AN EFFECTIVE LUNG TUMOR DETECTION SYSTEM BASED

ON DEEP LEARNING TECHNIQUES

Submitted in partial fulfilment of the requirements for the award of

Bachelor of Engineering degree in Electronics and Communication Engineering

by

G JAYASURYA (41130194)
JIBU ALEX KOSHY (41130199)

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

SCHOOL OF ELECTRICAL AND ELECTRONICS

SATHYABAMA
INSTITUTE OF SCIENCE AND TECHNOLOGY
(DEEMED TO BE UNIVERSITY)
Accredited with Grade “A++” by NAAC

JEPPIAAR NAGAR, RAJIV GANDHI SALAI, CHENNAI – 600119

APRIL 2025
DEPARTMENT OF ELECTRONICS AND COMMUNICAITON ENGINEERING

BONAFIDE CERTIFICATE

This is to certify that this Project Report is the bonafide work of

G JAYASURYA (41130194) and JIBU ALEX KOSHY (41130201) who
carried out the project entitled “AN EFFECTIVE LUNG TUMOR DETECTION
SYSTEM BASED ON DEEP LEARNING TECHNIQUES” under my
supervision from November 2024 to April 2025.

Internal Guide

Dr. [Link] M.E., Ph.D.,

Head of the Department

Dr. T. RAVI, M.E., Ph.D.,

Submitted for Viva voce Examination held on

Internal Examiner External Examiner

ii
 DECLARATION

We G. JAYASURYA (41130194) and JIBU ALEX KOSHY (41130201) hereby

declare that the Project Report entitled “AN EFFECTIVE LUNG TUMOR
DETECTION SYSTEM BASED ON DEEP LEARNING TECHNIQUES” done by us
under the guidance of Dr. K. SRILATHA M.E., Ph.D., Associate Professor,
Department of Electronics and Communication Engineering is submitted in
partial fulfillment of the requirements for the award of Bachelor of Engineering degree
in Electronics and Communication Engineering.

DATE:

PLACE: CHENNAI SIGNATURE OF THE CANDIDATE(S)

1.

2.

iii
 ACKNOWLEDGEMENT

We are pleased to acknowledge our sincere thanks to The Board of

Management of Sathyabama Institute of Science and Technology for their
kind encouragement in doing this project and for completing it successfully. We
are grateful to them.

We convey our thanks to Dr. N. M. NANDHITHA, M.E., Ph.D., Professor

& Dean, School of Electrical and Electronics and Dr. T. RAVI, M.E., Ph.D.,
Professor & Head, Department of Electronics and Communication
Engineering for providing us necessary support during the progressive reviews.

We would like to express our sincere and deep sense of gratitude

to our internal guide Dr. K. SRILATHA , M.E., Ph.D., Associate Professor,
Department of Electronics and Communication Engineering for her
valuable guidance, suggestions and constant encouragement paved way for
the successful completion of our project work.

We wish to express our thanks to all Teaching and Non-teaching staff

members of the Department of Electronics and Communication Engineering
who were helpful in many ways for the completion of the project.

iv
 ABSTRACT

Lung tumor remains a leading cause of cancer-related deaths worldwide.

Early detection is crucial for improving patient outcomes. This project proposes
a novel lung tumor detection system utilizing deep learning techniques,
specifically designed to analyze medical images, such as Computed
Tomography (CT) scans and X-rays. By harnessing the power of deep learning,
the system aims to accurately detect lung nodules and potential cancerous
growths with high precision. Computed tomography plays a pivotal role in
assessing tumor conditions by evaluating factors such as tumor size, location,
and morphology. The system incorporates a comprehensive evaluation
process, including image preprocessing, segmentation, and feature extraction.
During preprocessing, images are standardized and noise is minimized.

In the segmentation phase, lung regions and potential tumors are isolated
from the background, enhancing the focus on relevant features. Feature
extraction involves identifying key characteristics of the tumor, such as shape
and texture, which are crucial for accurate diagnosis. These features are then
used to train various machine learning classification algorithms, including
Support Vector Machines (SVM), Random Forest, and Artificial Neural Networks
(ANN). Each algorithm's performance is evaluated based on accuracy,
sensitivity, specificity, and other metrics to determine the presence of
malignancy or benignity in tumors. The potential impact of this system extends
to improving lung cancer screening and early diagnosis. Its integration into
clinical settings could enhance the efficiency of cancer screening programs and
support timely interventions, ultimately leading to improved survival rates and
quality of life for patients.

v
 TABLES OF CONTENTS
CHAPTER NO. TITLE PAGE NO.

ABSTRACT v

LIST OF FIGURES viii

LIST OF TABLES ix

1 INTRODUCTION 1

2 LITERATURE SURVEY

2.1 Review on Literature Survey 4

2.2 Inferences and Challenges 7

3 EXISTING SYSTEM

3.1 Description of Existing systems 10

3.1.1 Chest X-Ray
3.1.2 Computed Tomography
3.1.3 Magnetic Resonance Imaging
3.1.4 Positron Emission Tomography
3.1.5 Sputum Cytology
3.1.6 Bronchoscopy with Biopsy
3.1.7 Fine-Needle Aspiration
3.1.8 Artificial Intelligence (AI) in Early
CAD Systems

3.2 Aim 16

3.2 Scope 18

vi
4 PROPOSED SYSTEM

4.1 Proposed Block Diagram 22

4.2 Description of Block Diagram 23

4.2.1 Data Acquisition & Preprocessing
4.2.2 2D to 3D Data Conversion
4.2.3 Feature Extraction & Selection
4.2.4 Model Training & Classification
4.2.5 Output Generation & Evaluation
4.3 Selected Methodologies 27
4.3.1 3D Convolutional Neural Network
4.3.2 Transfer Learning
4.3.3 Preprocessing &Data Augmentation
4.3.4 Model Training & Validation
4.3.5 Deployment and Integration

4.4 Software requirements 31

5 RESULT AND DISCUSSIONS 33

5.1 Constraints and Tradeoff 41

6 SUMMARY AND CONCLUSIONS

6.1 Summary 42
6.2 Conclusions 43
REFERENCES 45

vii
 LIST OF FIGURES

FIG NO. TITLE PAGE NO.

3.1 Chest X ray 11

3.2 CT scan 12

3.3 MRI Image of lung 12

3.4 PET Scan of lung 13

3.5 Sputum Cytology 14

3.6 Bronchoscopy with Biopsy 14

3.7 Fine-Needle Aspiration (FNA) Biopsy 15

3.8 AI in Early CAD Systems 16

4.1 Architecture Design 21

4.2 Architecture Diagram 26

5.1 Original Lung image 33

5.2 Output of proposed system 34

5.3 Tumor Highlighted lung image 35

5.4 Tumor lung converted to 3D 35

5.5 Tumor affected area 36

5.6 Ribcage of lung 37

5.7 Comparison Parameters 38

5.8 Comparison between 3D and 2D 39

viii
 LIST OF TABLES

TABLE NO. TITLE PAGE NO.

5.1 Performance Analysis of 3D-CNN and 38

2D-CNN

ix
 CHAPTER 1

INTRODUCTION

Lung cancer remains one of the most prevalent and fatal forms of cancer
worldwide, causing a significant percentage of cancer deaths every year. Early and
accurate detection of the disease is the key to improving survival rates since early
treatment profoundly increases the effectiveness of treatment methods like surgical
procedure, chemotherapy, and radiation therapy. However, the process of diagnosing
lung cancer is complex and requires careful analysis of medical imaging data,
particularly CT scans, which provide detailed pictures of the lungs. Traditional
diagnostic methods, like radiologists' visual examination of images, are prone to human
fallibility and inconsistency, thereby underlining the increasing demand for automatic
and reliable detection methods. New developments in deep learning algorithms, i.e.,
Convolutional Neural Networks (CNNs) and U-Net architectures, have proved effective
tools in the field of medical image analysis, offering improved accuracy in lung tumor
detection and classification.

Recent advances in artificial intelligence (AI) have revolutionized many

industries, and the healthcare industry is one of the most impacted sectors. The use of
machine learning and deep learning models in medical imaging has come with huge
potential in assisting radiologists in early detection of diseases, as well as decreasing
false positive and false negative rates. These computational methods not only enhance
the quality of diagnosis but also ease the workload for healthcare professionals, hence
allowing them to focus on complex cases. The use of CNN-based models in lung cancer
diagnosis has been researched widely, with studies comparing different architectures,
preprocessing methods, and classification approaches. Researchers have tried several
CNN models, including VGG19, ResNet50, and EfficientNetB7, in a bid to ensure the
reliability of tumor classification. Such architectures make it easy to extract complex
features from CT scans, which leads to better discrimination between malignant and
benign nodules.

1
 Even though deep learning of lung cancer detection has made significant
progress in automatic detection, some challenges exist. The performance of CNN-
based models is highly data-dependent on the diversity and quality of the training set.
Medical images are also subject to variations in noise levels, contrast, and resolution,
affecting the accuracy of automated systems of detection. Huge amounts of annotated
data are also needed for deep learning models to generalize across various patient
bases. It is, however, challenging to obtain high-quality annotated medical data
because of ethical issues, patient confidentiality legislation, and annotation expertise.
Due to the requirement of strong and generalizable models, researchers have turned
to hybrid methods that integrate deep learning with conventional image processing
methods, including morphological operators and genetic algorithms, to perform
segmentation and classification better.

Another critical step in lung tumor identification is segmentation, a method of

isolating tumors from other lung tissue in medical images. U-Net models, especially
those with dilated convolutions, have become increasingly popular for this purpose.
The models are trained to learn both local and global tumor features and thus improve
segmentation accuracy. However, challenges still exist in distinguishing tumors from
other pulmonary diseases, such as infections, fibrosis, and benign nodules. The
different geometries and sizes of tumors present an additional complexity to
segmentation that requires sophisticated models with flexibility to different tumor
attributes. Researchers have attempted multi-scale feature extraction and attention
mechanisms to enhance the performance of U-Nets; however, computational cost and
real-time processing are still areas of research.

In addition to segmentation, classification is also necessary in lung cancer

diagnosis to differentiate between benign and malignant tumors. CNN-based
classification models have been very precise, but their accuracy is sensitive to network
depth, parameter tuning, and data augmentation techniques. Transfer learning, using
pre-trained models, has been used to address data sparsity problems and enhance

2
classification accuracy. Fine-tuning pre-trained networks for medical use, however,
needs domain knowledge to guarantee that the features learned from general data are
applicable to lung cancer detection. Furthermore, interpretability is still a major
challenge, as deep learning models are "black boxes," and it is hard to comprehend
the reasoning behind their predictions. Researchers are developing explainable AI
methods to offer more transparent insights into CNN decision-making processes, which
will improve trust and uptake in clinical environments.

The rapid development of deep learning technology continues to advance lung

cancer detection capability with ongoing research focused on improving existing
frameworks and developing new methods. One of the major areas of research has
been the use of three-dimensional convolutional neural networks (3D CNNs) as a
possible method of utilizing spatial information in computed tomography (CT) scans to
improve tumor classification accuracy. Furthermore, hybrid models that integrate
machine learning-based classifiers and deep learning-based feature extraction
methods have been shown to improve diagnostic accuracy. While deep learning-based
approaches have improved medical image analysis in a variety of areas, effective
application in real-world clinical practice requires overcoming issues of data quality,
model explainability, and computational complexity. With ongoing research in this area,
collaborative effort between artificial intelligence researchers, clinicians, and medical
centers will be critical in bridging these technological advances into feasible solutions
for early lung cancer detection and treatment planning.

3
 CHAPTER 2

LITERATURE SURVEY

2.1 REVIEW ON LITERATURE SURVEY

D. Muruga Radha Devi et al., (2024) has pioneered convolutional neural networks
to find lung cancer using CT scans. The authors employed a pre-processed Kaggle
dataset and yielded more than 80% accuracy. The research emphasizes the strong
performance of ZF NET than other models on account of more pixel information
contained in smaller-sized filters and resulting in better performance in detection.

Mani and Viswanadhapalli, (2021) the article present a DCNN-based method for
the classification of lung tumors as malignant or benign using varying architectures
with varying patch sizes. The approach is more effective than traditional classifiers like
ANNs with high accuracy in early tumor detection and classification. The study shows
the potential of deep learning algorithms to enhance medical image analysis.

Shahruk Hossain et al. (2022) emphasize a DCNN-based approach for classifying

lung tumors as malignant vs. benign using different architectures with different patch
sizes. The method is more effective than conventional classifiers such as ANNs, with
high classification accuracy and tumor detection in the early stage. The research
illustrates the capability of deep learning methods to facilitate better medical image
analysis.

Young-Jin Park et al. (2023) suggests an approach for the identification of lung
cancer tumors via preprocessing of DICOM images, lung and soft tissue region
segmentation, and improvement of image clarity. The method achieved impressive
performance metrics with 97.96% sensitivity, 99.23% precision, and F1-score of
98.56%.The results indicate that the method significantly enhances the

4
neural network learning process for lung cancer identification. The models,
nevertheless, can be tested by instances of high fluctuation or rapid changes in the
levels of accuracy, which requires models that are capable of coping with dynamic
change.

Spyrou et al. (2021) compare CNN and LSTM models in forecasting CT scans at
a port area, concluding that LSTM performs better than CNN in handling temporal
relationships. The capability of LSTM to deal with non-linearities and sequential data
ensures it is well-suited for intricate time-series forecasting. Yet, the need for vast
amounts of training data and computational power remains a challenge. In addition,
LSTM models are likely harder to interpret than conventional statistical models such
as CNN, which can deter their real-world use.

Humaira Jahan Tuba et al. (2022) suggests a state-of-the-art approach for lung
tumor detection and segmentation based on U-Net architecture augmented with dilated
convolutions. The authors utilize 3D lung CT scans of the NSCLC- Radiomics Dataset
to develop a pipeline that enhances segmentation accuracy while decreasing
computational complexity. Using a weighted binary cross- entropy loss function instead
of the normal dice loss allows the model to obtain a 53% dice coefficient and overall
accuracy of 78%. Better segmentation with lesser computational resources in
comparison to LungNet makes the proposed model more of an enhanced innovation
in technologies for early detection of lung cancer.

Mst. Farhana Khatun et al. (2022) suggests a deep learning model intended to
improve lung cancer detection and classification from medical images. By combining
sophisticated neural network methods, the model produces high classification
accuracy, which renders it a sound tool for minimizing diagnostic mistakes. The study
highlights the significance of deep learning in medical diagnosis by showing that the
suggested model is capable of differentiating between benign and malignant tumors.

5
 P Jayapriya et al. (2023) investigate machine learning methods for detecting lung
tumors at an early stage using CT scans. The paper is centered on the importance of
preprocessing and using sophisticated machine learning algorithms to enhance the
accuracy of tumor detection. The suggested method has a high potential to detect
tumors at an early stage, which is crucial to enhance survival rates. The study
highlights the application of machine learning to improve the accuracy of medical
diagnosis, providing a useful insight into the development of early detection of cancer.

T. Senthil Kumar et al. (2022) investigates the use of Deep Convolutional Neural
Networks (DCNN) in classifying lung tumors based on differentiating benign from
malignant tumors. The authors have come up with three unique DCNN architectures,
with different patch sizes, to fine-tune the classification accuracy. The models were
validated on the LIDC database, wherein the DCNN showed better performance than
conventional classifiers such as Artificial Neural Networks (ANN). This study affirms
the efficacy of DCNN in producing accurate tumor classification that is critical in the
early and accurate detection of lung cancer.

M. Ariful Haque et al.(2020) method for automated lung tumor detection and
segmentation based on a hybrid 3D convolutional neural network. The authors present
a new architecture that blends 2D and 3D convolutions to better extract 3D structural
information in lung CT scans. The model was validated on the NSCLC-Radiomics
Dataset and obtained an average dice coefficient of 65.7%, outperforming other state-
of-the-art models such as LungNet. The research contributes considerably to the area
of medical imaging by providing a superior approach to tumor segmentation, which is
fundamental to precise diagnosis and treatment planning.

6
2.1 INFERENCES AND CHALLENGES IN EXISTING SYSTEM

2.2.1 Inferences

Recent studies have demonstrated significant advancements in lung tumor

detection and classification using deep learning techniques, particularly Convolutional
Neural Networks (CNNs) and their specialized variations such as Deep CNNs, U-Net,
and hybrid models. One of the most notable trends is the use of 3D CNNs for analyzing
volumetric CT scan data, which allows for better spatial feature extraction compared
to traditional 2D CNNs. Additionally, dilated convolutions have proven effective in
improving the segmentation accuracy of lung tumors by capturing fine details in CT
images. Several researchers have also explored the integration of knowledge-based
genetic algorithms (KBGA), SVM classifiers, and morphological operators, all of which
enhance tumor segmentation and classification. These AI- driven approaches have
achieved high accuracy rates, with models such as ResNet50 and ZF Net surpassing
90% accuracy, demonstrating their potential for early diagnosis.

The effectiveness of CNN-based methods has also been reinforced by the use of
transfer learning and image pre-processing techniques, which improve model
generalization, even with limited medical imaging datasets. Some studies have
proposed automated pipelines combining 2D and 3D feature extractions, which refine
segmentation precision and reduce false positives in lung cancer detection.
Additionally, hybrid approaches that combine CT and PET scan data have been
explored, leveraging the strengths of both imaging modalities for improved tumor
classification. The integration of deep learning models into web applications and real-
world clinical environments is also gaining traction, making AI-assisted diagnosis more
accessible for radiologists and medical professionals. The implementation of these
technologies in medical imaging could significantly enhance diagnostic accuracy,
reduce radiologist workload, and facilitate earlier interventions, ultimately improving
patient outcomes.

7
2.2.2 Challenges

Despite In spite of the outstanding progress being made in AI-based lung tumor
detection, numerous challenges still need to be overcome in order to make these
approaches more dependable and universally applicable. One critical challenge lies in
the computational cost of training deep learning models, particularly architectures
employing dilated convolutions, hybrid 3D networks, and ZF Net, which demand large
processing capacity and memory. Availability of large, varied, and well-labeled
datasets is also a big challenge since deep learning models are dependent on high-
quality labeled medical images for successful training. The accuracy of segmentation
also depends on image quality, with the model having a difficult time distinguishing
between tumors and overlapping lung structures, noise, and differences in imaging
conditions.

he other challenge is generalizability of these models across various datasets

and clinical settings. There are indications from some studies that performance
declines when models are evaluated on unseen data, calling for more resilient and
adaptive architectures. The selection of feature extraction methods, the choice of
kernels in SVM classifiers, and the combination of several imaging modalities make
the process even more complicated since different methods may provide different
sensitivities and specificities. Furthermore, interpretable AI remains a critical concern,
as many deep learning-based models function as black-box systems, making it difficult
for radiologists to understand how decisions are made. Ensuring that AI-assisted
diagnostic tools are clinically trustworthy and transparent is essential for their adoption
in real-world medical practice.

Moreover, optimization methods and pre-processing techniques also need to be

precisely tuned for uniform tumor detection in heterogeneous populations of patients.
Weighted loss functions and balancing methods have been identified by certain
researchers as crucial to counteracting dataset imbalances, especially when
discriminating between benign and malignant tumors.
8
 Lastly, the effective implementation of AI-driven lung tumor detection devices
depends on seamless integration with current medical protocols, so that the devices
support, not substitute for, the acumen of human radiologists.

9
 CHAPTER 3

EXISTING SYSTEM

3.1 DESCRIPTION OF EXISTING SYSTEMS

Previously, technologies used for detecting lung tumors have improved over
time with advances in medical diagnostics and imaging. Chest X-ray (CXR) was one of
the first techniques used, delivering a simple 2D representation of the lungs but with
poor detection of small or initial-stage tumors. For better accuracy, computed
tomography (CT) scans were developed, providing sophisticated cross-sectional views
of the lungs, which enabled more accurate detection of tumors. In a few instances,
magnetic resonance imaging (MRI) was utilized, particularly when differentiation of soft
tissues was required, although it was not as frequently utilized for the diagnosis of lung
cancer. Another important development was the positron emission tomography (PET)
scan, usually in conjunction with CT (PET-CT), which identified metabolically active
cancerous tissue and tumor spread.

In addition to imaging, sputum cytology was employed to scan lung mucus for
cancer cells, but it was only effective against tumors that released measurable cells.
More invasive methods of diagnosis comprised bronchoscopy with biopsy, in which a
flexible scope with a camera was passed down the airways to obtain tissue specimens,
and fine-needle aspiration (FNA) biopsy, which involved passing a needle into the lung
to obtain cells for analysis. There were also early computer-aided detection (CAD)
systems based on artificial intelligence (AI) that were used to aid radiologists in
detecting lung abnormalities, though with limited accuracy at first. These conventional
techniques have, over time, given way to advanced deep learning methods like 3D
convolutional neural networks (3D CNNs) that can detect 3D medical images with
higher accuracy, enhancing early diagnosis and detection of lung tumors.

10
3.1.1 Chest X-Ray – The Fig:3.1 demonstrates that the Chest X-ray (CXR) is among
the oldest and most popular imaging methods for the detection of lung tumors. It offers
a two-dimensional projection of the lungs, hence an inexpensive and rapid initial
screening tool. Nevertheless, its principal disadvantage is that it is low in sensitivity to
detect small tumors, particularly at the initial stages of lung cancer. Most lung
pathology, including benign nodules and infections, looks the same on an X-ray, so
malignancy cannot easily be ruled in or out based on this mode of imaging. In spite of
these limitations, X-ray persists in widespread usage in everyday practice because it is
cheap and universally available.

Fig:3.1: Chest X ray

3.1.2 Computed Tomography (CT) Scan – Computed Tomography (CT) scanning is

a significant advancement over X-ray imaging, providing detailed cross-sectional
images of the lungs. As shown in Fig:3.2, High-Resolution CT (HRCT) can detect small
nodules and is particularly effective in distinguishing between benign and malignant
tumors. The introduction of Low-Dose CT (LDCT) has enhanced lung cancer screening,
particularly for individuals at high risk like heavy smokers. Compared to X-rays, CT
scans provide more visibility of lung anatomy and are a diagnostic method of choice for
lung tumor identification. Repeated exposure to radiation, however, is a limiting factor,
preventing it from being used frequently for follow-up scans.

11
 Fig:3.2: CT scan
.
3.1.3 Magnetic Resonance Imaging (MRI) – Magnetic Resonance Imaging or MRI is
a non-invasive imaging method which employs strong magnetic fields and radio waves
to create high-quality images of soft tissue. MRI, although not used initially for detection
of lung cancer, is instrumental in evaluating spread of tumors into neighboring
structures like chest wall or mediastinum. As apparent from Fig:3.3, MRI, unlike CT
scans, does not involve ionizing radiation, thus it is a safer method for repeated
imaging. However, MRI suffers from limitations in lung imaging because of lower
concentration of hydrogen atoms in lung tissues, which impacts the clarity of the image.
Therefore, it is generally employed as an adjunct tool in situations where further tissue
characterization is desired.

Fig:3.3: MRI Image of lung

12
3.1.4 Positron Emission Tomography (PET) Scan – Positron Emission
Tomography (PET) is a nuclear medicine imaging modality that measures metabolic
activity in tissues. As demonstrated in Fig:3.4, it requires the injection of a radiotracer,
usually fluorodeoxyglucose (FDG), which concentrates in very active cancer cells,
highlighting them as "hot spots" on the image. PET scans are usually combined with
CT scans (PET-CT) for better localization and diagnostic accuracy. This method is
especially valuable for the discrimination of malignant from benign lung nodules and
the cancer spread evaluation (staging). PET scans have a very good sensitivity but at
times can raise false-positive interpretations in infection and inflammatory disorders.
These must then be further validated by other types of tests..

Fig:3.4: PET Scan of lung

3.1.5 Sputum Cytology – Sputum cytology is a non-invasive and straightforward

diagnostic procedure in which a patient gives a sputum (mucus) sample that is analyzed
under a microscope for cancer cells. As illustrated in Fig:3.5, it is most useful in the
detection of centrally located lung cancers that develop within the airways.
Nevertheless, it is less sensitive and may miss tumors in the outer parts of the lung
(peripheral tumors). It does not exclude lung cancer since there are tumors which do
not release detectable cells into the sputum. Although once popular for detecting early
lung cancer, this method is now considered less reliable compared to advanced
imaging techniques.

13
 Fig:3.5: Sputum Cytology

3.1.6 Bronchoscopy with Biopsy In the Fig:3.8 it is illustrated that, Bronchoscopy is

an invasive test in which a flexible tube (bronchoscope) with a camera is inserted via
the mouth or nose into the airways to inspect lung tissues. A biopsy can be done by
removing a small sample of tissue for lab testing if an abnormal growth is found. This
method is especially effective in diagnosing tumors in the central airways and offers a
direct visual evaluation of lung pathology. Bronchoscopy, however, is limited in
accessing peripheral lung tumors and necessitates other means like CT-guided
biopsies. In spite of its invasiveness, bronchoscopy is still an important tool for
diagnosing lung cancer.

Fig:3.6: Bronchoscopy with Biopsy

14
3.1.7 Fine-Needle Aspiration (FNA) Biopsy – Fine-Needle Aspiration (FNA) biopsy
is an intra thoracic procedure involved to take samples of lung tissue for microscopic
analysis. A thin needle is guided through the chest wall by imaging modalities like CT
or ultrasound to obtain a sample from the involved tumor area as illustrated in Fig:3.7.
It is a very good method of proving malignancy and is routinely employed when lung
nodule is identified by imaging. FNA biopsy is not without risks, though - pneumothorax
(pneumonia) or bleeding - and is thus a procedure to be carefully performed. It is still
one of the best diagnostic methods for tumors of the lung, even with such risks.

Fig:3.7: Fine-Needle Aspiration (FNA) Biopsy

3.1.8 Artificial Intelligence (AI) in Early CAD Systems – Early Computer-Aided

Detection (CAD) systems were developed to assist radiologists in identifying suspicious
lung nodules on X-ray and CT scans. As shown in Fig:3.8, such AI systems employed
conventional image processing methods and pattern recognition rules to mark
suspicious regions. Despite this, the early CAD systems suffered from flaws such as
excessive false-positive readings and poor generalization across disparate imaging
conditions. Though they assisted radiologists by serving as a second opinion, they were
15
not accurate enough for independent diagnosis. These initial CAD methods set the
stage for contemporary deep learning-based methods, including 3D Convolutional
Neural Networks (3D CNNs), which offer significantly improved accuracy in lung tumor
detection.

Fig :3.8: AI in Early CAD Systems

3.2 AIM

The Goal of these past and ongoing research studies in lung tumor detection is
to develop computational models that work on medical imaging data to identify and
classify lung tumors. Researchers have tried many methods to improve accuracy and
efficiency of tumor detection with the help of advanced imaging methods like CT scans
and X-rays. The research studies are mostly aimed at using machine learning and deep
learning models to enhance diagnostic accuracy and minimize human interpretation
dependency.

16
 Previous methods for detecting lung tumors used conventional image
processing methods, where radiologists visually examined tumor features like shape,
texture, and intensity changes. Even though these methods were rudimentary, they
were highly subject to human expertise and could not be applied to big data sizes. To
address this, machine learning models were proposed, wherein algorithms like Support
Vector Machines (SVM), k-Nearest Neighbors (k-NN), Decision Trees, and Random
Forests were used to classify lung tumors. These models utilized handcrafted features
of medical images, enhancing classification performance but with reduced adaptability
and automation.

Another method investigated in recent studies is the use of 3D Convolutional

Neural Networks (3D CNNs) to identify lung tumors. As medical image data such as
CT scans contain many slices with various cross-sections of the lungs, 3D CNNs were
used to handle volumetric data better than conventional 2D CNNs. The method enables
more precise analysis of tumor structures to enhance classification performance. The
use of 3D CNNs is, however, computationally expensive and demands huge well-
annotated training data, which is difficult in model development and deployment.

Segmentation-based methods have also been studied extensively to enhance

the precision of the detection of tumors. U-Net and Mask R-CNN networks have been
used to segment the tumor of the lung from background tissue, leading to more precise
diagnoses. The methods are effective in tumordelineating regions from noisy
backgrounds and, hence, enhancing classification results. Hybrid approaches
combining segmentation models with classification networks have been the interest of
some studies, enabling a more precise investigation of tumor characteristics.

In addition, researchers have also looked into using transfer learning in improving
17
tumor detection models. Using pre-trained CNN models pre-trained on large image
datasets, it has been demonstrated that it is possible to significantly improve the
detection performance of the models by fine-tuning the models using lung tumor
datasets.

This practice is used to reduce issues associated with small medical image
datasets while attaining stable classification performance. Even with these
advancements, current approaches are still burdened by the issues of data shortage,
variability of tumor appearance, and the requirement for more computational efficiency.
Current research continues to uncover emerging techniques, such as attention
mechanisms and combined deep neural network architectures, in order to further
enhance lung tumor detection and classification.

3.3 SCOPE

From a comprehensive survey of the literature, a wide range of studies have

been discovered in lung tumor detection and classification. Several investigations have
tested a range of different methodologies, including both conventional and deep
learning- oriented techniques, for enhanced accuracy in early detection. Previous work
includes both classification and segmentation of tumors, classifying between benign
and malignant tumors with growing accuracy. These investigations employ various
imaging modalities, including X-rays, CT scans, and PET scans, and place special
emphasis on the conversion of 2D images to 3D volumetric representations. The
processing of 2D radiographic data into 3D stacks has been duly mentioned since it
facilitates the more detailed analysis of tumor properties through deep learning models.

Much of the literature is focused on the creation of efficient data preprocessing

methods for medical imaging. Different preprocessing techniques, such as noise
reduction, contrast enhancement, normalization, and dimensionality reduction, have
been investigated to enhance dataset quality. Interpolation methods have been
extensively studied to transform 2D images into 3D stacks, which improve the ability of
convolutional neural networks (CNNs) to process spatial features more efficiently.
18
Experiments have also pointed out the use of various machine learning and deep
learning models, i.e., ResNet, AlexNet, VGGNet, and ZFNet, which have shown
significant increase in the accuracy of classification.

Research has also emphasized various model training methodologies and

optimization methods for improving classification performance. These have included
hyperparameter optimization, regularization methods, and various loss functions for
minimizing model loss and maximizing model accuracy.

The performance of various architectures has been extensively tested across

large-scale datasets, with emphasis on optimal performance metrics such as accuracy,
sensitivity, specificity, and precision. The aim of these studies has been to create robust
and efficient models that are capable of high accuracy while being computationally
feasible.

Some studies have also touched upon post-processing methods and

interpretability of results. Techniques like heatmaps and gradient-based class
activation mapping have been suggested to increase transparency in deep learning
models so that they become more interpretable for clinicians. Furthermore, several
studies have investigated the incorporation of lung tumor detection systems into clinical
practice, evaluating issues surrounding real-time processing, model scalability, and
creating user-friendly interfaces for clinicians. The emphasis has been on developing
systems that not only have high accuracy in detecting tumors but are also feasible for
real-world deployment.

There is an emerging focus in the literature on the clinical application of AI-based

models for lung tumor detection. Research has pointed to several challenges, such as
real-time data handling, scalability, and applicability in clinical environments.
Overcoming these challenges is important to allow the developments in deep learning
and medical imaging to convert to clinical practice successfully, enhancing early
diagnosis and treatment outcomes.

19
 The state of the art remains to develop, mirroring the development of ongoing
initiatives aimed at improving the accuracy and practical applicability of computerized
lung cancer detection systems.

The methods are especially useful in separating the tumor areas from noisy
backgrounds to enhance the classification results. Hybrid methods that integrate
segmentation models and classification networks have been the focus of some studies,
enabling a more precise exploration of tumor characteristics.

20
 CHAPTER 4

PROPOSED SYSTEM

Conventional medical testing tools like 2D X-ray and CT scans cannot reveal all
the elements of tumors because they only see in two dimensions. The suggested model
uses 3D Convolutional Neural Networks (3D CNNs) to find lung tumors from volumetric
medical imaging data. The approach conforms to IEEE P2801 standards which
enables trusted AI-based medical imaging systems.

Our system starts by processing both 2D X-ray images and 3D CT scan datasets
which serve as input data. Preprocessing begins with image enhancement and lung
area isolation steps before reducing the image dimensions with both Principal
Component Analysis and autoencoders. The 3D processed data helps train our unique
3D CNN architecture through labeled datasets. Model performance depends on how
we tune its essential parameters. After training the model it analyzes new lung scan
pictures to find and identify tumors. Our testing depends on key performance measures
such as accuracy, precision, recall and F1-score to determine reliability of diagnostic
output. The last stage includes showing and validating tumor finding results as well as
generating diagnosis reports to confirm the model predictions against doctor
assessments.

Using IEEE P2801-standard AI systems makes sure our method is transparent,

easy to duplicate and uniformly followed in tumor finding from lungs. By adding 3D
CNNs this detection method shows better results which strengthens CAD systems for
lung cancer assessment. Future research should add capabilities from multiple medical
data sources and develop AI systems that doctors can understand to make the model
more practical in medical work.

21
4.1 PROPOSED BLOCK DIAGRAM

Fig: 4.1: Proposed block Diagram

22
4.2 DESCRIPTION OF BLOCK DIAGRAM

The block diagram represents the step-by-step workflow for lung tumor detection
and classification using deep learning techniques. The entire process is structured into
distinct phases, each playing a critical role in ensuring accurate tumor identification.
These phases include Data Acquisition & Preprocessing, 2D to 3D Data Conversion,
Feature Extraction & Selection, Model Training & Classification, and Output Generation
& Evaluation. Each of these stages is further subdivided into key tasks that contribute
to the efficiency and reliability of the system.

4.2.1. Data Acquisition & Preprocessing

The first stage involves acquiring medical imaging data from publicly available
datasets such as LIDC-IDRI, NSCLC-Radiomics, and Kaggle repositories. These
datasets contain lung CT scans and X-ray images used for model training and
evaluation. The data acquisition process ensures that a diverse range of images,
including both benign and malignant cases, is collected for comprehensive analysis.

Once the images are acquired, they undergo preprocessing to enhance quality
and prepare them for deep learning models.
This includes:
Noise Reduction: Medical images often contain noise due to scanning artifacts.
Various filters, such as Gaussian and median filtering, are applied to remove noise and
enhance image clarity.
Contrast Enhancement: Since medical images can have poor contrast, techniques
such as histogram equalization and adaptive contrast stretching are applied to improve
visibility.
Normalization: Image intensity values are normalized to ensure uniformity across
different samples, preventing bias during model training.
Dimensionality Reduction: To improve computational efficiency, techniques like
principal component analysis (PCA) are used to remove redundant features.

23
This phase ensures that high-quality, standardized images are fed into the system for
accurate tumor detection and classification.

4.2.2. 2D to 3D Data Conversion

Since the proposed approach involves analyzing tumors in a 3D spatial context,

2D medical images are converted into 3D stacks before feature extraction. This step is
crucial because deep learning models, particularly 3D CNNs, require volumetric data
for precise tumor localization. The conversion process involves:

Interpolation Techniques: Missing slices in CT scan sequences are reconstructed

to create a continuous 3D representation of lung structures.
Voxelization: The 3D representation is divided into small volumetric units (voxels),
allowing the model to process spatial information effectively.
Normalization & Resizing: The converted 3D data is resized to a uniform shape to
maintain consistency across all samples.

By leveraging 3D data instead of purely 2D images, the system improves its

ability to differentiate between normal tissues and tumors, ensuring more reliable
classification.

4.2.3. Feature Extraction & Selection

Once the data is in a suitable format, essential tumor features are extracted for
analysis. Feature extraction is a key step in deep learning-based classification, as it
helps the model learn critical characteristics of lung tumors. This involves:

Shape & Texture Analysis: Tumor shape, edges, and texture variations are extracted
to differentiate between benign and malignant growths.
Intensity & Density Features: CT scans provide information about tissue density,
which helps in tumor classification. Hounsfield unit (HU) values are analyzed to
distinguish cancerous tissues from healthy lung structures.

24
Deep Feature Extraction via CNNs: The convolutional layers in deep neural networks
automatically extract hierarchical features, such as edges, contours, and patterns, that
are crucial for accurate tumor identification.

Additionally, feature selection techniques like mutual information and recursive

feature elimination (RFE) are applied to retain only the most relevant features, reducing
computation time and improving model performance.

4.2.4. Model Training & Classification

In this phase, the preprocessed and feature-extracted images are fed into a
Convolutional Neural Network (CNN)-based model for classification.

The training process involves Splitting Data into Training & Testing Sets. The dataset
is divided into training, validation, and testing subsets to ensure proper model
evaluation. Different CNN architectures, such as ResNet and VGGNet are considered
to determine the most effective model for classification.
Training Process: The model is trained using backpropagation and optimization
techniques, such as Adam and Stochastic Gradient Descent (SGD), to minimize
classification errors.
Loss Function Optimization: To improve classification accuracy, loss functions such
as cross-entropy loss and weighted binary cross-entropy are used.
During training, the model learns to recognize complex patterns in tumor images,
refining its ability to distinguish between malignant and benign cases. The trained
model is then tested on unseen images to evaluate its performance.

4.2.5. Output Generation & Evaluation

The final phase involves generating classification results and evaluating the
model’s performance. This includes:
Classification Output: The model predicts whether the detected tumor is benign or
malignant, providing confidence scores for each prediction.

25
Performance Metrics: The system is evaluated using various metrics, such as:
Accuracy: Measures the overall correctness of predictions.
Sensitivity (Recall): Indicates the model’s ability to detect malignant tumors.
Precision: Reflects how often predicted malignant cases are actually malignant.
F1-Score: A balance between precision and recall.

Additionally, heatmaps and Grad-CAM visualization techniques are used to highlight

the regions in the CT scans that influenced the model’s decision, improving result
interpretability.

Visualize tumor predictions: This involves generating visual representations of the

identified tumors, such as highlighting the tumor regions on the CT scan images. This
allows radiologists to easily assess the location, size, and shape of the detected
abnormalities.
Generate diagnostic report: The system compiles the findings into a comprehensive
report, summarizing the detected tumors, their characteristics, and the model's
confidence in the diagnosis. This report aids clinicians in making informed decisions
about patient care.
Validate against expert evaluations: The model's predictions are compared with
assessments made by experienced radiologists or pathologists. This validation step is
crucial for ensuring the accuracy and reliability of the AI-powered system and for
building trust in its performance.

This outlines the structured workflow for lung tumor detection using deep
learning. By incorporating data preprocessing, 2D to 3D conversion, feature extraction,
model training, and performance evaluation, the system aims to enhance the accuracy
and reliability of tumor classification. The use of CNNs and 3D volumetric data ensures
improved detection performance, making this approach a valuable contribution to the
field of medical image analysis. The structured pipeline facilitates real-world
implementation, bridging the gap between AI-based diagnostics and clinical
applications.

26
 4.3 SELECTED METHODOLOGIES

4.3.1 3D Convolutional Neural Network (3D CNN):

3D Convolutional Neural Networks (3D CNNs) are a type of deep learning
architecture that extends the capabilities of traditional 2D CNNs to process three-
dimensional data. This makes them particularly well-suited for applications involving
volumetric data, such as medical imaging (e.g., CT scans, MRI scans). Here's a
breakdown of key aspects:

Fig:4.2: Architecture Design

Extension of 2D CNNs:

• Traditional 2D CNNs operate on images, which are two-dimensional arrays of

pixels.

• 3D CNNs, on the other hand, work with three-dimensional data, such as voxel
data from CT scans, where each voxel represents a point in 3D space.

3D Convolutions:

• The core operation in a CNN is convolution, where filters (or kernels) slide
across the input data to extract features.

• In 3D CNNs, these filters are also three-dimensional, allowing them to capture

spatial features in all three dimensions. This is crucial for understanding the
volumetric structure of data.

27
Volumetric Feature Extraction:

• 3D CNNs excel at capturing spatial relationships and dependencies within 3D

data. This is particularly important in medical imaging, where the shape, size,
and location of tumors or other abnormalities are critical for diagnosis.

Applications in Medical Imaging:

• Lung cancer detection: 3D CNNs are widely used to analyze CT scans for the
detection and classification of lung nodules.

• Brain imaging: They are also used for analyzing MRI scans of the brain to
detect abnormalities such as tumors or lesions.

• Other medical applications: 3D CNNs are used in various other medical

imaging applications, including cardiac imaging and pathology.

Advantages of 3D CNNs:

• Improved Spatial Understanding: 3D CNNs can capture the complete 3D

context of volumetric data, leading to more accurate feature extraction.

• Enhanced Detection Accuracy: In applications like lung cancer detection, 3D

CNNs have shown to improve the accuracy of nodule detection and
classification compared to 2D CNNs.

• Direct Volumetric Analysis: They can directly process volumetric data,

eliminating the need for slice-by-slice analysis, which can lose important spatial
information.

Challenges:

• Computational Cost: 3D CNNs are computationally intensive, requiring

significant processing power and memory.

• Data Requirements: Training 3D CNNs requires large datasets of labeled 3D

data, which can be challenging to obtain in some medical applications.

28
 • Overfitting: Due to the large number of parameters, 3D CNNs are prone to
overfitting, especially when trained on small datasets.

4.3.2 Transfer Learning:

Transfer learning is a machine learning method that uses pre-trained models,

already learned on big datasets, to enhance performance on novel, similar tasks,
particularly when data is scarce. Rather than starting from a zeroth model and training
it, transfer learning consists of fine-tuning a pre-trained model by holding its initial
layers constant, having learned general features, and adapting its final layers to the
particular task, and then training on the new dataset. This method greatly shortens
training time, enhances generalization, and enables efficient model building even from
small datasets, making it especially useful in areas such as medical imaging, where
labeled data tends to be limited, by facilitating the transfer of strong, pre-learned
features to a particular diagnostic task, e.g., lung cancer detection.

4.3.3 Preprocessing and Data Augmentation:

Preprocessing and data augmentation are critical methods in deep learning for
lung cancer detection, especially with medical imaging data. Preprocessing improves
data quality by normalizing pixel values, removing noise, resampling to uniform voxel
spacing, performing CT windowing to emphasize relevant tissues, and segmenting out
extraneous information, thereby enhancing model input. Data augmentation
synthetically increases the training dataset by applying transformations such as
rotation, translation, flipping, scaling, shearing, adding noise, and elastic deformations,
and in 3D, volumetric transformations, which are essential for reducing overfitting and
enhancing model generalization, particularly considering the usually small size of
medical imaging datasets. These methods allow the model to learn strong features and
identify lung nodules under varying conditions, ultimately improving detection accuracy
and clinical utility.

29
4.3.4 Model Training and Validation:

Model training and validation and are critical to the creation of resilient lung
cancer detection systems. Training entails iteratively tuning model parameters with a
training set, with guidance from a loss function and optimization algorithm, while
tracking performance on a validation set to optimize hyperparameters and avoid
overfitting. The data is usually divided into training, validation, and test sets, with
validation making sure the model performs well on unseen data through metrics such
as accuracy, precision, recall, F1-score, AUC, Dice score, and sensitivity/specificity,
particularly important in medical imaging. Cross-validation adds further robustness to
the evaluation by employing several data folds. Lastly, the test set gives an objective
measurement of the final performance of the model, a measure of clinical usefulness
and reliability in lung cancer detection.

4.3.5 Deployment and Integration:

After validation, the system can be incorporated into clinical workflows, which can
benefit radiologists by automating tumor detection and marking areas of interest in lung
scans. This can greatly minimize the time needed for diagnosis and enhance patient
outcomes by allowing early detection and treatment. In conclusion, our proposed
system combines the cutting-edge feature extraction features of 3D CNNs with the
efficiency of transfer learning to develop a strong and precise lung tumor detection
device. This method has the potential to improve the diagnostic process and aid
healthcare professionals in offering timely and accurate cancer treatment.

Deployment and integration are the last essential steps in bringing a successful lung
cancer detection model from the research lab into operational clinical use. Deployment
entails encapsulating the trained model into an operational form, usually in the form of
a software application or web-based tool, so that it can run on new patient CT scans
with optimal speed. Integration focuses on smoothly integrating this AI tool into current
hospital workflows and imaging systems, necessitating keen attention to data security,
patient privacy, and compatibility with electronic health records. This process requires
extensive testing in operational clinical environments, overcoming possible technical
30
issues, and including proper training for radiologists and medical personnel for proper
utilization. Finally, successful

deployment and integration seek to increase diagnostic reliability, lower radiologist

workloads, and enable earlier interventions, resulting in better patient outcomes.

4.4 SOFTWARE REQUIREMENTS

• Google Colab Environment:

o Colab provides a pre-configured environment with Python, common data science
libraries, and GPU/TPU support.
o You'll primarily interact with Colab through Jupyter notebooks.

• Python Libraries (Pre-installed or Easily Installable):

o TensorFlow or PyTorch: Colab supports both frameworks. You can choose based
on your preference.
o NumPy: For numerical computations.
o Pandas: For data manipulation.

• Medical Imaging:
o SimpleITK: For medical image processing (install with !pip install SimpleITK).
o Nibabel: For working with neuroimaging data (install with !pip install nibabel).
o pydicom: For reading DICOM files (install with !pip install pydicom).

• Visualization:
o Matplotlib: For basic plotting.
o Seaborn: For enhanced statistical plots.
o Plotly: for interactive plots.

31
 CHAPTER 5

RESULTS AND DISCUSSION

The application of 3D Convolutional Neural Networks (CNNs) for the detection
of lung tumors from CT scans has yielded significant contributions to medical imaging.
The medical field benefits greatly from using 3D Convolutional Neural Networks (CNNs)
to scan CT scans for lung tumors. CT scans must be evaluated throughout their whole
3D space to overcome 2D CNN problems which study individual picture layers. 3D CNN
results in a better understanding of lung nodule characteristics. Researchers find 3D
CNNs effective although they need much processing power and work poorly for
variations in CT scan quality. Experts develop better network designs while making
their systems perform better to help diminish false alarms in their long-term mission to
provide trustworthy lung tumor detection tools.

Fig:5.1: Original Lung image

The Fig:5.1 represents a CT scan of the chest, commonly used in medical imaging
for lung tumor detection and diagnosis. The scan is an axial (cross-sectional) view,
showing detailed internal structures of the chest cavity. Key anatomical features visible
include the lungs, trachea, and major blood vessels. The lungs appear as large, dark
regions on either side, indicating air-filled spaces, while the surrounding

32
lighter areas represent soft tissues, bones, and organs. Circular structures in the center
denote major blood vessels such as the aorta and pulmonary arteries.

Fig:5.2: Output of the Proposed system

The Fig:5.2 represents the 2D output of a lung tumor detection and segmentation
model applied to a CT scan. The left side, labeled "Original Image," shows a grayscale
cross-sectional CT scan of the lungs, while the right side, labeled "Tumor Highlighted,"
displays the same scan with red-colored regions indicating detected tumor areas.

The below Fig:5.3 represents a CT scan of the chest with tumor regions highlighted,
serving as an output from our lung tumor detection [Link] scan is a 2D axial cross-
section of the lungs, with red-marked areas indicating detected tumor regions. These
highlighted areas suggest abnormal tissue growth, which may correspond to malignant
lung tumors or other pathological changes.

33
 Fig:5.3: Tumor Highlighted lung image

The tumor detection process involves applying a 3D Convolutional Neural Network

(3D CNN), which analyzes CT scan slices to identify and segment suspicious regions.
The CNN model processes volumetric medical images, distinguishing between normal
and abnormal lung tissues based on learned patterns. In this image, the model has
successfully localized tumor regions, aiding in visualization and diagnosis.

Fig:5.4: Tumor lung converted to 3D

34
 This Fig:5.4 represents a 3D reconstructed model of the lungs, generated as an
extension of the 2D lung tumor detection output from our project. The visualization
showcases a volumetric representation of the lungs, enabling a more comprehensive
analysis of the tumor distribution. By converting 2D CT scan slices into a 3D model.

By integrating 3D Convolutional Neural Networks (3D CNNs), our project

enhances traditional diagnostic methods by enabling automated tumor detection in
volumetric lung data. This step marks a significant advancement in medical imaging,
allowing for a more detailed and accurate assessment of lung tumors, ultimately
improving early detection and patient outcomes.

Fig:5.5: Tumor affected area

The Fig:5.5 represents a 3D visualization of tumor spread within the lungs, plotted using
volumetric medical imaging data, likely from a CT scan. The blue points illustrate the
detected tumor regions distributed across the lung structure. The grid background and
axis labels indicate the spatial dimensions of the scan, helping to map the tumor's
extent accurately.

35
 Fig:5.6: Ribcage of lung

The above Fig:5.6 represents a 3D reconstructed model of the rib cage, obtained
as part of our lung tumor detection project using CT scan data. The visualization
showcases the skeletal structure surrounding the lungs, including the rib bones and
portions of the thoracic cavity. This reconstruction is generated from volumetric medical
imaging by applying 3D segmentation and rendering techniques to extract bone
structures from the original CT scan data.

The purpose of this rib cage reconstruction is to provide a clear anatomical

reference, enabling better localization of lung tumors concerning the surrounding
skeletal framework. This is particularly useful in medical imaging and surgical planning,
as it helps radiologists and surgeons assess the spatial relationships between lung
abnormalities and bony structures.

By integrating 3D Convolutional Neural Networks (3D-CNNs) into our project, we

can efficiently analyze volumetric lung and rib cage data, facilitating more precise tumor
detection while preserving anatomical details. This step marks a significant

36
advancement in AI-assisted medical imaging, improving diagnostic capabilities and
supporting better clinical decision-making.

Fig: 5.7: Comparison Parameters of 3D-CNN, 2D-CNN and RNN

From the Fig:5.7:

• 3D CNN performs the best across all metrics, achieving the highest accuracy,
precision, recall, and confidence, indicating its superior capability in tumor
detection and classification.
• 2D CNN shows a noticeable decline in performance, performing worse than 3D
CNN but better than RNN. This suggests that while 2D CNNs are effective, they
lack the ability to fully capture spatial and volumetric information.
• RNN has the lowest performance among the three models, suggesting that
recurrent neural networks, which are typically suited for sequential data, are not
optimal for medical image classification tasks.
37
 Table: [Link] Performance Analysis of 3D-CNN and 2D-CNN

The Table:5.1 presents a comparative analysis of the performance metrics

between 3D Convolutional Neural Networks (3D-CNN) and 2D Convolutional Neural
Networks (2D-CNN) for lung tumor detection. The evaluation is based on multiple
performance indicators, including accuracy, precision, recall, F1-score, probability of
correct classification, false positive rate (FPR), and false negative rate (FNR).

From the data, the 3D-CNN model outperforms the 2D-CNN model in all key
performance metrics. The accuracy of 3D-CNN is 90%, which is higher than the 85%
achieved by 2D-CNN. Similarly, precision, which measures the proportion of correctly
identified positive cases among all predicted positives, is 88% for 3D-CNN, compared
to 82% for 2D-CNN, indicating better reliability in detecting tumors. The recall, which
signifies the model’s ability to identify actual tumor cases, is significantly better in 3D-
CNN (91%) than in 2D-CNN (80%), suggesting that the 3D model is more effective in
detecting true positives and reducing false negatives.

38
 Additionally, the F1-score, which balances precision and recall, is 90.5% for 3D-
CNN, compared to 81% for 2D-CNN, further highlighting the robustness of the 3D
model. The probability of correct classification is also higher in 3D-CNN (90-95%) than
in 2D-CNN (80-85%), reflecting its superior predictive capability. In terms of error rates,
3D-CNN exhibits a lower False Positive Rate (FPR) and False Negative Rate (FNR),
both classified as low, whereas 2D-CNN has medium levels of FPR and FNR, indicating
a higher chance of incorrect classifications.

Fig: 5.8: Comparison between 3D and 2D

The comparison between 3D-CNN and 2D-CNN models for lung tumor detection
reveals that 3D-CNN outperforms 2D-CNN in several key metrics. The accuracy of 3D-
CNN is 90%, which is higher than the 85% achieved by 2D-CNN, indicating better
overall classification performance. Precision, which measures the model’s ability to
39
correctly identify positive cases while minimizing false positives, is also higher in 3D-
CNN (88%) compared to 2D-CNN (82%). Similarly, recall, which determines how well
the model identifies actual positive cases, is significantly better in 3D-CNN (91%) than
in 2D-CNN (80%), demonstrating superior sensitivity to lung tumors. Additionally, the
F1-score, which balances precision and recall, is 90.5% for 3D-CNN and only 81% for
2D-CNN, further proving the robustness of the 3D model. The probability of correct
classification also favors 3D-CNN, ranging from 90-95%, whereas 2D-CNN achieves
only 80-85%. Furthermore, the false positive rate (FPR) and false negative rate (FNR)
are lower in 3D-CNN, whereas 2D-CNN has a medium level of FPR and FNR, indicating
a higher likelihood of misclassifications. Overall, the results suggest that 3D-CNN is a
more effective model for lung tumor detection, offering higher accuracy, better
classification probabilities, and lower error rates compared to 2D-CNN.

Overall, this comparative study demonstrates that 3D-CNN provides significant

improvements in lung tumor detection by enhancing accuracy, precision, recall, and
overall classification reliability while reducing misclassification rates. This confirms the
effectiveness of 3D deep learning models in analyzing volumetric medical imaging,
making them a superior choice for lung tumor detection compared to traditional 2D-
CNN models.

5.1 CONSTRAINTS AND TRADEOFF

Developing a lung tumor detection system using 3D CNN involves balancing

multiple constraints and tradeoffs to ensure accuracy, efficiency, and reliability. One
major constraint is the limited availability of high-quality, annotated medical imaging
datasets, which impacts model training and generalization. The computational
complexity of processing 3D medical images requires significant hardware resources,
making real-time analysis challenging. Deep learning models, especially 3D CNNs,
demand substantial processing power and memory, which can limit deployment on
standard computing systems. The need for high-performance GPUs and large storage
capacities increases the overall cost of system development. Additionally, deep
learning models often lack interpretability, creating difficulties in gaining trust

40
from medical professionals. Preprocessing steps, such as noise reduction and
normalization, add further complexities in standardizing input data.

In terms of tradeoffs, there is a constant balance between accuracy and

computational efficiency, more complex models yield higher precision but require
greater processing power, while simpler architectures improve speed but may
compromise detection performance. Training a model on a highly curated dataset can
yield high accuracy for a specific imaging environment, but it may not perform well on
images acquired from different medical centers with varying scanning protocols.
Simpler machine learning models, such as decision trees or traditional classifiers, offer
better interpretability but lack the performance of deep learning models. Similarly,
ensuring model robustness across diverse patient data may reduce specificity in certain
cases. Addressing these constraints and tradeoffs is essential for creating a practical
and clinically acceptable tumor detection system.

41
 CHAPTER 6

SUMMARY AND CONCLUSIONS

6.1 SUMMARY

Early diagnosis of lung tumors becomes very important because this disease kills
more patients around the world than any other disease. Trials that depend on chest X-
rays CT scans MRI and PET scans plus biopsy tests for lung tumor identification need
much human participation from radiologists who might mistake results because of
complex lung tumor structures. This project creates an automated system based on
deep learning through 3D Convolutional Neural Networks to better detect lung tumors
from 3D CT scan data.

The proposed 3D-CNN model works with multiple CT images to find important
image details to pinpoint tumor areas and tell malignant from non-malignant tissue. 3D-
CNN analysis outperforms 2D-CNN methods because it evaluates complete volumes
of images to find tumors better. We begin with image preparation that involves reducing
noise and enhancing contrast while segmenting lung areas before sending processed
images into our model.

Our study found that 3D-CNN generated better results when compared to 2D-CNN
by important performance standards. Tests show that the 3D-CNN model delivers 90%
accuracy with 88% precision and 91% recall for an F1-score of 90.5% which proves
superior results when compared with the 2D-CNN model achieving 85% in accuracy
while scoring 82% precision, 80% recall, and 81% F1-score. The new 3D-CNN
approach showed lower FPCR and FNR rates which made disease identification more
reliable for doctors.

By creating three-dimensional renderings of lungs and rib structures we could

better see where lesions grew in the body. These can use these created images to
42
review how far and deep cancer cells have moved in patients to help doctors decide
what treatments should happen first.

The project results show that 3D-CNNs enhance lung tumor detection better than
traditional methods through improved accuracy and faster processing. The research
supports AI systems for faster lung cancer detection with deep learning and medical
image analysis which helps to prevent deaths from this disease.

6.2 CONCLUSIONS

Our project shows that 3D-CNN helps medical imaging by making better tumor
diagnosis and helping computers spot tumors. The study shows that 3D deep learning
methods detect tumors more effectively than standard 2D systems because they
process complete data volumes better than flat images. The new procedure lets
doctors find tumors earlier when lung cancer screening works best to lower mortality
rates from lung cancer.

The 3D model of the lungs plus rib cage improves diagnosis results for radiologists
and oncologists by providing them with clear anatomical information. Thanks to
improved 3D-CNN performance the model needs less human interpretation to enhance
medical diagnosis activities. The investigation displays excellent performance but
requires additional research using extensive varied medical images to boost accuracy
and decrease possible biases.

The system must work correctly with healthcare partners to show effectiveness
before medical teams start using it at their clinics. New research should examine better
model designs and include multiple imaging scans like PET-CT and MRI along with
creating AI tools to help medical experts in their work. The project shows clear AI
potential to detect lung tumors by using 3D-CNNs as an early cancer detection tool.

43
 Medical imaging can achieve better diagnostic results and lower the risk of
mistakes through deep learning applications and enhanced treatment planning
benefits patients. AI healthcare technologies are advancing to help defeat lung cancer
and save lives from dangerous diseases.

44
 REFERENCES

[1] Alakwaa, W., Nassef, M., Badr, A., "Lung cancer diagnosis and classification
using 3D convolutional neural networks," International Journal of Advanced
Computer Science and Applications, 2017.
[2] Armato, S. G., Giger, M. L., Moran, C. J., Blackburn, J. T., Doi, K., MacMahon,
H., "Computerized detection of pulmonary nodules on CT scans," RadioGraphics,
vol. 19, no. 5, pp. 1303-1311, 1999.
[3] Armato, S. G., Giger, M. L., Moran, C. J., Blackburn, J. T., Doi, K., MacMahon,
H., "Computerized detection of pulmonary nodules on CT scans," RadioGraphics,
vol. 19, no. 5, pp. 1303-1311, 1999.
[4] Bhuvaneswari, K., Bharathy, S., Padmasri, K. K., Tanushree, G., "A deep learning
approach to detect and classify lung cancer," International Journal of Machine
Learning and Cybernetics, vol. 14, no. 3, pp. 711-722, 2023.
[Link]
[5] Brown, A., Wilson, E., & Davis, F. (2022). Interpretable AI for Lung Cancer
Diagnosis: Understanding Model Decision-Making. Nature Medicine.
[6] Chon, A., Lu, P., "Deep CNN for lung cancer diagnosis," Stanford University,
2017.
[7] Daassi, I., Sayadi, M., Ben Slama, A., Trabelsi, H., Barbaria, S., "An Effective
Method for Lung Tumor Screening Using CT Dataset," 2023.
[8] Gao, H., Yi, M., Yu, J., Li, J., Yu, X., "Character segmentation-based coarse-fine
approach for automobile dashboard detection," IEEE Transactions on Industrial
Informatics, vol. 15, no. 10, pp. 5413-5424, 2019.
[9] Garcia, R., Lopez, A., & Martinez, C. (2023). Integration of AI-Assisted Diagnosis
into Clinical Workflows for Lung Cancer Screening. Radiology.
[10] Hossain, S., Najeeb, S., Shahriyar, A., Abdullah, Z. R., Haque, M. A., "Lung
cancer classification using deep convolutional neural networks," International
Journal of Computer Applications, vol. 182, no. 42, pp. 25-31, 2023.
[Link]

45
46