PUBLICATION

PHASE-I

P. Shanmugam , Keshav SR, Mithilesh Kumaar JS (2024), “Improve Medical

Image Diagnosis in Healthcare Utilizing a Framework for the Web
Application”, International Conference on Smart Technology, Artificial
Intelligence and Computer Engineering (ICSTAICE-2023), paper got accepted
and published.
 PHASE-II

P. Shanmugam , Keshav SR, Mithilesh Kumaar JS (2024), “Medical Image

Diagnosis In Healthcare Utilizing A Web Application”, was submitted to
5th IEEE India Council International Subsections Conference 2024.
REPORT PLAGIARISM:
 Improve Medical Image Diagnosis In
Healthcare Utilizing A Framework For The
Web Application

Dr.P.Shanmugam Keshav SR Mithilesh Kumaar JS

Assistant Professor Student Student
Department of Department of Department of
Computer Science and Computer Science and Computer Science and
Engineering Engineering Engineering
Rajalakshmi Engineering Rajalakshmi Engineering Rajalakshmi Engineering
College Chennai, India College Chennai, India College Chennai, India
shanmugam.p@rajalakshmi.e 200701124@rajalakshmi.edu. 200701145@rajalakshmi.edu.
du.in in in

ABSTRACT: The need for precise and effective medical diagnosis has grown in importance in the healthcare industry in
today's fast-paced world.. The core concept driving clinical diagnosis is to minimize human error in medical settings, a
principle that extends beyond healthcare to other domains such as earth observation via satellites and comprehending
activities in outer space. The primary motivation behind the development of our project lies in providing doctors with a
reliable tool to predict and address potential health issues through a user-friendly web platform. Additionally, we aim to
enhance the overall user experience by implementing features that manage user history and preferences. The technologies at
the heart of our solution involve cutting-edge convolutional neural networks (CNN) and the powerful EfficientNet B3 for
image processing, combined with the versatility of React.js for crafting an interactive web front-end. Our project is firmly
grounded in image data, addressing common challenges in image processing, including overfitting, hyperparameter
sensitivity, and time consumption. By tackling these issues head-on, we aim to empower medical professionals with rapid and
accurate diagnostic results, aligning with their need for swift decision-making and optimal patient care.

INTRODUCTION and advanced techniques to enhance, analyze and interpret

medical images. It includes a variety of imaging scans from
In today's world, medical imaging uses digital technology X-rays to MRI and CT scans, and facilitates operations
such as segmentation, feature extraction and
computer-aided diagnosis. The combination of machine are identical, it leads to the creation of symmetric B-CNNs,
learning and deep learning helps in recognizing the reminiscent of second-order pooling (O2P), a common
disease, plan treatment and predict outcomes. technique used in semantic segmentation However, this
approach comes with some limitations, including
The main purpose behind image processing techniques is to sensitivity to hyperparameters, risk of overfitting, limited
revolutionize healthcare by improving the accuracy, applicability to specific scenarios and problems in
efficiency and usability of medical imaging. These systems achieving interpretable functions.
enable early detection of the disease, allowing timely and
accurate diagnosis. It supports individualized treatment
plans, reduces procedural complications and improves Cancer Detection Using CNN (2018)[2] by Ishani Dabral
patient outcomes. Medical imaging also extends medical et.al proposed a convolutional neural network (CNN)
care to areas underserved by telemedicine and contributes architecture for cancer detection focusing on the Invasive
to research and medical education. By integrating Ductal Carcinoma (IDC) dataset. This innovative
intelligence, these systems help inform decision-making, architecture combines convolutional, dropout, and fully
increase efficiency, engage patients, and ultimately control connected layers along with specific data preprocessing
healthcare costs. The main goal is to improve the quality of steps. However, it faces challenges such as the demand for
patient care, diagnosis and treatment while making a large dataset, model complexity, excessive risks, and the
healthcare more efficient and patient-focused. complexity of implementing transfer learning techniques.

Technologies used in this project are CNN and In A Novel Attention-Based CNN-RNN Hybrid
EfficientNet. The next column explains the technologies we Architecture for sEMG-Based Gesture Recognition
use. (2018)[3], the proposed architecture combines
convolutional neural networks (CNNs) and recurrent neural
Convolutional Neural Network (CNN) greatly helps with networks (RNNs) with an attentional mechanism that
image processing and pattern recognition. It uses recognizes gestures based on surface electromyography
convolutional layers to automatically learn hierarchical signals (sEMG). CNNs extract features from images, while
features of data. CNNs perform the most accurate and RNNs, especially Long Short-Term Memory (LSTM) units,
adaptive computer vision tasks. model temporal dependencies in data. An attention
mechanism is introduced to highlight relevant information
EfficientNet is a family of convolutional neural networks in sEMG signals. The architecture uses a combination of
(CNN) optimized for computer vision. It uses a loss functions, including attention loss and target
combination of various techniques to balance depth, width replication loss.
and resolution, making the model effective with accuracy.
EfficientNet is well-known in the deep learning community Challenges in this approach include the vanishing and
for its ability to deliver high performance while using fewer exploding gradient problem, which can hinder training and
computational resources. make it difficult to effectively capture long-range
dependencies in sequences. RNNs also face limitations
This article aims to provide an overview of the system by related to their short-term memory, excelling at tasks with
touching on its architecture, functionality and benefits. By relatively short sequences but struggling to remember
exploring its potential and practical applications, we aim to information from distant past events.
demonstrate the transformative power of this technology,
which integrates the highest level of intelligence with
humans to edit medical images. Through this research, we Wei Fang et.al[4] proposed a Deep Convolutional
aim to demonstrate the potential of this system to transform Generative Adversarial Network (DCGAN) including a
customer loyalty and error-free operation in healthcare, discriminative network and a generative network, both
ushering in a new era of efficiency and consumer analysis. consisting of four layers. The generator transforms noise
vectors into realistic samples, while the discriminator
distinguishes between real and generated data. The training
2. RELATED WORKS process involves a non-convex game, and an adaptive
learning rate reduction strategy is used to increase the
Tsung-Yu Lin et.al[1], discussed the architecture bilinear training efficiency. However, this approach encounters
pooling to combine the outer product of two features several problems, including training instability, regime
through global averaging, resulting in a covariance matrix collapse (where the generator produces limited diversity),
that captures pairwise interactions. When these two features hyperparameter sensitivity, and the need for significant
computational resources due to its intensive nature.

Abien et.al[5] explores an experimental approach that In Improved Classification of MRI Images for Cervical
combines SVM and CNN for image classification. This Cancer Using CONVOLUTIONAL NEURAL
hybrid architecture uses SVM for binary classification, NETWORKS [9] the paper presents an approach for the
which determines the optimal hyperplane for separating automatic identification of hepatocellular carcinoma. It
two classes. The study considers both L1-SVM and uses an ensemble learning strategy with seven machine
L2-SVM, using the hinge loss as a standard loss function learning algorithms and uses collaborative representation
for SVM. However, this paper highlights problems classification with a boosting technique to categorize
including a lack of comprehensive understanding, potential hyperspectral images. However, this method faces
bias and fairness issues, absence of moral or ethical challenges, including a lack of domain knowledge and
considerations, and long-term memory-related limitations potential limitations in portability to different domains.
in the classification process.
Yiwen Liu, et.al[10] presented the paper that outlines the
An Image Classification Algorithm using AlexNet [6], the methodology for identifying crop diseases and insect pests.
authors present a method for image classification using an It uses transfer learning with VGG16 and
improved version of the AlexNet convolutional neural Inception-ResNet-v2 (CNN) convolutional neural
network (CNN). This approach solves the problem of networks. However, this approach faces challenges,
traditional CNNs with a high number of parameters, including resource intensiveness, the need for model
especially in fully connected layers, which can lead to updates, and complexity associated with fine-tuning.
problems such as overfitting and increased computational
complexity. The proposed enhancement introduces Sota Kato, et.al[11] proposed a study involving 50 cell
deconvolution layers into the AlexNet architecture, images with ground truth provided by Kyoto University.
effectively reducing the number of parameters, especially These images are obtained using a fluorescent marker on
in the fully connected layer. However, it is important to the cell membrane and nucleus of mouse liver and are
note that this improvement comes with increased 256×256 pixels in size. Among these images, 40 were used
computational complexity, the risk of overfitting due to for training, five for validation, and the remaining five for
model complexity, and the requirement for a significant testing. However, it should be noted that the proposed
amount of labeled training data. Achieving a balance method did not significantly improve the cell membrane
between these factors is essential for successful segmentation accuracy.
implementation of the improved AlexNet-based image
classification algorithm. Image Processing for Malaria Screening Using a Deep
Learning Approach [12] by Fetulhak Abdurahm, et.al. The
Yanan Sun, et.al, presented a method that uses a genetic study aims to improve the detection of P. falciparum in
algorithm (GA) to automatically optimize the architecture thick-smear microscopic images using an advanced deep
of convolutional neural networks (CNN). The process starts learning object . detection models. The aim is to address the
by initializing a population of CNN architectures, encoding challenges that were previously identified in this context.
them using skip and pooling layers and omitting fully To achieve this, the paper proposes a tile-based image
connected layers. Although this approach appears processing approach that enables better detection of small
promising, it faces challenges, including reproducibility objects in high-resolution images that exceed the input
issues, domain knowledge requirement, cross-domain resolution limits of the network. However, it should be
transferability difficulties, generalizability limitations, and noted that this method may not achieve a very high level of
data and time resource requirements. detection performance, especially when resizing the
high-resolution input image for detection.
Shengyi Zhao, et.al[8] presented an attention module to
improve model performance with minimal computational Md. Ashiq Mahmood and Tamal Joyti Roy[13], conducted
cost. This module uses the Squeeze-and-Excitation an experiment involving image analysis to establish an
Networks (SENet) method, which automatically assigns assessment process for identifying COVID-19-related lung
weight coefficients to feature channels based on their infection. This approach first involves analyzing the images
importance. This recalibration process reinforces relevant without converting them into binary segments.
feature channels while suppressing irrelevant ones. Subsequently, the authors convert normal images into
However, this approach comes with problems, including binary and collect the relevant values. The experiment also
computational complexity, reduced interpretability, and the uses parameter-based neural networks. However, one of the
need for careful hyperparameter tuning. challenges facing this study is the availability of
insufficient datasets for comprehensive analysis. Image
division is controlled by local segmentation.

Soto Kato and Kazuhiro Hotta[14], proposed a study using

50 cell images that come with ground truth data provided
by Kyoto University. These images are captured using a
fluorescent marker on the cell membrane and nucleus of a
mouse liver with a size of 256×256 pixels. Of these images,
40 were allocated for training, five for validation and the
remaining five for testing. However, it was observed that 3.1 SYSTEM ARCHITECTURE
the proposed method did not significantly improve the cell
membrane segmentation accuracy. 3.1.1 SYSTEM DESIGN

Convolutional Neural Network with Wavelet Domain

Inputs[15] the model starts with an RGB input that
transforms images into 448x448 YCbCr format. It then
applies a 3-level Wavelet Packet Transform (WPT) or
Dual-Tree Complex Wavelet Transform (DTCWT) to
generate 192 channels. To mitigate noise and computational
complexity, it selects and consolidates key channels into a
Cx56x56 tensor. This modified tensor is then processed by
the ResNet-50 model, which adapts to the changing C to
improve the classification performance, especially focusing System Design
on low-frequency information. A challenge highlighted in
the study is the absence of a suitable data set for this Users will receive login credentials, to prevent unwanted
approach and insufficient frequency. access. Then after the user logs in, they will be provided
The method described in the article [16] by Ahmad Yahya with a homepage that will have the Diagnoses that are
Dawod, Aniwat Phaphuangwittayakul and Salita available, then in that particular diagnosis, the user can
Angkurawaranon in 2022 involves a series of steps, upload an image, which is passed to the model from there,
including image acquisition, preprocessing, Simple Linear after which the result is presented to the user.
Iterative Clustering (SLIC) , edge detection, region
growing, dilation and hybrid method. However, this
approach has certain drawbacks, including
resource-intensiveness, the need for model updates, and the
complexity associated with fine-tuning the methodology.
3.1.2 SYSTEM OVERVIEW
3. PROPOSED WORK The web application for medical image diagnosis leverages
We have used a model which will have a base model as an a technology stack that combines the front-end capabilities
EfficientNetB3 and using the imageNet weights. The input of React.js for a responsive and dynamic user interface with
which a image of any imaging scan is sent into the model, the back-end powered by Python, utilizing Flask as the web
after which the convolution operation done on the image, framework, and TensorFlow for deep learning with the
and the features are highlighted. Then we apply some noise EfficientNetB3 model. This seamlessly integrated tech
to the data for regularization, to prevent overfitting, and to stack enables users to upload medical images, and the
help the model to focus on only the essential part. Then we system's Flask-based server facilitates the interaction
use batchNormalization that normalizes the activation of between the front-end and the deep learning model,
each layer. allowing for efficient image processing and diagnosis.
 model is a fundamental step in developing healthcare
diagnostic or analytical tools. This process initiates with the
acquisition of medical images using specialized equipment,
like X-ray machines, MRI scanners, or ultrasound devices.
Ensuring equipment accuracy and quality through regular
maintenance and calibration is critical to obtain reliable
data.

3.2.2 DATA LABELING:

System
flow Medical images often require manual annotation, which
includes identifying regions of interest (ROIs) and
3.1.3 USER INTERFACE DESIGN: abnormalities. Expert radiologists or trained annotators
usually perform this task to ensure the quality and accuracy
In the user interface for the medical image diagnosis web
of annotated data. Establishing ground truth data with
application, the primary input is an image upload feature.
accurate annotations and diagnoses for each image is
The homepage will feature a list of diagnosis, upon
critical as it forms the basis for training the AI ​model and
selecting one diagnosis, they are sent to a page which will
allowing it to make diagnostic predictions.
feature an Upload Image button that allows users to easily
select and submit medical images for diagnosis. Once the
image is uploaded, the interface presents the image for 3.2.3 DATA PREPROCESSING:
review and subsequently displays the diagnosis results,
offering a streamlined and user-friendly experience focused To standardize the data and to improve the image quality
solely on image-based Diagnosis we use data preprocessing techniques such as noise
reduction, normalization and augmentation. Medical
images can be noisy due to various factors, and noise
mitigation is essential for model accuracy. Ensuring
consistent intensity and resolution across images is
essential for accurate analysis. Data augmentation
techniques like rotating the data and brightness
adjustments increase the amount of data in the dataset and
helps the AI ​model make more reliable predictions. Image
registration is also used when working with multiple image
modalities or images taken at different times.

3.3 MODEL DEVELOPMENT AND TRAINING

3.3.1 EfficientNet B3

EfficientNet is a family of CNN architectures designed for

efficient and effective deep learning. It is trained on
ImageNet-1K at a resolution of 300X300. It uses simple
compound coefficients to achieve better results.
User interface design
EfficientNet B3 is known for its good balance between
model size and performance, making it suitable many
computer vision tasks,

3.2. DATA COLLECTION AND

3.3.2 Transfer Learning
PREPROCESSING

Leveraging pre-trained models like EfficientNet B3 is a

3.2.1 COLLECTION OF DATA
common approach in deep learning. Pre-trained model
weights can be used in large datasets as a starting point for
Collecting medical images data for feeding into an AI
many medical image analysis tasks. This can significantly 3.4.3 IMAGE UPLOAD AND PROCESSING:
speed up training and potentially improve performance,
especially if you have limited labeled medical data. The image upload and processing stage is a critical part of
the workflow, ensuring that the system can efficiently
3.3.3 Fine-Tuning: analyze medical images. Users are guided to click the
Upload Image button, followed by the selection of a
After initializing the model with pre-trained weights, medical image from their device. Once uploaded, the
fine-tuning the model on the medical image dataset is system proceeds to process the image for further analysis.
essential. Adjust the top layers of the model to adapt it to This preprocessing may include resizing and normalization
your specific task. The output layer needs to be modified to to ensure the image is in the appropriate format and size
the number of output classes for the diagnostic task. required by the machine learning model.

3.3.4 Data Augmentation: 3.4.4 DIAGNOSIS PROCESS:

Augmenting dataset during training by applying After the image is submitted, the image is handed over to
transformations like rotation, scaling, and flipping we can the deep learning model, EfficientNetB3, for the actual
get a wider variety of data for training which will help in diagnosis. The model analyzes the image to generate a
generalizing the model and prevent overfitting. Data prediction for the medical condition and produces a
augmentation helps the model become robust to variations confidence score to assess the certainty of the diagnosis.
in input data. This step is the core of the application, where the machine
learning model leverages its capabilities to provide
valuable medical insights.
3.3.5 Regularization:
3.4.5. RESULT PRESENTATION:
To prevent overfitting, we have added regularization
techniques such as dropout and weight decay to the model.
EfficientNet B3 has a significant number of parameters, The diagnosis results are presented to the user in a
and regularization can help improve generalization. user-friendly and understandable format. This presentation
includes the predicted medical condition, the associated
confidence score.
3.4. SYSTEM WORKFLOW
3.4.6. DATA SAFETY AND PRIVACY:
3.4.1 USER LOGIN:
Throughout the workflow, robust data safety and privacy
Upon accessing the application, users are presented with measures are upheld. This includes secure data
the login screen. Here, users input their credentials, transmission protocols, temporary data storage during the
typically consisting of an email address and password, user's session, and user authentication. The application
which are securely verified for authentication. Once prioritizes data protection to safeguard sensitive patient
successfully logged in, users gain access to the application's information and medical images, following industry
features. standards and regulations to ensure user data remains
confidential and secure.
3.4.2 HOMEPAGE AND IMAGE UPLOAD :
3.4.7 LOGOUT AND USER MANAGEMENT:
Upon successful login, users are directed to the
application's homepage, which serves as the central hub for Users have the option to log out of their accounts when
initiating the diagnosis process. The homepage is designed they have completed the diagnosis. Additionally, the
to be intuitive and user-friendly, featuring a conspicuous application offers user management features that empower
Upload Image button or area where users can easily select users to maintain their accounts, including the ability to
and submit their medical images for analysis. This step is reset passwords and configure account settings. This
crucial for initiating the diagnosis process and streamlining provides a seamless and user-centric experience while
the user experience. allowing users to have control over their accounts.
4. CONCLUSION [7] Yanan Sun, Bing Xue, Mengjie Zhang, Gary and
Jiancheng Lv, “Automatically Designing CNN
In the culmination of this project, we have successfully Architectures Using Genetic Algorithms for Image
developed a user-centric medical image diagnosis web Classification”. arXiv:1808.03818v3, [cs.NE], 27 march
application. By removing the individual user registration 2020
process we have made sure that there is no unauthorized
access to the tool. This streamlined workflow, coupled with [8] Shengyi Zhao, Yun Peng, Jizhan Liu * and Shuo Wu.
efficient image processing and clear result presentation, “Tomato Leaf Disease Diagnosis Based on Improved
facilitates prompt and comprehensive medical image Convolution Neural Network by Attention Module”.
diagnoses. The integration of the EfficientNetB3 deep Agriculture 2021, 11, 651. 11 July 2021
learning model enhances the accuracy and reliability of the
results, contributing to informed healthcare decisions. As [9] S. Gowri, Judith Justin and R. Vanithamani. S. Gowri,
we move forward, ongoing monitoring and maintenance Judith Justin and R. Vanithamani. “An Improved
will be essential to uphold the application's continued Classification of Mr Images for Cervical Cancer Using
reliability and accuracy. This project exemplifies the Convolutional Neural Networks”. ICTACT journal on
delicate balance between user-friendliness and data image and video processing, volume: 12, issue: 02,
security, underlining its potential as a valuable tool in the november 2021
realm of healthcare, benefiting both medical professionals
and patients.
[10] Yiwen Liu , Xian Zhang, Yanxia Gao,Taiguo Qu, and
Yuanquan Shi. “Improved CNN Method for Crop Pest
Identification Based on Transfer Learning”. Hindawi
Computational Intelligence and Neuroscience Volume
REFERENCE: 2022, Article ID 9709648, Published 16 March 2022

[1] Tsung-Yu LIN, Subhransu Maji, Improved Bilinear [11] Sota Kato Kazuhiro Hottal, “Cell Segmentation by
Pooling with CNNs, arXiv:1707.06772v1 [cs.CV], 21 july Image-to-Image Translation using Multiple Different
2017 Discriminators''.DOI: 10.5220/0009170103300335, Volume
4: BIOSIGNALS, published in 2022

[2] Ishani Dabral,Maheep Singh,Krishan Kumar, Cancer

Detection using CNN, published in 2018 [12] Fetulhak Abdurahm Kinde Anlay Fante. “Tile-based
microscopic image processing for malaria screening using
deep learning approach”. April 19th, 2022
[3] Yu HuID1, Yongkang WongID2, Wentao Wei1, Yu
Du1, Mohan Kankanhalli3, Weidong GengID1, “A novel
attention-based hybrid CNN-RNN architecture for [13] Md. Ashiq Mahmood Tamal Joyti Roy. “A Hybrid
sEMG-based gesture recognition”. PLoS ONE 13(10): Approach to Find COVID-19 Related Lung Infection
e0206049. October 30, 2018 Utilizing 2 Bit Image Processing, 19 February 2022.

[4] Wei Fang, Feihong Zhang1, Victor S. Sheng3 and [14] Sota Kato Kazuhiro Hotta. Cell Segmentation by
Yewen Ding1. “A Method for Improving CNN-Based Image-to-Image Translation using Multiple Different
Image Recognition Using DCGAN”. CMC, vol.57, no.1, Discriminators. DOI: 10.5220/0009170103300335, Volume
pp.167-178, 2018 4: BIOSIGNALS, published in 2022

[5] Abien Fred M. Agarap. “An Architecture Combining [15] Luyuan Wang Yankui Sun Image classification using
Convolutional Neural Network(CNN) and Support Vector convolutional neural network with wavelet domain inputs.
Machine (SVM) for Image Classification.” IET Image Process, DOI: 10.1049/ipr2.12466, 21 February
arXiv:1712.03541v2, 7 Feb 2019 2022

[6] Shaojuan Li, Lizhi Wang, Jia Li and Yuan Yao. “Image [16] Ahmad Yahya Dawod Aniwat Phaphuangwittayakul
Classification Algorithm Based on Improved AlexNet”. Salita Angkurawaranon. A hybrid method for traumatic
Shaojuan Li et al 2021 J. Phys.: Conf. Ser. 1813 012051 brain injury lesion segmentation. International Journal of
Electrical and Computer Engineering (IJECE) Vol.12,
No.2, April 2022, pp.1437~1448
PLAGIARISM REPORT FOR PHASE 1 PAPER
 Medical Image Diagnosis In Healthcare
Utilizing A Web Application

Dr.P.Shanmugam Keshav S.R Mithilesh Kumaar J.S

Assistant Professor Student Student
Department of Department of Department of
Computer Science and Computer Science and Computer Science and
Engineering Engineering Engineering
Rajalakshmi Engineering Rajalakshmi Engineering Rajalakshmi Engineering
College Chennai, India College Chennai, India College Chennai, India
shanmugam.p@rajalakshmi.e 200701124@rajalakshmi.edu. 200701145@rajalakshmi.edu.
du.in in in

Abstract:
In today's fast-paced world, the need for accurate and effective diagnoses in the healthcare industry has become increasingly
important. The main idea behind diagnostics is to reduce human error in medicine; This is a principle that extends medicine
to other fields, such as satellite Earth observation and understanding operations in space. The main purpose in developing
our work is to provide doctors with reliable tools to predict and solve customers' health problems on the online platform. We
also aim to improve the overall user experience by implementing functions to manage user history and preferences. The core
technologies of our solution include the powerful EfficientNet B3 for neural networks (CNN) and image processing, as well
as extensive use of React.js to create an interactive web front end. Our work focuses on image data and solves problems such
as overhead, hyperparameter sensitivity, and time consumption in image processing. By addressing these issues, our goal is to
provide physicians with fast, accurate results that meet their rapid decision-making and care needs, good patients.

Related Work overuse, limited use for certain situations, and problems
Tsung-Yu Lin et al. [1] discussed the design of bilinear with the use of defined roles.
pooling, which combines the products of two faces with
global competition to form a different matrix leading to the Sing CNN for Cancer Detection (2018) [2] Ishani Dabral et
interaction effect. Although these two features are similar, al. proposed a neural network (CNN) architecture for
they lead to the parallelism of B-CNN, reminiscent of cancer diagnosis targeting the Invasive Ductal Cancer
second-order (O2P), a technique used in semantic (IDC) dataset. This innovation combines convolution,
classification. However, this approach suffers from several output and deconvolution techniques with custom data
limitations, such as sensitivity to hyperparameters, risk of priors. However, it faces problems such as the need for
large data sets, complex models, excessive risk, and A method using the image classification algorithm AlexNet
difficulty in implementing the transformation process. [6], the author proposed a method using the development of
AlexNet convolutional neural network (CNN) for image
A tracking system that recognizes gestures based on gesture classification. This method solves many inconsistency
recognition and surface electromyographic signals (sEMG) problems of traditional CNN, especially in the fully
based on the CNN-RNN hybrid architecture (2018). While connected process, which will cause excessive interference
CNNs extract features from images, RNNs, especially and increase the complexity of the process. This
short-term (LSTM) units, model the body of the object. development introduced the deconvolution process into the
Tracking techniques have been introduced to highlight AlexNet architecture, specifically reducing the number of
important information in sEMG signals. The architecture parameters in all layers. However, it is worth noting that
uses a combination of redundancy, including shock and this development also brings with it an increase in
target redundancy. computer complexity, the risk of overfitting due to model
complexity, and the need for large amounts of training data.
Challenges with this approach include fading and breakage Achieving a balance of these factors is critical for the
gradient issues, which can disrupt training and make successful development of the AlexNet-based image
long-term progress difficult to make. RNNs also face classification algorithm.
limitations regarding short time intervals; They are good at
working with short segments but have difficulty Yanan Sun et al. proposed a method to optimize the
remembering information about events in the distant past. performance of convolutional neural networks (CNN) using
genetic algorithms (GA). The process starts by initializing a
set of CNN architectures, coding them layer by layer and
Wei Fang et al. The generator transforms noise vectors into removing entire layers. Although this approach seems
real patterns, while the operator distinguishes between real promising, it faces challenges such as reproducibility
and generated data. Training methods include non-convex issues, information requirements, flexible adaptations,
games and adaptive learning reduction strategies are used overall limits, material and time requirements.
to improve learning performance. However, this approach
suffers from several problems such as training instability, Shengyi Zhu et al. This model uses the Compression and
poor performance (the generator is diverse), excessive Excitation Network (SENet) method to assign weight
hyperparameter sensitivity, and requires significant coefficients to channels based on their importance. This
computer services due to its severe nature. recalibration process improves the performance of the film
without any particular problems. However, this approach
Abien et al[5] discussed an experimental method still suffers from many problems, such as computational
combining SVM and CNN for image classification. This complexity, reduced interpretability, and the need for
hybrid architecture uses SVM for binary classification to careful tuning of hyperparameters.
determine the best hyperplane separating two classes. This
work considers both L1-SVM and L2-SVM by using hinge In improving the classification of Cervical Cancer MRI
loss as the loss model for SVM. However, this article Images Using Convolutional Neural Networks [9] A report
highlights issues such as lack of understanding of the on a method for identifying Hepatocellular carcinoma. It
distribution process, issues of injustice and justice, ethics or uses a learning concept combined with seven machine
lack thereof, and limitations with long-term memory. learning algorithms and uses collaborative representation
classification and optimization techniques to classify
hyperspectral images. However, this approach faces
challenges such as lack of information and the ability to Kumar, et.al[14] proposed a facial recognition system using
limit movement between different locations. Local Binary Pattern algorithm; it focuses on local patterns,
comparing each pixel to its neighbors and encoding the
Yiwen Liu et al. It uses VGG16 and Inception-ResNet-v2 results into a binary format. This method might be sensitive
(CNN) convolutional neural networks for transfer learning. to change in the lighting and the change in poses. While
However, this approach faces challenges such as resource efficientNetB3 being a deep learning model can learn the
consumption, need for updated models, and difficulties hierarchical features in an efficient way, and can handle the
with optimization. change in pose, expressions and lighting.

Sota Kato et al. [11] presented a study from Kyoto

University involving 50 mobile phones. Images were Convolutional Neural Network with Wavelet Domain Input
obtained using fluorescent markers of mouse liver cell [15] This model starts with RGB input and converts the
membranes and nuclei and were 256 x 256 pixels in size. image to 448x448 YCbCr format. It then uses the 3rd-order
40 of these images were used for training, 5 for validation, wavelet packet transform (WPT) or double-tree complex
and the remaining 5 for testing. However, it is worth noting wavelet transform (DTCWT) to create 192 channels. It
that the scheme cannot improve the accuracy of cell selects the main channels and combines them into Cx56x56
membrane segmentation. tensors to reduce noise and computational complexity. This
modified tensor is then processed by the ResNet-50 model;
Image processing for malaria screening using deep learning This model modifies C to improve the performance of
[12], Fetulhak Abdurahm et al. This study aims to use deep classification by paying special attention to low-frequency
learning material to improve the detection of Plasmodium data. The biggest challenge in research is the lack of
falciparum in thick smear microscopy. Sample check. The suitable and insufficient data for this approach.
aim is to solve previously identified problems in this area.
To achieve this goal, the reporting process is based on The process described in the article [16] published by
images that can see small objects in higher resolution Ahmad Yahya Dawod, Aniwat Phaphuangwittayakul and
images that exceed the resolution of the network. However, Salita Angkurawaranon in 2022 involves several steps,
it should be noted that this approach may not achieve a including finding images, first simple linear iterative
high level of detection performance, especially when clustering (SLIC), edge detection, growing region,
changing the image to be detected. expansion . and hybrid methods. However, this approach
has some disadvantages, such as resource usage, the need
Art. Ashiq Mahmood and Tamal Joyti Roy [13] conducted for up-to-date models, and difficulties with quality systems.
an experiment with image analysis to develop an
assessment system to identify COVID-19-associated lung 3. PROPOSED WORK
diseases. This method primarily involves analyzing images
The model we will use will be EfficientNetB3 as the base
without converting them into binary parts. Then the authors
model and we will mainly use imageNet. Transfer an image
converted the normal image to binary and recorded the
from scanning to an input on the model; this then convolves
corresponding results. This experiment also uses
the image and highlights features. We then apply some
parameter-based neural networks. However, one of the
regular noise to the data to prevent overfitting and help the
problems faced by this study is that there is not enough data
model focus on only important values. We then use
for analysis. Image segmentation is controlled by local
BatchNormalization to normalize the performance of each
segmentation.
layer.
3.1 SYSTEM ARCHITECTURE

3.1.1 SYSTEM DESIGN

Fig 3. System flow

3.1.3 USER INTERFACE DESIGN:

The main concept in the user interface of the medical

imaging diagnostic web application is the image upload
function. A list of diagnoses will be displayed on the home
page, and once a diagnosis is selected, they will be sent to a
page with image uploads, allowing users to easily select
and upload medical images for diagnosis. Once the image
is loaded, the interface displays the image for review and
Fig 1. System Design then displays diagnostic results, providing links and user
experience.
Users will be issued a login certificate to prevent
unauthorized access. Then, once the user logs in, they are
presented with a home page with available diagnoses and
then on a particular diagnosis the user can send images
from there which are transferred to the model and the
results are then presented to the user.

3.1.2 SYSTEM OVERVIEW

The web application for medical diagnostics uses a set of

technologies that combine the front-end capabilities of
React.js to respond to user interactions with the back-end
driven by Python using Flask as a web platform and
TensorFlow for deep learning. EfficientNetB3 model. This
integrated technology allows users to transmit medical
images, and the Flask-based server system, front-end and
absorption of dynamic learning models facilitate the
interaction of image processing and diagnostics.

Fig 2. User interface design

3.2. DATA COLLECTION AND PREPROCESSING

3.2.1 COLLECTION OF DATA

Collecting clinical data to feed cognitive models is an

important step in the development of a diagnosis or
diagnostic tool. The process begins with capturing medical
images using specialized equipment such as x-ray It is important to weight the model before training and then
machines, MRI scanners or ultrasound equipment. Ensuring adjust the model based on clinical data. Customize the top
product accuracy and performance through regular layer of the template to fit your specific project. For
monitoring and calibration is essential to obtain reliable diagnosis, the urine method must be adjusted according to
data. the number of urine groups.

3.2.2 DATA LABELING: 3.3.4 Data Augmentation:

Medical imaging often requires manual guidance that By improving the data during training using
includes identifying regions of interest (ROIs) and transformations such as rotation, scaling, and translation,
abnormalities. This work is usually done by an electrician we can obtain more training data, which will help make the
or trained scribe to ensure the quality and accuracy of the model larger and avoid collisions. Data augmentation helps
recorded data. Establishing a ground truth with accurate the model make changes to the input data.
descriptions and diagnoses for each image is important
because this forms the basis for training the AI ​model and
3.3.5 Regularization:
allows it to make future diagnoses earlier.

3.2.3 DATA PREPROCESSING: To prevent overfitting, we add regularization parameters

such as dropout and weighting to the model. EfficientNet
B3 has many parameters and continuous operation helps
We use data processing techniques such as noise reduction,
increase overall efficiency.
normalization and enhancement to create data and improve
image quality. Medical images can be noisy for a variety of
reasons, and it is important to reduce noise for accurate 3.4. SYSTEM WORKFLOW
samples. Keeping the image consistent and resolved is
important for accurate analysis. Data enhancement 3.4.1 USER LOGIN:
techniques such as data transformation and brightness
adjustment increase the information in the dataset and help Once in the application, users will see the login screen.
AI models make more reliable predictions. The registration Here, users enter credentials that are checked for
form can also be used when taking different types of authentication, usually including email address and
photographs or photographs taken at different times. password. After successfully logging in, the user can access
the functions of the application.
3.3 MODEL DEVELOPMENT AND TRAINING
3.4.2 HOMEPAGE AND IMAGE UPLOAD :
3.3.1 EfficientNet B3
Upon successful login, the user is sent to the main page of
EfficientNet is a family of CNN architectures designed for the application and this forms the basis for starting the
deep and efficient learning. Trained on ImageNet-1K at diagnostic process. The homepage design is intuitive and
300X300 resolution. It uses simple coefficients to get better user-friendly, featuring a prominent image upload button or
results. EfficientNet B3 is known for its good balance area where users can easily select medical images and
between model size and performance, making it suitable for submit them for analysis. This step is important to start the
many computer vision systems. diagnostic process and improve the user experience.

3.3.2 Transfer Learning 3.4.3 IMAGE UPLOAD AND PROCESSING:

Using advanced learning models such as EfficientNet B3 is The image loading and processing phase is the main work
one method in deep learning. Pre-trained weight models to ensure that the system can analyze the medical image
can be used as a starting point for many types of image well. Tell users to click the "Upload Image" button and
analysis on large data sets. This can increase training speed select a medical image from their device. Once uploaded,
and potentially improve performance, especially if you the system continues to process images for further analysis.
have limited clinical experience. This preprocessing may include transforming and
normalizing to ensure that the image is in the appropriate
3.3.3 Fine-Tuning: format and size required by the machine learning model.
3.4.4 DIAGNOSIS PROCESS:

After the image is sent, the image is sent to the

EfficientNetB3 deep learning model for actual diagnosis.
This model evaluates images to generate predictions of
clinical conditions and increases confidence in assessing
the accuracy of diagnosis. This step is the main part of the
application where the machine learning model uses its
ability to provide treatment recommendations.

3.4.5. RESULT PRESENTATION:

Diagnostic results are presented to users in a user-friendly After that, the user will have to select the diseases they are
and easy-to-understand format. This display includes willing to diagnose. There will be a list available form
predicted treatments and confidence scores. which the user can select.

3.4.6. DATA SAFETY AND PRIVACY:

We comply with data security and confidentiality measures

throughout our work. This includes security data transfers,
temporary data storage during user sessions, and user
authentication. The app focuses on data protection to
protect patient data and medical images in accordance with
industry standards and regulations to ensure the privacy and
security of user data.
Then there will be a dialog where the user can select the
3.4.7 LOGOUT AND USER MANAGEMENT: image, and upload it to the server which will give the result
in a human readable format.
Once the diagnosis is complete, users can cancel their
account. In addition, the app provides user management
features that allow users to manage their accounts,
including resetting passwords and creating accounts. This
allows users to manage their money while providing
consistency and a great user experience.

4. IMPLEMENTATION AND RESULT

After the login there is the Dashboard in which all the

medical image diagnoses are present. The user can select In the above image, the X Ray is selected. Where there is a
from the listed diagnoses. list of diseases that can be diagnosed with the X ray. Then
The user should upload the image of a chest X ray.
 [1] Tsung-Yu LIN, Subhransu Maji, Improved Bilinear
Pooling with CNNs, arXiv:1707.06772v1 [cs.CV], 21 july
2017

[2] Ishani Dabral,Maheep Singh,Krishan Kumar, Cancer

Detection using CNN, published in 2018

[3] Yu HuID1, Yongkang WongID2, Wentao Wei1, Yu

Du1, Mohan Kankanhalli3, Weidong GengID1, “A novel
attention-based hybrid CNN-RNN architecture for
sEMG-based gesture recognition”. PLoS ONE 13(10):
e0206049. October 30, 2018

[4] Wei Fang, Feihong Zhang1, Victor S. Sheng3 and

Yewen Ding1. “A Method for Improving CNN-Based
Then when the image is set to the server. Where the model Image Recognition Using DCGAN”. CMC, vol.57, no.1,
predicts the disease and then the result is sent in a human pp.167-178, 2018
readable format.
[5] Abien Fred M. Agarap. “An Architecture Combining
Convolutional Neural Network(CNN) and Support Vector
Machine (SVM) for Image Classification.”
arXiv:1712.03541v2, 7 Feb 2019

[6] Shaojuan Li, Lizhi Wang, Jia Li and Yuan Yao. “Image
Classification Algorithm Based on Improved AlexNet”.
Shaojuan Li et al 2021 J. Phys.: Conf. Ser. 1813 012051

[7] Yanan Sun, Bing Xue, Mengjie Zhang, Gary and

Jiancheng Lv, “Automatically Designing CNN
Architectures Using Genetic Algorithms for Image
Classification”. arXiv:1808.03818v3, [cs.NE], 27 march
2020

5. CONCLUSION
[8] Shengyi Zhao, Yun Peng, Jizhan Liu * and Shuo Wu.
“Tomato Leaf Disease Diagnosis Based on Improved
With the completion of this project, we completed the Convolution Neural Network by Attention Module”.
development of the client's web analytics application. By Agriculture 2021, 11, 651. 11 July 2021
removing the user registration process, we ensure that there
is no unauthorized access to the device. This simple
[9] S. Gowri, Judith Justin and R. Vanithamani. S. Gowri,
process, combined with image processing and presentation
Judith Justin and R. Vanithamani. “An Improved
of clear results, facilitates timely and successful diagnoses.
Classification of Mr Images for Cervical Cancer Using
Integration of the EfficientNetB3 deep learning model
Convolutional Neural Networks”. ICTACT journal on
helps make clinical decisions by increasing the accuracy
image and video processing, volume: 12, issue: 02,
and reliability of results. As we move forward, ongoing
november 2021
monitoring and maintenance is critical to maintain the
reliability and accuracy of the application. By
demonstrating the balance between user-friendliness and [10] Yiwen Liu , Xian Zhang, Yanxia Gao,Taiguo Qu, and
information security, the project demonstrates its potential Yuanquan Shi. “Improved CNN Method for Crop Pest
to be a useful, important tool in medicine for doctors and Identification Based on Transfer Learning”. Hindawi
patients. Computational Intelligence and Neuroscience Volume
2022, Article ID 9709648, Published 16 March 2022
REFERENCE:
[11] Sota Kato Kazuhiro Hottal, “Cell Segmentation by 2023,pp.1-6, doi: 10.1109/ViTECoN58111.2023.10157843.
Image-to-Image Translation using Multiple Different
Discriminators''.DOI: 10.5220/0009170103300335, Volume [15] Luyuan Wang Yankui Sun Image classification using
4: BIOSIGNALS, published in 2022 convolutional neural network with wavelet domain inputs.
IET Image Process, DOI: 10.1049/ipr2.12466, 21 February
[12] Fetulhak Abdurahm Kinde Anlay Fante. “Tile-based 2022
microscopic image processing for malaria screening using
deep learning approach”. April 19th, 2022 [16] Ahmad Yahya Dawod Aniwat Phaphuangwittayakul
Salita Angkurawaranon. A hybrid method for traumatic
[13] Md. Ashiq Mahmood Tamal Joyti Roy. “A Hybrid brain injury lesion segmentation. International Journal of
Approach to Find COVID-19 Related Lung Infection Electrical and Computer Engineering (IJECE) Vol.12,
Utilizing 2 Bit Image Processing, 19 February 2022. No.2, April 2022, pp.1437~1448

[14] K. P, S. L. T. A and S. R, "Face Recognition

Attendance System Using Local Binary Pattern
Algorithm," 2023 2nd International Conference on Vision
Towards Emerging Trends in Communication and
Networking Technologies (ViTECoN), Vellore, India,
PLAGIARISM REPORT FOR PHASE 2 PAPER
 CO-PO MAPPING

CO PO ATTAINMENT FOR PROJECT

1. On completion the students are capable of executing the proposed plan and
become aware of and overcome the bottlenecks throughout every stage.

2. On completion of the project work students could be in a role to take in any

difficult sensible issues and locate answers through formulating the right
methodology.

3. Students will attain a hands-on level in changing a small novel idea / method
right into an operating model / prototype related to multidisciplinary abilities and /
or understanding and operating in a team.

4. Students will be able to interpret the outcome of their project. Students will take
on the challenges of teamwork, prepare a presentation in a professional manner,
and document all aspects of design work.

5. Students will be able to publish or release the project to society.

PROJECT WORK COURSE OUTCOME (COs):

CO1: On completion the students are capable of executing the proposed plan and
become aware of and overcome the bottlenecks throughout every stage.

CO2: On completion of the project work students could be in a role to take on any
difficult sensible issues and locate answers through formulating the right
methodology.

CO3: Students will attain a hands-on level in changing a small novel idea / method
right into an operating model / prototype related to multi- disciplinary abilities and
/ or understanding and operating in a team.

CO4: Students will be able to interpret the outcome of their project. Students will
take on the challenges of teamwork, prepare a presentation in a professional
manner, and document all aspects of design work.

CO5: Students will be able to publish or release the project to society.

PROGRAM OUTCOMES (POs):
PO1: Engineering knowledge: Apply the knowledge of mathematics, science,
engineering fundamentals, and an engineering specialization to the solution of
complex engineering problems.
PO2: Problem analysis: Identify, formulate, review research literature, and analyze
complex engineering problems reaching substantiated conclusions using first
principles of mathematics, and engineering sciences.
PO3: Design / development of solutions: Design solutions for complex
engineering problems and design system components or processes that meet the
specified needs with appropriate consideration for the patients health and safety,
and medical needs.
PO4: Conduct investigations of complex problems: Use research - based
knowledge and research methods including design of experiments, analysis and
interpretation of data, and synthesis of the information to provide valid
conclusions.
PO5: Modern tool usage: Create, select, and apply appropriate techniques,
resources, and modern engineering and IT tools including prediction and modeling
to complex engineering activities with an understanding of the limitations.
PO6:The engineer and society: Apply reasoning informed by the contextual
knowledge to assess societal, health, safety, legal and cultural issues and the
consequent responsibilities relevant to the professional engineering practice and
medical practices.
PO7: Environment and sustainability: Understand the impact of the professional
engineering solutions in societal and environmental contexts, and demonstrate the
knowledge of, and need for sustainable development.
PO8: Ethics: Apply ethical principles and commit to professional ethics and
responsibilities and norms of the engineering practice.
PO9: Individual and team work: Function effectively as an individual, and as a
member or leader in diverse teams, and in multidisciplinary settings.
PO10: Communication: Communicate effectively on complex engineering
activities with the engineering community and with society at large, such as, being
able to comprehend and write effective reports and design documentation, make
effective presentations, and give and receive clear instructions.
PO11: Project management and finance: Demonstrate knowledge and
understanding of the engineering and management principles and apply these to
one’s own work, as a member and leader in a team, to manage projects and in
multidisciplinary environments.
PO12: Life-long learning: Recognize the need for, and have the preparation and
ability to engage in independent and life-long learning in the broadest context of
technological change.
PO13: Learning: Recognize the need for, and have the preparation and ability to
engage in independent and life-long learning in the broadest context of
technological change.
PROGRAM SPECIFIC OUTCOMES (PSOs)
PSO1: Foundation Skills: Ability to understand, analyze and develop computer
programs in the areas related to algorithms, system software, web design, machine
learning, data analytics, and networking for efficient design of computer-based
systems of varying complexity. Familiarity and practical competence with a broad
range of programming languages and open-source platforms.

PSO2: Problem-Solving Skills: Ability to apply mathematical methodologies to

solve computational tasks, model real world problems using appropriate data
structure and suitable algorithms. To understand the Standard practices and
strategies in software project development using open-ended programming
environments to deliver a quality product that has been done.

PSO3: Successful Progression: Ability to apply knowledge in various domains to

identify research gaps and to provide solutions to new ideas, inculcate passion
towards higher studies, creating innovative career paths to be an entrepreneur and
evolve as an ethically socially responsible computer science professional.

0/PS PO PO PO PO PO PO PO PO PO PO PSO PSO PSO

OCO 1 2 3 4 5 6 7 9 11 12 1 2 3
CO1 3 3 3 3 3 1 2 2 2 2 - 1 2
CO2 2 1 3 1 2 1 2 1 2 1 1 2 1
CO3 3 2 3 1 2 3 2 1 2 - 3 2 1
CO4 3 1 3 2 1 2 3 1 2 3 1 2 2
CO5 2 3 1 2 3 1 2 3 1 1 2 2 3
average 2.6 2 2.6 1.8 2.2 1.6 2.2 1.6 1.8 1.4 1.4 1.8 1.8