86.0/82.0/84.0/90.0/90.0 41.4/82.6/77.7/81.9/88.9/88.9 0.01/1.83/0.01/7.22/0.01/0.01 0.01/0.03/0.02/3.77/0.01/0.01
Fig. 4: Example of segmentation of the images of the VGG19
MobileNet
60.0/86.0/82.0/82.0/86.0/86.0
92.0/92.0/90.0/90.0/92.0/92.0
37.5/84.1/79.5/78.1/84.3/84.0
91.5/91.5/88.1/89.3/91.5/91.5
0.01/0.65/0.01/7.03/0.01/0.01
0.01/0.74/0.01/6.76/0.01/0.01
0.01/0.01/0.02/3.74/0.01/0.01
0.01/0.03/0.02/3.69/0.01/0.01

considered databases. The top row shows the original images ResNet50
InceptionV3
60.0/84.0/78.0/84.0/82.0/82.0
86.0/86.0/80.0/82.0/86.0/86.0
37.5/80.4/72.2/81.0/80.1/80.1
84.5/84.5/76.3/78.3/84.5/84.5
0.01/12.23/0.01/6.79/0.02/0.02
0.01/2.40/0.01/7.29/0.03/0.03
0.01/0.06/0.03/3.73/0.02/0.02
0.02/0.04/0.02/3.92/0.02/0.02
Xception 80.0/84.0/76.0/82.0/86.0/86.0 71.5/82.3/71.1/79.9/84.7/84.7 0.01/1.52/0.01/7.27/0.03/0.03 0.01/0.04/0.02/3.78/0.02/0.02
and the bottom row the segmented images. InceptionResNetV2
DenseNet121
78.0/86.0/74.0/78.0/86.0/86.0
62.0/88.0/86.0/90.0/92.0/90.0
69.3/82.2/66.4/72.7/82.6/82.6
41.4/86.5/83.7/88.9/91.8/89.5
0.01/4.01/0.01/7.08/0.02/0.02
0.01/0.47/0.01/6.82/0.01/0.02
0.01/0.05/0.02/3.73/0.02/0.02
0.01/0.01/0.03/3.69/0.02/0.02
DenseNet201 68.0/92.0/94.0/92.0/92.0/92.0 51.5/91.6/93.2/90.9/91.6/91.6 0.01/1.01/0.01/7.11/0.02/0.02 0.01/0.01/0.03/3.78/0.02/0.02
DenseNet169 70.0/90.0/86.0/84.0/90.0/90.0 56.8/89.2/85.5/81.7/89.6/89.5 0.01/1.25/0.01/7.11/0.02/0.02 0.02/0.02/0.02/3.78/0.02/0.02
NASNetLarge 60.0/76.0/70.0/72.0/76.0/76.0 37.5/72.6/65.9/66.4/73.5/73.5 0.01/3.71/0.01/7.01/0.04/0.05 0.01/0.05/0.04/3.71/0.02/0.02
X. RESULTS AND D ISCUSSION NASNetMobile 80.0/84.0/76.0/84.0/84.0/84.0 78.4/82.5/73.6/82.3/81.9/81.9 0.01/1.63/0.01/6.97/0.02/0.02 0.01/0.03/0.02/3.72/0.02/0.02

In all tests, Accuracy (Acc) and F1-Score (F1S) metrics were TABLE IV: The best results of the proposed system for all the
used for the combinations of all CNN architectures with all databases considered.
classifiers, seeking to evaluate the performance of the Database Extractor / Classifier Acc (%) F1S (%) Processing Time (s)
framework. In addition to the metrics listed, the training and
test times of the classifiers are also presented.
The 12 (twelve) CNN architectures used in this work as
extractors were: VGG16, VGG19, MobileNet, ResNet50,
For the CVA database the system obtained 100% in accuracy
InceptionV3, Xception, InceptionResNetV2, DenseNet121,
and F1-Score when the kNN and SVM (RBF) classifiers were
DenseNet201, DenseNet169, NASNetLarge and NASNetMo-
used in conjunction with all the CNN architectures. The
bile, and the 6 (six) algorithms used as classifiers were:
training times of Bayes with VGG16, VGG19, MobileNet,
Bayesian Classifier (Bayes) / Multi-layer Perceptron (MLP)
InceptionV3, Xception, InceptionResNetV2, DenseNet121,
/ k-Nearest Neighbor (kNN) / Random Forest (RF) / Support
DenseNet201, DenseNet169 and NASNetMobile and kNN in
Vector Machine with Linear kernel: SVM (Linear) / Support
association with VGG16, VGG19 and DenseNet169 were the
Vector Machine with RBF kernel: SVM (RBF), with the results
fastest taking 0.1s. The classification times of Bayes were also
arranged in this order in Tables I, II and III.
the fastest in association with all CNN architectures, taking
In the classification processes, Bayesian classifier operated
only 0.2s.
with the Gaussian probability density function. The hyperpa-
The results for the lung nodule classifications showed 89.1%
rameter k from the kNN classifier was chosen through a grid
of accuracy with SMV (RFB) and MLP in combination with
search, testing the odd values from 1 to 11. Random Forest
InceptionV3 and Xception, respectively. The highest score
training process used 3000 estimators. MLP carried out its
obtained for F1-Score was 85.8% by combining SVM (RFB)
training using the Levenberg-Marquardt method and neurons
and InceptionV3. While SVM (RFB) and SVM (Linear) scored
ranging from 2 to 1000 in the hidden layer. SVM used linear
85.5% when associated with Xception. The best times for clas-
and RBF kernels and hyper parameters C and γ with 2−5, 2−4,
sification and test were also obtained by Bayes in association
2−3, ...,215 and 2−15, 2−14, ...,23 values, respectively.
with all CNN architectures, and was less than 0.1s.
The LINDA framework executes all the combinations of
The skin database results for DenseNet201 in combination
extractors and classifiers and chooses the best result for each
with k-Nearest Neighbor were 94.0% for accuracy and 93.2%
project. Table I shows the combination of extractor and
for the F1-Score. MobileNet obtained 90.0% accuracy when
classifier that represents the best result for a specific database,
associated with anyone of the system classifiers and the SVM
using accuracy and F1-Score as the criteria, in this order. If
(Linear) scored more than 88.0% for the F1-Score when
these two criteria are equal, then the next criterion to be
combined to DenseNet121, DenseNet169 or VGG16. The
compared is the total time, the sum of the training and test times.
fastest processing time was obtained when MobileNet was
TABLE I: Results of accuracy (Acc), F1-Score (F1S), training combined with Bayes and was less than 0.01s. The best results
and test times for CVA database. for each database are presented in the Table IV.
Extractor Acc (%) F1S (%) Training Time (s) Test Time (s)

A. Evaluating the Collaborative Network

VGG16 72.2/100.0/100.0/100.0/100.0/100.0 57.5/100.0/100.0/100.0/100.0/100.0 0.1/97.2/0.1/31.9/1.0/1.0 0.2/2.0/1.2/32.5/0.3/0.4
VGG19
MobileNet
72.9/99.8/100.0/100.0/99.8/100.0
91.6/100.0/100.0/100.0/100.0/100.0
58.3/99.7/100.0/100.0/99.7/100.0
90.7/100.0/100.0/100.0/100.0/100.0
0.1/34.2/0.1/30.7/1.0/1.4
0.1/117.6/0.3/33.8/2.3/2.5
0.2/0.5/1.2/33.0/0.4/0.5
0.2/3.3/3.5/33.6/0.9/1.0
One of the main distinctions of this work is the integration of
ResNet50 71.7/100.0/100.0/99.3/100.0/100.0 55.5/100.0/100.0/99.1/100.0/100.0 0.2/384.5/0.2/44.0/3.1/2.4 0.2/3.7/3.1/31.4/1.0/0.8
InceptionV3
Xception
71.7/71.7/100.0/100.0/100.0/100.0
71.7/100.0/100.0/99.8/100.0/100.0
55.6/55.5/100.0/100.0/100.0/100.0
55.5/100.0/100.0/99.7/100.0/100.0
0.1/0.1/368.5/0.2/84.0/2.1 .
0.1/675.3/0.4/38.2/2.3/2.4
0.2/0.5/3.1/2.5/32.5/0.9
0.2/4.7/2.5/32.2/0.7/0.7
a medical image processing platform with a social network.
InceptionResNetV2
DenseNet121
68.5/97.3/100.0/99.8/100.0/100.0
71.7/97.8/100.0/99.8/100.0/100.0
53.3/97.0/100.0/99.7/100.0/100.0
55.5/97.7/100.0/99.7/100.0/100.0
0.1/497.4/0.2/81.2/1.6/1.9
0.1/403.1/0.2/48.5/1.6/1.4
0.2/2.5/2.0/32.8/0.6/0.7
0.2/1.8/1.7/31.5/0.7/0.6 This integration aims to improve the precision of the results
DenseNet201 71.7/98.1/100.0/99.8/100.0/100.0 55.5/98.0/100.0/99.7/100.0/100.0 0.1/674.1/0.2/35.8/2.7/3.1 0.2/2.9/3.4/33.8/0.9/1.1
DenseNet169
NASNetLarge
71.7/98.8/100.0/100.0/100.0/100.0
71.7/99.8/100.0/100.0/100.0/100.0
55.4/98.8/100.0/100.0/100.0/100.0
55.6/99.7/100.0/100.0/100.0/100.0
0.1/795.3/0.1/30.4/1.7/2.0
0.2/385.5/0.4/29.9/8.2/6.5
0.2/2.9/2.5/34.0/0.7/0.8
0.2/3.7/4.5/35.2/1.5/1.3
through the collaboration of other users of the system that allow
NASNetMobile 91.6/100.0/100.0/100.0/100.0/100.0 90.7/100.0/100.0/100.0/100.0/100.0 0.1/117.6/0.3/33.8/2.3/2.5 0.2/3.3/3.5/33.6/0.9/1.0
the use of their databases. In a resource sharing system,
databases must be made available.

6 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 7

XI. C ONCLUSION
This work proposed the development of a collaborative
social network to aid in medical diagnoses. The aim was to
make better use of machine learning technologies for those who
are already knowledgeable in the field, as well as making it
possible for people with no knowledge in the areas of
computational intelligence, mathematical learning and
convolutional neural networks to use these resources in a
simple and systematic way. The integration of these
technologies with a framework that harnesses the collaborative
potential can generate a large knowledge base for each specific
problem and foster collaborative problem solving.
At the end of the work there is a significant gain provided by
the tool that enables users to carry out a proof of concept, using
deep learning and other advanced computer-based techniques,
by including their databases and configurations. Moreover,
Fig. 5: Comparison of the results of collaboration with the they can integrate their data with other collaborators / users of
AVC and Lung Nodule bases. the social network. Thus, at the end of the process, the users
have evaluated their hypothesis, and at the same time have a
In order to evaluate the gains through collaboration, the ready-to-use product.
development group was divided into three teams with distinct The usability of the system was well evaluated by 87% of the
development areas, and each to carry out the image processing users who responded to the questionnaire. Furthermore, the
based on the available bases. The objective here was not to integration of the social network with the application, which
achieve greater accuracy, but to verify how much collaboration
uses the concepts of deep machine learning, allowed the use of
would lead to improved results.
artificial intelligence methods in a collaborative and centralized
In the first test the AVC base was divided into three parts,
way. In addition to allowing the interchange of various users in
with 2 (two) teams with 47 samples from each base and 1 (one)
with 46 from each. When the procedures were carried out a project, examples were added and the developed tool was
individually, the accuracies obtained were 81.2%, 87.7% and used in a collaborative way.
78.9%. When the two teams with the most samples shared their Finally, the accuracy of the results, which were already good
bases with each other, the accuracy increased to 93.6%, which with the simple use of the existing algorithms, was reinforced.
represented a gain of approximately 15.27% over the worst They were improved by more than 22% when the image
result, however, the processing time increased. Accuracy is databases were enlarged through collaboration provided by the
considered to be of greater importance than variations in social network. It is also worth mentioning that, in addition to
processing time, especially when such variations do not even the databases of medical images cited here, the tool was applied
reach one second (1s). to other types of image databases, such as industrial processes
A second test was then carried out with the pulmonary and evaluation of equipment. The results here were also good,
nodule base, following the same methodology as the previous and accuracy gains of more than 15%, were obtained with the
test, of dividing the group into 3 (three) teams. The individual tool, but were not presented in more detail because they were
accuracy results obtained were 65.2%, 73.4 and 69.9%. Again not the focus of this work. However they also reinforce the
on sharing the bases provided by the proposed solution, the best benefits of using social networks in this context.
result of combining two of the two bases rose to 80.1%, which As future works, we intend to analyze other databases
represented an accuracy gain of 22.85% over the worst result. focused on the medical field, such as cardiac image processing,
The results of the lung nodule and AVC databases can be seen in order to detect and/or classify cardiovascular diseases. An
in Figure 5. integration of the proposals presented in [37], [38] and [39],
The last test performed with the health-related bases was initially, would be very useful for the growth and, consequently,
with melanoma, but because of the smaller number of samples evolution of the platform, making it more comprehensive and
it was only divided into 2 (two) parts not 3 (three). The results robust. Future works will also address aspects related to data
obtained in relation to accuracy, for the individual processing transport optimization, which may compromise the platform
was 88.4% and 90.1%, and there was an accuracy gain of 6.33% integration with some devices. In addition, every system will
when using the whole base, thus reaching 94.0%. These results undergo an interface evaluation in order to be simple to use by
verify the gain that collaborative work can give, bringing professionals without much computer knowledge and be
benefits both to those who use the solution for diagnosis, by adapted for use by people with disabilities.
obtaining more accurate results, and for those who work with
ACKNOWLEDGMENT
image processing, and are then able to use larger study bases to
test their algorithms. The authors acknowledge the sponsorship from the FUN-
CAP, CAPES and CNPq (via grants no. 304315/2017-6 and

7 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSAC.2020.3020598, IEEE Journal
on Selected Areas in Communications
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, AUGUST XXXX 8

430274/2018-1). [20] Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E.

REFERENCES Hubbard, and L. D. Jackel, “Handwritten digit recognition with a back-
propagation network,” in Advances in Neural Information Processing
Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-
[1] M. A. A. da Cruz, J. J. P. C. Rodrigues, J. Al-Muhtadi, V. V. Korotaev, Advances in neural infor- mation processing systems, pp. 1097–1105.
and V. H. C. de Albuquerque, “A reference model for internet of things [22] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
middleware,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 871– L. Fei-Fei, “Large-scale video classification with convolutional neural
883, April 2018. networks,” in Proceedings of the IEEE conference on Computer Vision
[2] J. J. P. C. Rodrigues, D. B. D. R. Segundo, H. A. Junqueira, M. H. Sabino, and Pattern Recognition, 2014, pp. 1725–1732.
R. M. Prince, J. Al-Muhtadi, and V. H. C. D. Albuquerque, “Enabling [23] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans-
technologies for the internet of health things,” IEEE Access, vol. 6, pp. actions on knowledge and data engineering, vol. 22, no. 10, pp. 1345–
13 129–13 141, 2018. 1359, 2010.
[3] I. Yaqoob, E. Ahmed, I. A. T. Hashem, A. I. A. Ahmed, A. Gani, [24] D. Zhang, D. Meng, and J. Han, “Co-saliency detection via a self-paced
M. Imran, and M. Guizani, “Internet of things architecture: Recent multiple-instance learning framework,” IEEE Transactions on Pattern
advances, taxonomy, requirements, and open challenges,” IEEE Wireless Analysis and Machine Intelligence, vol. 39, no. 5, pp. 865–878, May 2017.
Communications, vol. 24, no. 3, pp. 10–16, June 2017. [25] D. Zhang, J. Han, C. Li, J. Wang, and X. Li, “Detection of co-salient objects
[4] D. S. Linthicum, “Cloud computing changes data integration forever: by looking deep and wide,” International Journal of Computer Vision, vol.
What’s needed right now,” IEEE Cloud Computing, vol. 4, no. 3, pp. 50– 120, no. 2, pp. 215–232, Nov 2016.
53, 2017. [26] R. R. Guimaraes, L. A. Passos, R. H. Filho, V. H. C. de Albuquerque,
[5] P. P. R. Filho, R. M. Sarmento, G. B. Holanda, and D. de Alencar Lima, J. J. P. C. Rodrigues, M. M. Komarov, and J. P. Papa, “Intelligent network
“New approach to detect and classify stroke in skull CT images via security monitoring based on optimum-path forest clustering,” IEEE
analysis of brain tissue densities,” Computer Methods and Programs in Network, 2018.
Biomedicine, vol. 148, pp. 27–43, 2017. [27] J. E. Siegel, S. Kumar, and S. E. Sarma, “The future internet of things:
[6] X. J. Cai, J. I. E. Ignacio, E. F. Mendoza, D. J. F. Rabino, R. P. G. Secure, efficient, and model-based,” IEEE Internet of Things Journal,
Real, and E. A. Roxas, “Iot-based gait monitoring system for static 2018.
and dynamic classification of data,” in 2018 IEEE 10th Interna- [28] V. Radisavljevic-Gajic, S. Park, and D. Chasaki, “Vulnerabilities of
tional Conference on Humanoid, Nanotechnology, Information Tech- control systems in internet of things applications,” IEEE Internet of
nology,Communication and Control, Environment and Management Things Journal, 2018.
(HNICEM), Nov 2018, pp. 1–4. [29] G. Harerimana, B. Jang, J. W. Kim, and H. K. Park, “Health big data
[7] M. Vanan, M. Balamurugan, H. L, N. , and M. Reddy, “Iot based analytics: A technology survey,” IEEE Access, 2018.
continuous monitoring of cardiac patients using raspberry pi,” vol. 2039, [30] S. Sidana, S. Amer-Yahia, M. Clausel, M. Rebai, S. T. Mai, and
11 2018, p. 020025. M. Amini, “Health monitoring on social media over time,” IEEE
[8] D. C. Yacchirema, D. Sarabia-Jacome, C. E. Palau, and M. Esteve, Transactions on Knowledge and Data Engineering, 2018.
“System for monitoring and supporting the treatment of sleep apnea using [31] C. M. Dourado, S. P. da Silva, R. V. da N óbrega, A. C. Barros, A. K.
iot and big data,” Pervasive and Mobile Computing, vol. 50, pp. 25–40, Sangaiah, P. P. R. Filho, and V. H. C. de Albuquerque, “A new approach
2018. for mobile robot localization based on an online iot system,” Future
[9] S. K. Lakshmanaprabu, S. N. Mohanty, S. S. Rani, S. Krishnamoorthy, Generation Computer Systems, 2019.
U. J., and K. Shankar, “Online clinical decision support system using [32] C. M. D. Jr, S. P. P. da Silva, R. V. M. da N óbrega, A. C. da S. Barros,
optimal deep neural networks,” Appl. Soft Comput., vol. 81, 2019. P. P. R. Filho, and V. H. C. de Albuquerque, “Deep learning iot system
[10] M. A. Sayeed, S. P. Mohanty, E. Kougianos, and H. P. Zaveri, “Neuro- for online stroke detection in skull computed tomography images,”
detect: A machine learning-based fast and accurate seizure detection Computer Networks, vol. 152, pp. 25 – 39, 2019.
system in the iomt,” IEEE Trans. Consumer Electronics, vol. 65, no. 3, [33] K. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of transfer learning,”
pp. 359–368, 2019. Journal of Big Data, vol. 3, no. 1, p. 9, 2016.
[11] I. Mehmood, A. Ullah, K. Muhammad, D. Deng, W. Meng, F. Al- [34] P. P. R. Filho, E. de S. Rebouças, L. B. Marinho, R. M. Sarmento,
Turjman, M. Sajjad, and V. H. C. de Albuquerque, “Efficient image J. M. R. Tavares, and V. H. C. de Albuquerque, “Analysis of human tissue
recognition and retrieval on iot-assisted energy-constrained platforms densities: A new approach to extract features from medical images,”
from big data repositories,” IEEE Internet of Things Journal, pp. 1–1, Pattern Recognition Letters, vol. 94, pp. 211 – 218, 2017.
2019. [35] M. B. Rodrigues, R. V. M. Da N óbrega, S. S. A. Alves, P. P. R. Filho,
[12] M. S. Hossain, M. F. Alhamid, and G. Muhammad, “Collaborative anal- J. B. F. Duarte, A. K. Sangaiah, and V. H. C. De Albuquerque, “Health of
ysis model for trending images on social networks,” Future Generation things algorithms for malignancy level classification of lung nodules,”
Computer Systems, vol. 86, pp. 855 – 862, 2018. IEEE Access, vol. 6, pp. 18 592–18 601, 2018.
[13] N. Moradianzadeh, P. M. Zadeh, Z. Kobti, S. Hansen, and K. Pfaff, [36] F. F. X. Vasconcelos, A. G. Medeiros, S. A. Peixoto, and P. P. R. Filho,
“Using social network analysis to model palliative care,” Journal of “Automatic skin lesions segmentation based on a new morphological
Network and Computer Applications, vol. 120, pp. 30 – 41, 2018. approach via geodesic active contour,” Cognitive Systems Research, vol.
[14] S. Enshaeifar, P. Barnaghi, S. Skillman, A. Markides, T. Elsaleh, S. T. 55, pp. 44 – 59, 2019.
Acton, R. Nilforooshan, and H. Rostill, “The internet of things for [37] C. Xu, L. Xu, Z. Gao, S. Zhao, H. Zhang, Y. Zhang, X. Du, S. Zhao,
dementia care,” IEEE Internet Computing, 2018. D. N. Ghista, H. Liu, and S. Li, “Direct delineation of myocardial
[15] S. Peng, S. Yu, and P. Mueller, “Social networking big data: Opportuni- infarction without contrast agents using a joint motion feature learning
ties, solutions, and challenges,” Future Generation Computer Systems, architecture,” Medical Image Analysis, vol. 50, pp. 82–94, 2018.
vol. 86, pp. 1456 – 1458, 2018. [38] S. Zhao, Z. Gao, H. Zhang, Y. Xie, J. Luo, D. N. Ghista, Z. Wei, X. Bi,
[16] M. B. Rodrigues, R. V. M. D. N óbrega, S. S. A. Alves, P. P. R. Filho, H. Xiong, C. Xu, and S. Li, “Robust segmentation of intima-media
J. B. F. Duarte, A. K. Sangaiah, and V. H. C. D. Albuquerque, “Health of borders with different morphologies and dynamics during the cardiac
things algorithms for malignancy level classification of lung nodules,” cycle,” IEEE J. Biomedical and Health Informatics, vol. 22, no. 5, pp.
IEEE Access, vol. 6, pp. 18 592–18 601, 2018. 1571–1582, 2018.
[17] P. Pace, G. Aloi, R. Gravina, G. Caliciuri, G. Fortino, and A. Liotta, “An [39] L. B. Marinho, N. de Medeiros Mendonça e Nascimento, J. W. M. Souza,
edge-based architecture to support efficient applications for healthcare M. V. Gurgel, P. P. R. Filho, and V. H. C. de Albuquerque, “A novel
industry 4.0,” IEEE Transactions on Industrial Informatics, pp. 1–1, 2018. electrocardiogram feature extraction approach for cardiac arrhythmia
[18] H. Ye, F. Gao, Y. Yin, D. Guo, P. Zhao, Y. Lu, X. Wang, J. Bai, K. Cao, classification,” Future Generation Comp. Syst., vol. 97, pp. 564–577,
Q. Song, H. Zhang, W. Chen, X. Guo, and J. Xia, “Precise diagnosis of 2019.
intracranial hemorrhage and subtypes using a three-dimensional joint
convolutional and recurrent neural network,” European Radiology, Apr
2019.
[19] Z. Gao, Y. Li, Y. Sun, J. Yang, H. Xiong, H. Zhang, X. Liu, W. Wu,
D. Liang, and S. Li, “Motion tracking of the carotid artery wall from
ultrasound image sequences: a nonlinear state-space approach,” IEEE
Transactions on Medical Imaging, vol. 37, no. 1, pp. 273–283, Jan 2018.
30, 1989], 1989, pp. 396–404.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet
classification with deep convolutional neural networks,” in
8 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
0733-8716 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 02,2020 at 12:47:21 UTC from IEEE Xplore. Restrictions apply.