BIOMEDICAL IMAGE ANALYSIS USING DEEP

LEARNING FOR DISEASE DETECTION

P.OYYATHEVAN 1 ,CH.S.S.SUDARSHAN 2 , S.ANUPAMA 3 , K.N.V.RAMJI 4 , K.KEERTHANA 5

ASSO.PROFESSOR 1, U.G 2,3,4,5

DEPARTMENT OF COMPUTER SCIENCE AND TECHNOLOGY

SASI INSTITUTE OF TECHNOLOGY AND ENGINEERING, ANDHRA PRADESH, INDIA

ABSTRACT

Lung cancer remains one of the leading causes of cancer-related deaths worldwide, highlighting the critical
need for accurate and efficient diagnostic methods. This study presents a novel approach for lung tumor
detection using the VGG16 convolutional neural network architecture. The VGG16 model is leveraged for
feature extraction from lung imaging data, focusing on its deep feature representation capabilities to enhance
tumor classification accuracy. Preprocessed CT scan images are input into the VGG16 model, which extracts
hierarchical features through its convolutional layers. The extracted features are then classified using
conventional machine learning techniques, such as support vector machines (SVM) or logistic regression, to
distinguish between malignant and benign tumors. The proposed method demonstrates improved detection
performance, emphasizing the importance of deep feature extraction in medical image analysis. The results
indicate that VGG16-based feature extraction combined with traditional classification methods provides a
reliable and efficient solution for lung tumor detection.

KEYWORDS: Lung tumor detection, VGG16, feature extraction, CT scan, convolutional neural network,
classification, medical imaging.

1. INTRODUCTION

Biomedical image analysis is crucial for techniques, however, often face challenges in
diagnosing, monitoring, and treating various accurately identifying complex patterns and subtle
medical conditions. It provides detailed insights anomalies within these images. [2]. Automated
into the internal structures and functions of the systems using Convolutional Neural Networks
human body, enabling accurate disease detection (CNNs) have proven highly effective in analyzing
and effective treatment planning. Advanced lung images from modalities like X-rays and CT
techniques like Machine Learning (ML) and Deep scans. By leveraging deep learning's ability to
Learning (DL) enhance the capabilities of extract meaningful patterns, these systems can
biomedical imaging, offering improved accuracy identify abnormalities such as lung nodules, pleural
and speed in medical diagnostics. Biomedical effusion, or other pathological signs at early stages.
image analysis plays a crucial role in modern This allows for timely intervention and better
healthcare by facilitating the early detection, management, ultimately enhancing diagnostic
diagnosis, and treatment of various diseases, accuracy and enabling improved patient care in
including cancer. The analysis of medical images clinical settings. [3]. In the context of lung disease
such as CT scans, MRI, and X-rays provides detection, CNNs can process high-resolution
valuable insights into the structure and function of medical images such as chest X-rays and CT scans
tissues and organs. Traditional image analysis to identify abnormalities. By feeding these images
into the CNN, the network processes the data behind my work are rooted in leveraging
through multiple layers of convolutional filters advancements in deep learning to create efficient
designed to detect features such as textures, edges, and reliable systems for lung disease detection.
and patterns associated with various lung Lung diseases, such as cancer and pneumonia, are
conditions. These features help the system classify among the leading causes of mortality and require
images into categories, identifying early indicators early and accurate diagnosis to improve patient
of diseases such as pneumonia, lung cancer, or outcomes.[9]. Another key intention behind my
COVID-19. This automated process can work is to leverage deep learning advancements to
outperform traditional, manual diagnostic methods, enhance the precision and efficiency of lung
offering high precision while reducing subjectivity disease diagnosis, addressing the growing demand
and time constraints.[4]. Implementing CNNs for for reliable and automated medical imaging
lung disease detection involves collecting high systems. Lung diseases like cancer and pneumonia
resolution chest X-rays and CT scan images, require timely and accurate detection, yet
labeling them with relevant disease categories, and traditional diagnostic methods often encounter
preprocessing them to enhance quality and reduce challenges such as variability in interpretations and
noise. These images are then used to train the CNN, delays [10]. The medical field is increasingly
which processes data through multiple relying on artificial intelligence to enhance
convolutional layers to identify features like diagnostic accuracy and patient care. Deep
textures and edges indicative of lung abnormalities. learning, a subset of AI, is particularly well-suited
Once trained, the model is evaluated for accuracy for medical image analysis, where it can identify
and robustness, ensuring reliable automated subtle patterns and features that may be missed by
detection of conditions like pneumonia and lung human interpretation [11]. In the era of big data,
cancer. [5]. The use of deep learning in disease healthcare is increasingly turning to deep learning
detection is not confined to a single application. for the analysis of complex medical images. By
Approaches such as those utilizing Convolutional automating the interpretation of imaging data, deep
Neural Networks (CNNs) can extend to various learning models are revolutionizing how clinicians
types of diseases, improving diagnostic accuracy diagnose diseases, improving both speed and
and efficiency. Combining IoT devices with accuracy [12]. Deep learning is at the forefront of
synergic deep learning models allows for advanced innovations in medical imaging, offering
classification techniques, such as detecting and unparalleled accuracy and efficiency in diagnosing
classifying medical anomalies.[6]. The primary diseases. This journal examines how deep learning
goal of this research is to develop a robust system techniques are reshaping the landscape of medical
for the early detection and classification of Lung image analysis, making healthcare more precise
diseases using deep learning, particularly and personalized [13]. As deep learning technology
Convolutional Neural Networks (CNNs). By continues to evolve, its applications in medical
leveraging high-resolution CT scan images, the image analysis are becoming increasingly
proposed system aims to accurately identify lung sophisticated. From detecting early signs of disease
abnormalities and assess severity levels. This to enhancing the quality of imaging data, deep
involves creating a computational tool capable of learning is redefining what is possible in medical
automatically detecting and classifying lung tissues diagnostics [14]. The complexity of medical
to assist medical professionals in diagnosing imaging data often makes it difficult for healthcare
efficiently. [7]. The motivation behind my work professionals to analyze manually. Deep learning is
lies in advancing the capabilities of deep learning providing a solution by automating the process and
for medical imaging, focusing on lung disease enabling faster, more accurate diagnoses,
detection and diagnosis. Lung diseases, such as especially in areas such as oncology, cardiology,
cancer, pneumonia, and chronic obstructive and neurology [15]. Deep learning has become a
pulmonary disease (COPD), pose significant global transformative force in medical image analysis,
health challenges. Early detection and accurate allowing healthcare providers to extract valuable
diagnosis are crucial for improving patient insights from complex imaging data. By leveraging
outcomes and reducing healthcare burdens. deep neural networks, clinicians can now make
However, traditional diagnostic methods often rely faster, more informed decisions that improve
on manual interpretation, which can be time- patient care [16].
consuming and subjective.[8]. The intentions
2.RELATED WORK and efficiency. This comprehensive review
underscores the transformative impact of deep
Machine learning and deep learning approach learning in medical diagnostics and its potential to
for medical image analysis: diagnosis to enhance disease detection, personalized treatment
detection plans, and overall healthcare delivery.[2]

In this journal Meghavi Rana & Megha Bhushan Medical image analysis based on deep learning
discuss the integration of machine learning (ML) approach
and deep learning (DL) techniques in medical
image analysis for disease detection and diagnosis. In this journal, Muralikrishna Puttagunta and S.
It highlights the growth of computer-aided Ravi explore the advancements and applications of
detection using DL and ML, emphasizing the deep learning in medical image analysis,
importance of medical images for accurate emphasizing its role in addressing challenges in
diagnosis. The study reviews various ML and DL traditional image processing. The paper highlights
approaches, including Convolutional Neural the use of deep learning approaches such as
Networks (CNNs) and Generative Adversarial convolutional neural networks, recurrent neural
Networks (GANs), and addresses challenges such networks, and autoencoders, which are extensively
as noise, artifacts, and inter-patient variability. It applied across medical imaging modalities,
provides a detailed analysis of 40 primary studies including X-rays, CT scans, MRI, mammography,
and presents experimental results using MRI and digital histopathology. It discusses applications
datasets. The paper concludes that DL techniques, like disease detection, classification, segmentation,
despite challenges, offer significant potential for and pattern recognition, showcasing the efficiency
improving medical image analysis. [1] of deep learning in enhancing accuracy and
reducing analysis time. The authors provide a
A review on deep learning in medical image detailed review of the evolution of artificial neural
analysis networks and their implementation in medical
diagnostics, addressing limitations and future
In this journal, S. Suganyadevi, V. Seethalakshmi, potential. This comprehensive study underscores
and K. Balasamy emphasize the use of the transformative impact of deep learning
Convolutional Neural Networks (CNNs) as a key technologies in improving healthcare through more
deep learning technique for medical image precise and efficient medical image analysis. [3]
analysis. CNNs are highlighted for their ability to
process 2D and 3D data effectively, making them Deep learning-based image processing in optical
suitable for tasks like image classification and microscopy
feature extraction. The paper explores various deep
learning models and their effectiveness in In this journal, Sindhoora Kaniyala Melanthota,
identifying, classifying, and quantifying clinical Dharshini Gopal, Shweta Chakrabarti, Anirudh
patterns across a wide range of medical fields, Ameya Kashyap, Raghu Radhakrishnan, and
including neuroimaging, retinal analysis, Nirmal Mazumder discuss the transformative role
pulmonary studies, digital pathology, and imaging of deep learning techniques in optical microscopy.
of the breast, heart, bone, stomach, and The authors highlight how deep learning has
musculoskeletal systems. It highlights key significantly improved image processing, enabling
challenges in medical image processing, such as tasks such as classification, segmentation, and
the need for extensive labeled datasets and resolution enhancement with higher accuracy and
computational resources, and discusses state-of- efficiency compared to traditional methods.
the-art deep learning approaches like convolutional Various optical microscopy techniques, including
neural networks, recurrent neural networks, and fluorescence, phase-contrast, and non-linear
autoencoders. The authors emphasize the microscopy, are explored in the context of deep
significance of integrating these advanced learning applications. The review emphasizes
techniques with big data and knowledge-based advancements in medical and biological research,
prediction models to improve diagnostic accuracy particularly in automated analysis and mobile-
phone-based microscopy for remote diagnostics. It
also addresses challenges such as computational diagnosis by analyzing color features, thereby
costs and the need for large datasets, while eliminating the limitations of traditional, subjective
showcasing state-of-the-art deep learning models methods. Simulation results on benchmark datasets
that have enhanced image quality and enabled demonstrate the superior diagnostic performance
innovative applications across diverse fields. [4] of the ASDL-TCI model, achieving high precision,
recall, and accuracy, making it a promising tool for
Deep Learning and the Future of Biomedical reliable disease detection and classification in
Image Analysis healthcare. [6]

In this journal, Monika Jyotiyana and Nishtha A Review on Alzheimer’s Disease Through
Kesswani explore the advancements and future Analysis of MRI Images Using Deep Learning
potential of deep learning techniques in biomedical Techniques
image analysis. The paper highlights the reliability
and accuracy of deep learning models, particularly In this journal, Battula Srinivasa Rao and Mudiyala
in medical applications such as image processing, Aparna provide a comprehensive review of deep
segmentation, classification, fusion, retrieval, and learning techniques applied to the analysis of MRI
computer-aided diagnosis. It discusses how images for the early detection and diagnosis of
methods like convolutional neural networks, Alzheimer’s Disease (AD). The study emphasizes
recurrent neural networks, and encoder decoder the significance of early identification to
architectures significantly enhance the analysis and implement preventative measures and discusses the
diagnosis of complex diseases. The authors advantages of deep learning over traditional
emphasize the transformative impact of these machine learning for segmenting brain structures
techniques in handling vast datasets and extracting and categorizing AD. Convolutional neural
precise features for tailored treatment plans. They networks (CNNs) are highlighted for their
also address the challenges, including data quality, effectiveness in analyzing brain anatomy through
computational complexity, and the need for high- segmented MRI scans. The paper also reviews
quality datasets, while pointing towards future various publicly available datasets, including
developments in leveraging deep learning for OASIS, ADNI, and MICCAI, used in brain MRI
improved patient outcomes and healthcare segmentation and AD diagnosis. Challenges in
efficiency. Overall, the journal underscores the brain MRI segmentation, such as noise, low
pivotal role of deep learning in revolutionizing the contrast, and anatomical complexity, are
field of medical imaging and diagnostics. [5] addressed, alongside potential improvements
offered by deep learning approaches. The authors
Internet of Things and Synergic Deep Learning further explore the role of brain MRI segmentation
Based Biomedical Tongue Color Image Analysis in improving classification accuracy for AD and
for Disease Diagnosis and Classification discuss the evaluation metrics for segmentation
methods. The study underscores the transformative
In this journal, Romany F. Mansour, Maha M. potential of deep learning in advancing AD
Althobaiti, and Amal Adnan Ashour propose a diagnosis and treatment. [7]
novel automated system for biomedical tongue
color image analysis to aid in disease diagnosis and Deep Learning in Medical Ultrasound Image
classification. The study integrates the Internet of Analysis: A Review
Things (IoT) and synergic deep learning techniques
to develop the ASDL-TCI model. The model This journal, authored by Xinke Ge, He Ma,
encompasses multiple stages, including data Shouliang Qi, Yu Wang, and Yudong Yao, reviews
acquisition using IoT devices, pre processing with the advancements in deep learning for medical
median filtering, feature extraction using a ultrasound image analysis. It highlights the role of
synergic deep learning network, classification deep learning in overcoming challenges such as
through a deep neural network, and parameter noise, artifacts, and subjective interpretations
optimization with enhanced black widow inherent in traditional ultrasound diagnostics. The
optimization. The approach emphasizes the study outlines common deep learning models like
significance of non invasive and efficient tongue convolutional neural networks, recurrent neural
networks, autoencoders, and transfer learning, the Echo State Network (ESN) classifier for cancer
which have been applied to tasks such as detection. Additionally, the model employs the
classification, object detection, segmentation, and Tuna Swarm Algorithm (TSA) and Adaptive Fuzzy
image reconstruction. They use image Artificial Optimization (AFAO) for
enhancement to improve the quality and reliability hyperparameter tuning and optimization,
of ultrasound data. By integrating these enhancing the accuracy and efficiency of the
technologies, the review showcases how deep system. Experimental results demonstrate the
learning is transforming ultrasound image analysis, model's superior performance, achieving a
making diagnostics more accurate and efficient.[8] maximum accuracy of 99.33% compared to
existing methods. The authors highlight the
Endoscopic Image Analysis for Gastrointestinal potential of this approach to revolutionize cancer
Tract Disease Diagnosis Using Nature Inspired diagnostics by providing a cost effective,
Algorithm With Deep Learning Approach automated, and highly accurate solution for early
detection and treatment planning.[10]
In this journal, Abdulrahman Alruban, Abdullah
Alharbi, and their team introduce the EIAGTD- Lung Cancer Classification Using Modified U-
NIADL system, a novel approach combining deep Net Based Lobe Segmentation and Nodule
learning and nature-inspired algorithms for Detection
analyzing endoscopic images of the
gastrointestinal (GI) tract. The method is designed This journal, authored by Iftikhar Naseer, Arfan
to improve the detection and classification of GI Jaffar, Sheeraz Akram, Tehreem Masood, and
diseases using a structured pipeline. Initially, Muhammad Rashid, presented a study on lung
bilateral filtering is applied for image pre cancer classification using a modified U-Net model
processing to enhance quality. Then, an improved combined with lobe segmentation and nodule
ShuffleNet model extracts features efficiently, detection. The authors propose an advanced
utilizing fewer computational resources. To segmentation technique aimed at improving the
optimize the model’s parameters, the authors accuracy of lung tumor identification from CT scan
propose an enhanced spotted hyena optimizer images. The modified U-Net model incorporates a
algorithm. Finally, disease classification is specialized lobe segmentation process, allowing
performed using a stacked long short-term memory for precise localization and analysis of nodules
(SLSTM) method, which handles sequential data within lung regions. This approach addresses
effectively. Experimental results, tested on the challenges such as variations in nodule size, shape,
Kvasir dataset, demonstrate the system's superior and location, which often hinder accurate
performance with an accuracy of 98.96%, diagnosis. The proposed model is designed to
surpassing other contemporary models. The study improve the accuracy of both benign and malignant
highlights the potential of the EIAGTD-NIADL tumor classification by refining the feature
system in advancing medical image analysis, extraction and segmentation process. The study
reducing diagnostic errors, and enabling precise explores how the modified U-Net model enhances
disease identification in the GI tract. [9] the detection of small and irregularly shaped
nodules, which are often overlooked by traditional
Biomedical Image Analysis for Colon and Lung diagnostic methods. The authors emphasize the
Cancer Detection Using Tuna Swarm potential of this model to transform clinical
Algorithm With Deep Learning Model practices by enabling early diagnosis and timely
intervention, which are critical for improving lung
In this journal, Munya A. Arasi, Marwa Obayya, cancer survival rates. Furthermore, the automated
Abdullah Mohamed, Nuha Alruwais, Raed Alsini, nature of the model reduces the burden on
and Ishfaq Yaseen propose an innovative model radiologists, allowing for faster and more
called BICLCD TSADL for the detection and consistent diagnoses. The study also discusses
classification of colon and lung cancer using future improvements, such as incorporating more
biomedical images. The study integrates advanced diverse datasets and refining the model’s ability to
techniques such as Gabor filtering for
preprocessing, GhostNet for feature extraction, and
handle complex cases involving overlapping or In this journal, Ruyi Zhang, Dingcheng Tian,
hidden nodules.[11] Dechao Xu, Wei Qian, and Yudong Yao provide a
comprehensive survey of the applications of deep
Identification of Lung Tumors in Nude Mice learning in wound image analysis, focusing on
Based on the LIBS With Histogram of tasks such as classification, detection, and
Orientation Gradients and Support Vector segmentation. The paper reviews publicly available
Machine datasets used in this field and highlights the
significance of preprocessing methods to enhance
In this journal, Qian-Lin Lian, Xiang-You Li, Bing model performance. It also investigates various
Lu, Chen-Wei Zhu, Jiang Tao Li, and Jian-Jun deep learning models, including convolutional
Chen introduced an innovative approach that neural networks (CNNs), and their applications to
leverages LIBS for spectral analysis of lung tissue different wound types like burns, diabetic foot
samples, extracting detailed chemical composition ulcers, and pressure ulcers. The challenges in
data. LIBS uses a high-energy laser pulse to create wound image analysis, such as data scarcity, noise,
a plasma from lung tissue samples, which emits and color variation, are discussed. The study
light that reflects the chemical makeup of the emphasizes the transformative role of deep
tissue. The emitted light is captured and analyzed learning in improving the accuracy, efficiency, and
to generate a spectral signature, which serves as the accessibility of wound diagnosis and treatment,
input data for the classification model. The HOG providing insights into future research and
technique is applied to analyze texture and development in this vital area of medical
structural features from the spectral data, enabling imaging.[13]
the identification of complex patterns and tumor-
specific markers. The extracted features are then Hybrid Metaheuristics With Deep Learning-
classified using an SVM model, which separates Based Fusion Model for Biomedical Image
malignant and benign tissue samples based on the Analysis
underlying spectral patterns. This approach
addresses key challenges in medical image In this journal, Marwa Obayya, Saud S. Alotaibi,
analysis, such as variations in tissue structure and Muhammad Kashif Saeed, Nuha Alruwais,
spectral noise, by combining physical and Mohammed Assiri, and Ahmed S. Salama propose
statistical analysis techniques. The authors validate the Hybrid Metaheuristics with Deep Learning-
the model’s performance through rigorous Based Fusion Model for Biomedical Image
experimentation, demonstrating its ability to Analysis (HMDL-MFMBIA). The study
achieve high sensitivity and specificity in tumor introduces a novel approach that combines Swin-
classification. The study highlights that the LIBS- UNet for image segmentation, Xception and
HOG-SVM approach outperforms traditional ResNet for feature extraction, and the hybrid salp
biopsy-based methods in terms of speed and swarm algorithm (HSSA) for optimal
accuracy, offering a non-invasive and reliable hyperparameter tuning. The model also integrates
diagnostic solution. Additionally, the model’s the gated recurrent unit (GRU) for classification.
ability to operate on small tissue samples enhances By addressing challenges such as improved
its clinical applicability, particularly for early stage accuracy and efficiency, the HMDL-MFMBIA
lung cancer detection. The study concludes that technique showcases its effectiveness through
combining LIBS with machine learning techniques experimental validation on datasets like ISIC 2017
offers a significant advancement in lung cancer and ISIC 2020, demonstrating superior
diagnosis, providing a fast, accurate, and non- performance compared to existing methods. The
invasive alternative to traditional diagnostic authors highlight the potential of this hybrid model
methods.[12] to significantly advance medical image
classification and enhance diagnostic precision in
A Survey of Wound Image Analysis Using Deep healthcare. The study highlights the significance of
Learning: Classification, Detection, and deep learning methods in biomedical imaging,
Segmentation addressing challenges like noise, artifacts, and
computational complexity. Experimental
validations were conducted on ISIC 2017 and ISIC to revolutionize medical diagnostics by improving
2020 datasets, demonstrating superior accuracy, efficiency, and patient outcomes. It
classification accuracy, precision, and recall underscores how these models have evolved with
compared to existing techniques. The authors the advent of big data and GPU advancements,
emphasize the transformative potential of allowing the discovery of hierarchical relationships
combining metaheuristic optimization with deep in medical images without manual feature crafting.
learning to enhance medical diagnostics, offering a The study explores the application of CNNs and
reliable and efficient solution for automated other deep learning models like recurrent neural
disease detection and classification in healthcare networks, autoencoders, and generative adversarial
systems.[14] networks across various medical imaging
modalities, including CT, MRI, ultrasound, and
Deep Learning Applications in Medical Image histology slides. [15]
Analysis
Unified deep learning models for enhanced lung
This journal, authored by Justin Ker, Lipo Wang, cancer prediction with ResNet 50–101 and
Jai Rao, and Tchoyoson Lim, provides a EfficientNet-B3 using DICOM images
comprehensive review of deep learning
applications in medical image analysis. It In this journal, the authors Vinod Kumar, Chander
emphasizes the transformative role of Prabha, Preeti Sharma, Nitin Mittal, S. S. Askar,
convolutional neural networks (CNNs) in tasks and Mohamed Abouhawwash explore the
such as classification, localization, detection, application of advanced deep learning models for
segmentation, and registration of medical images. lung cancer prediction. They utilized ResNet-50,
The study highlights the advantages of deep ResNet-101, and EfficientNet-B3, along with
learning in handling medical big data, enabling the transfer learning, to classify lung cancer images
discovery of hierarchical relationships without into four categories: Adenocarcinoma, Large Cell
manual feature crafting. It discusses various Carcinoma, Squamous Cell Carcinoma, and
imaging modalities, including CT, MRI, X-rays, Normal CT Images.The study addresses challenges
and histology slides, and explores the integration of such as high mortality rates due to late detection
supervised and unsupervised learning models like and the difficulty in distinguishing between similar
autoencoders, restricted Boltzmann machines, and cancer subtypes. Using a dataset of 1,000 DICOM
generative adversarial networks. The authors also images from the LIDC-IDRI repository, the
address challenges such as data scarcity, authors achieved significant results. The Fusion
computational limitations, and the need for large Model demonstrated 100% precision in classifying
labeled datasets, while proposing solutions like Squamous Cells, while ResNet-50, EfficientNet-
transfer learning and data augmentation. This B3, and ResNet-101 achieved 90% precision, with
review underscores the potential of deep learning slight variations in performance. [16]

Table 1: Comparison among Different Existing Systems

S. No Title Algorithm Challenges Data Sets Results

1 Machine learning and deep ML, DL Complexity MRI, 4D CNN achieved the highest
learning approach for of Medical DCE MRI
accuracy of 97.6%
medical image analysis: Imaging images
diagnosis to detection Data RF achieved an accuracy of
96.93%
2 Unified deep learning Resnet 50, Overfitting, LIDC- Accuracy of 90% with
models for enhanced lung resnet 101, late IDRI resnet 50
cancer prediction with EfficientNet detection
ResNet-50–101 and B3, Fusion
EfficientNet-B3 using Model
DICOM images
3 A review on deep learning Convolutiona High Lung CNN achieved the highest
in medical image analysis l Neural Computatio cancer accuracy of 97.6%,
Networks nal dataset, Random Forest (RF)
(CNNs), Requireme Heart-C achieved an accuracy of
Recurrent nts dataset 96.93%
Neural
Networks
(RNNs)
4 Medical Image Analysis CNN, RNN, Need of ChestX- CNN achieved the highest
Based on Deep Learning GANs large ray14,LU accuracy of 97.6%,
Approach labeled NA16 Random Forest (RF)
datasets achieved an accuracy of
96.93%

5 Deep learning-based image U-Net-based Data Fluoresce DeepLoc enhanced

processing in optical CNN, Deep variability nt automated categorization of
microscopy Loc and microscop protein subcellular
generalizab y images, localization in yeast cell
ility Holograp images, achieving 84%
hic precision
images of
living
cells

6 Deep Learning and the Long Short- Interpretabi EEG, CNN models showed an
Future of Biomedical Term lity of ECG, accuracy of up to 97.6%,
Image Analysis Memory model MEG Random Forest (RF)
Networks predictions (Magneto achieved an accuracy of
(LSTMs), encephalo 96.93%.
Encoder- gy)
Decoder
(ED)
architectures

Internet of Things and Median Data Benchmar

Synergic Deep Learning privacy and k tongue
Filtering
Based Biomedical Tongue security images
Color Image Analysis for (MF) for concerns dataset
Disease Diagnosis and consisting
image pre-
Classification (2021) of 936
processing, images
under 12 Precision: 98.4%, Recall:
Synergic
class
7 97.3%, Accuracy: 98.3%
Deep labels
Learning
(SDL) for
feature
extraction

8 A Review on Alzheimer’s Deep Neural Interpretabi OASIS, Dice Similarity Coefficient

Disease Through Analysis Networks lity of ADNI (DSC): Up to 98%, Jaccard
of MRI Images Using Deep (DNN), model Index (JI): 95%, Precision:
Learning Techniques Enhanced predictions
(2023) Black
 Widow 98.4%, Recall: 97.3%,
Optimization Accuracy: 98.3%
(EBWO)
9 Deep Learning in Medical Restricted Complexity CETUS CNNs showed improved
Ultrasound Image Analysis: Boltzmann of medical Challenge accuracy of 98% in
A Review (2021) Machines imaging dataset, classification tasks.
(RBMs), data Thyroid
Transfer dataset,
Learning Breast
(TL) ultrasoun
d image
dataset

10 Endoscopic Image Analysis Bilateral High Kvasir Precision: 95.85%, Recall:

for Gastrointestinal Tract Filtering computatio Dataset 95.85%, Accuracy: 98.96%
Disease Diagnosis Using (BF), nal
Nature Inspired Algorithm Improved requiremen
With Deep Learning Spotted ts.
Approach (2023) Hyena
Optimizer
(ISHO)
Biomedical Image Analysis Gabor Training LC25000
for Colon and Lung Cancer Filtering complexity, Dataset Accuracy: 99.33%,
Detection Using Tuna (GF), Need for
Precision: 98.31%, Recall:
Swarm Algorithm With Adaptive large
Deep Learning Model Firefly annotated 98.31%, F1 Score: 98.31%,
11 (2023) Algorithm datasets
AUC (Area Under Curve):
Optimization
(AFAO), 98.95%
Tuna Swarm
Algorithm
(TSA)

3. PROPOSED METHODOLOGIES meaningful predictions that can assist in clinical

decision-making. The stages are described below:
3.1 VGG16
1. Lung Dataset:
This project leverages the power of the VGG16
architecture, a well-established deep convolutional The foundation of this system lies in a carefully
neural network (CNN), to detect and classify lung curated lung CT image dataset, which contains
tumors from CT scan images. VGG16 is both healthy and tumor-affected lung scans. The
particularly effective in image classification tasks tumor cases are further divided into three distinct
due to its consistent use of small 3×3 convolutional categories representing various types of lung
filters and deep layered structure, which allows for cancer. These include Large cell carcinoma,
hierarchical feature extraction. Originally designed Squamous cell carcinoma, Adenocarcinoma.
for RGB images, the VGG16 model has been 2. Image Preprocessing:
adapted in this project to handle grayscale CT scans To ensure effective and uniform input for the neural
by resizing them to 224 × 224 × 3, ensuring network, each image undergoes a series of
compatibility with the pretrained architecture while preprocessing steps:
maintaining the visual structure of the input.
• Image Resizing: Adjusts images from their
The proposed system for lung tumor classification original size to uniform (224 x 224) size for
involves several sequential stages as outlined in the consistent analysis.
block diagram (Fig 1). Each stage is crucial in • Image Normalization: Enhances image
transforming raw medical imaging data into quality by scaling pixel values to a standard
 range(0 to 1), improving model the end of the convolutional stages, VGG16
performance. effectively encodes critical features that represent
3. Feature Extraction: the unique characteristics of various lung disease
types, which are then passed to fully connected
In the VGG16 architecture, feature extraction from layers for classification.
diseased lung images occurs through a deep
hierarchical structure of convolutional layers. The 4. Classification:
initial layers capture low-level features such as The extracted features are fed into VGG16 , a deep
edges, textures, and boundaries within the lung CT learning model for classification. The model
image. As the image passes through deeper layers, categorizes images into four classes:
the model identifies more complex patterns like
abnormal growths, nodules, or density variations • Normal (Healthy Lungs)
associated with tumors. Each convolutional block • Adenocarcinoma
extracts increasingly abstract and detailed • Large Cell Carcinoma
information, allowing the model to differentiate • Squamous Cell Carcinoma
between normal tissue and cancerous regions. By

Fig 1: Disease Detection Techniques using Biomedical Image Analysis

4 DATASETS
Table 2: Datasets used in biomedical Image Analysis.

S.NO Dataset Name No of Images Type

1. Oasis 87k MRI

2. Brats (Brain Tumor Segmentation) 57k MRI

3. Montgomery 139 X-Ray

4. Brain Tumor MRI Dataset 7k MRI

5. ADNI 18k MRI

6. JSRT(Chest X-Rays) 238 X-Ray

7. Covid-19 100 CSV

8. ISIC 2017 2000 Dermoscopy

9. ISIC 2020 33K Dermoscopy

10. GLAS 100 CSV

11 Chest CT scan Images Dataset 1000 CT scan

The dataset summary in above table 2 includes ray), and the Chest CT Scan dataset. Skin cancer
various medical imaging datasets used for disease diagnosis benefits from dermoscopy images in
detection. OASIS and ADNI provide large sets of ISIC 2017 and ISIC 2020. Additionally, GLAS
MRI brain images for neurological research, while supports colorectal cancer segmentation, and the
BRATS and the Brain Tumor MRI Dataset focus Covid-19 dataset offers basic CSV data for
on brain tumor analysis. Chest-related conditions infection analysis.
are targeted by Montgomery (X-ray), JSRT (X-
5 RESULTS AND DISCUSSION
(2 ∗ TP)
Mean Dice = (2 ∗ TP + FP + FN)
This section evaluates the performance of various
tumor classification models using standard metrics.
4.6 Specificity:
And also the web interface used to output the
predicted disease. Specificity is a metric that measures the ability of a
model to correctly identify negative outcomes. It
4.1 Accuracy:
represents the proportion of correctly predicted
Accuracy is a metric that measures how often a non-tumor cases out of all actual non-tumor cases,
model correctly predicts the outcome of a calculated as:
classification task. It represents the proportion of
correctly classified cases (both tumor and non- 𝑇𝑁
tumor) out of all cases, calculated as: Specificity = (TN + FP)

(𝑇𝑃 + 𝑇𝑁) Through keen observation of the below images we

Accuracy = (TP + TN + FP + FN)
know that we are getting a val accuracy of 98% and
4.2 Precision: val loss of 2%.

Precision is a metric that measures how accurately 4.7 Accuracy and Loss Plots
a model predicts positive outcomes. It represents
the proportion of correctly predicted tumor cases
out of all cases predicted as tumors, calculated as:
𝑇𝑃
Precision = (TP + FP)

4.3 Recall:
Recall is a metric that measures how well a model
identifies positive outcomes. It represents the
proportion of correctly predicted tumor cases out of
all actual tumor cases, calculated as:

𝑇𝑃
Recall = (TP + FN) Fig 2: Train and val accuracy

4.4 F1-Score: Train accuracy: It measures how well the model

performs on training data, while val accuracy
F1-Score is a metric that combines precision and shows performance on unseen validation data.
recall to provide a single measure of a model's Train loss: It indicates the model’s error on
performance. It represents the harmonic mean of training data, and val loss measures error on
precision and recall, calculated as: validation data. A large gap between train and val
(Precision ∗ Recall) accuracy suggests overfitting. Ideally, both
F1-Score = 𝟐 × (Precision + Recall)
accuracy should be high, and both losses should
decrease steadily.
• Train and val accuracy should be close.
4.5 Mean dice or Dice Coefficient:
• Train and val loss should both decrease
Mean Dice or Dice Coefficient is a metric used to steadily.
evaluate the similarity between two sets, • If val loss increases while train loss
commonly used in image segmentation tasks. It decreases, the model is overfitting.
represents the overlap between predicted and actual
tumor regions, calculated as:
 component. Once the image is uploaded, it is
displayed on the screen, giving immediate visual
feedback that the upload was successful.

(b):Fig 3: Train and val loss

4.8 Confusion Matrix
The confusion matrix provides a detailed
evaluation of the model's ability to classify lung
disease types, including adeno, large cell, normal,
and squamous. The majority of predictions are
correct, with 32 adeno cases, 21 large cell cases, 21
normal cases, and 33 squamous cases accurately Fig 5: Streamlit Input Interface
classified. Specifically, 2 adeno cases were
misclassified as large cell, while 1 large cell case
was incorrectly predicted as adenocarcinoma.

Fig 6: Streamlit Output Interface

Fig 4: Confusion Matrix From Figure 6 we can know that once the image is
shown, the system automatically resizes it to
224x224 pixels, normalizes it, and feeds it into the
4.9 Streamlit Input and Output Interface VGG16 model for prediction. The model processes
the input and outputs the probabilities for each class
The Streamlit interface for this lung tumor (Adenocarcinoma, Large Cell Carcinoma, Normal,
detection project provides a clean and intuitive user and Squamous Cell Carcinoma). This streamlined
experience. As shown in the Figure 5, at the top, it workflow allows users medical professionals or
displays the project title with a lung themed image. researchers to upload a scan and receive an efficient
Users are prompted to upload a CT scan image (in output.
JPG, JPEG, or PNG format) using the file uploader
 Table 3 : Performance of VGG16 vs. Existing Models

Model Accuracy(%) Precision(%) Sensitivity(%) Specificity(%)

Proposed VGG16
97 94 94 94
Model

Res-Net 50 95 90 - -

Res-Net 101 94 89 - -

From the Table 3 we can observe that the proposed accuracy of 94% and precision of 89%. The F1
VGG16 model demonstrates superior performance score and recall values were not explicitly
with an accuracy of 97%, precision of 94%, F1 mentioned in the base paper, which lacks a detailed
score of 94% and recall at 94%. This indicates that comparison in terms of balanced classification.
VGG16 effectively balances sensitivity and Although ResNet models leverage deep residual
specificity, making it highly reliable for lung learning to overcome vanishing gradient issues,
disease classification. The high precision suggests their lower accuracy in this context suggests that
that the model minimizes false positives, while the increasing network depth does not necessarily
recall value confirms its ability to correctly identify translate to better classification performance. The
actual disease cases. This performance can be marginal difference between ResNet50 and
attributed to VGG16’s deep convolutional layers, ResNet101 highlights that adding more layers may
which efficiently extract hierarchical spatial introduce complexity without significant
features from medical images. The structured performance gains. While ResNet models are
architecture of VGG16 ensures effective feature known for their robust feature extraction, VGG16
learning, enabling accurate differentiation between appears to be more effective in this specific lung
different lung disease categories. ResNet50 disease detection task due to its structured and
achieves an accuracy of 94% with a precision of efficient feature learning process.
90%, while ResNet101 attains a slightly lower

Table 4 : Performance of different types of diseases

Type of Disease Precision Recall F1-Score Support
Adeno 0.9487 0.9487 0.9487 39
Larg cell 0.9167 0.9565 0.9362 23
Normal 0.9310 0.9310 0.9310 29
Squamous 0.9667 0.9355 0.9508 31

The performance metrics in the above Table 4 demonstrate that the VGG16 model is robust, well-
highlight VGG16's strong ability to classify four generalized, and capable of learning important
lung tumor types accurately, with patterns in lung CT scan images. In contrast,
Adenocarcinoma, Normal, and Squamous Cell ResNet50 and ResNet101 performed slightly
Carcinoma all showing precision and F1-scores lower, likely due to their deeper and more complex
above 0.93. Large Cell Carcinoma also shows high architectures which tend to overfit on smaller
recall at 0.95, indicating effective identification of medical datasets without sufficient regularization.
true positive cases. These consistent results Moreover, as these models are pretrained on RGB
images from natural scenes, they may not fully contributing to its higher accuracy and reliability in
adapt to the grayscale CT scan domain without tumor detection. Overall, the results affirm that
extensive fine-tuning. VGG16’s sequential VGG16 offers a more optimal balance of
structure and smaller filters make it more effective performance and simplicity for medical image
in extracting subtle features from lung tissue, classification tasks like lung cancer detection.

5 CONCLUSION
Based on the state-of-the-art methods,
development of a Streamlit-based web application
[4] “Sindhoora Kaniyala Melanthota, Dharshini
using the VGG16 model for lung cancer detection Gopal, Shweta Chakrabarti, Anirudh Ameya
represents a significant advancement in medical Kashyap, Raghu Radhakrishnan, and Nirmal
diagnostics. The user-friendly steamlit interface Mazumder “, “Deep Learning-Based Image
allows healthcare professionals and researchers to Processing in Optical Microscopy”, Biophysical
easily upload lung CT scan images for analysis, Reviews, vol. 14, pp. 463–481, 2022.
facilitating early and accurate detection of different
types of lung cancer, including squamous cell
carcinoma, large cell carcinoma, and [5] “Monika Jyotiyana and Nishtha Kesswani,”
adenocarcinoma. Beyond classification, the “Deep Learning and the Future of Biomedical
application provides detailed information about the Image Analysis”, Deep Learning Techniques for
identified condition, including symptoms, possible Biomedical and Health Informatics, Studies in Big
causes, and recommended treatment strategies. By Data, vol. 68, pp. 329–342, 2020.
offering actionable insights, the system empowers
medical professionals to make informed decisions,
potentially improving patient outcomes and [6] “Romany F. Mansour, Maham Althobaiti, and
Amal Adnan Ashour”, “Internet of Things and
enabling timely interventions. The use of the
Synergic Deep Learning Based Biomedical Tongue
VGG16 model enhances the accuracy and
Color Image Analysis for Disease Diagnosis and
reliability of the classification process, ensuring Classification”, IEEE Access, vol. 9, pp. 94769-
consistent and precise results. This technology not 94777, 2021.
only supports early diagnosis but also serves as an
educational tool, increasing awareness about lung
cancer types and their management among medical [7] "Battula Srinivasa Rao and Mudiyala Aparna,
practitioners and researchers. 'A Review on Alzheimer’s Disease Through
Analysis of MRI Images Using Deep Learning
6.REFERENCES Techniques,' IEEE Access, vol. 11, pp. 71542-
[1] “Meghavi Rana and Megha Bhushan," Machine 71556, 2023.
Learning and Deep Learning Approach for Medical [8] Xinke Ge, He Ma, Shouliang Qi, Yu Wang, and
Image Analysis: Diagnosis to Detection, Yudong Yao, "Deep Learning in Medical
Multimedia Tools and Applications, vol. 82, pp. Ultrasound Image Analysis: A Review," IEEE
26731–26769, 2023. Access, vol. 9, pp. 54310-54325, 2021.

[2] “S. Suganyadevi, V. Seethalakshmi, and K. [9] Abdulrahman Alruban, Abdullah Alharbi,
Balasamy”, 'A review on deep learning in medical Eatedal Alabdulkreem, Majdy M. Eltahir, Imène
image analysis,' International Journal of Issaoui, and Ahmed Sayed, "Endoscopic Image
Multimedia Information Retrieval, vol. 11, no. 1, Analysis for Gastrointestinal Tract Disease
pp. 19–38, 2022. Diagnosis Using Nature-Inspired Algorithm With
Deep Learning Approach," IEEE Access, vol. 11,
pp. 130022-130029, 2023.
[3] “Muralikrishna Puttagunta and S. Ravi”,
“Medical Image Analysis Based on Deep Learning
Approach”, Multimedia Tools and Applications, [10] Munya A. Arasi, Marwa Obayya, Abdullah
vol. 80, pp. 24365–24398, 2021. Mohamed, Nuha Alruwais, Raed Alsini, and Ishfaq
Yaseen, "Biomedical Image Analysis for Colon Classification, Detection, and Segmentation,"
and Lung Cancer Detection Using Tuna Swarm IEEE Access, vol. 10, pp. 79502-79509, 2022.
Algorithm With Deep Learning Model," IEEE
Access, vol. 11, pp. 94705-94712, 2023.
[14] Marwa Obayya, Saud S. Alotaibi, Muhammad
Kashif Saeed, Nuha Alruwais, Mohammed Assiri,
[11] Iftikhar Naseer, Arfan Jaffar, Sheeraz Akram, and Ahmed S. Salama, "Hybrid Metaheuristics
Tehreem Masood, and Muhammad Rashid, "Lung With Deep Learning-Based Fusion Model for
Cancer Classification Using Modified U-Net Based Biomedical Image Analysis," IEEE Access, vol.
Lobe Segmentation and Nodule Detection," IEEE 11, pp. 117149-117155, 2023.
Access, vol. 11, pp. 60279-60287, 2023.

[15] Justin Ker, Lipo Wang, Jai Rao, and

[12] Qian-Lin Lian, Xiang-You Li, Bing Lu, Chen- Tchoyoson Lim, "Deep Learning Applications in
Wei Zhu, Jiang-Tao Li, and Jian-Jun Chen, Medical Image Analysis," IEEE Access, vol. 6, pp.
"Identification of Lung Tumors in Nude Mice 9375-9440, 2018.
Based on the LIBS With Histogram of Orientation
Gradients and Support Vector Machine," IEEE
Access, vol. 11, pp. 141915-141921, 2023. [16] Vinod Kumar, Chander Prabha, Preeti
Sharma, Nitin Mittal, S. S. Askar et.al, “Unified
deep learning models for enhanced lung cancer
[13] Ruyi Zhang, Dingcheng Tian, Dechao Xu, prediction with ResNet-50–101 and EfficientNet-
Wei Qian, and Yudong Yao, "A Survey of Wound B3 using DICOM images”, vol. 24, no. 1, pp. 1–21,
Image Analysis Using Deep Learning: 2024.