A MODERN MACHINE LEARNING-DRIVEN
FRAMEWORK FOR LUNG CANCER DETECTION AND
CLASSIFICATION
PRESENTED BY SUPERVISOR
N MANIKANDAN, DR.P.A GOWRI
SANKAR.,M.E.,PH.D.,
PH.D SCHOLAR, ASSOCIATE PROFESSOR,
ASSISTANT PROFESSOR KNOWLEDGE INSTITUTE OF
TECHNOLOGY(AUTONOMOUS),SALEM
SSM INSTITUTE OF ENGINEERING
AND TECHNOLOGY(AUTONOMOUS), DINDIGUL
ABSTRACT
• Lung cancer remains a leading cause of cancer related mortality worldwide, making early and accurate detection
critical for improving patient survival.
• In this paper, a django-based web platform is presented for automated lung cancer classification that integrates
advanced deep learning and probabilistic models.
• Existing methods—Convolutional neural networks (CNN) and Residual networks (ResNet)— are effective in
feature extraction but limited in handling diagnostic uncertainty and relational dependencies within medical data.
• To address these gaps, the proposed system introduces Bayesian networks for probabilistic reasoning and Graph
Neural Networks (GNN) for modeling spatial relationships among lung tissue regions.
• Using the LC25000 histopathological dataset and extensive preprocessing, the hybrid approach achieves superior
performance, with Bayesian networks reaching 98.01 % accuracy and 97.99 % f1-score, outperforming CNN
and ResNet baselines.
• Deployed as A secure, user-friendly django webserver, the system allows clinicians to upload CT images and
receive real-time, explainable predictions, reducing false negatives and enabling timely intervention. This platform
demonstrates A scalable, interpretable, and clinically viable solution for early lung cancer detection.
INTRODUCTION
• Modern, hybrid framework for lung cancer detection that integrates convolutional neural
networks (CNN), Bayesian networks (BN), and graph neural networks (GNN). While CNN
and Resnets have shown strong performance in image classification, they struggle with
uncertainty management and spatial relational modeling.
• To address these limitations, the proposed system combines probabilistic reasoning via BN
and spatial dependency modeling via GNN, resulting in A more interpretable and robust
diagnostic tool.
• Deployed through a secure Django-based web platform, the system allows clinicians to
upload ct or histopathological images and receive real-time, explainable predictions.
• This approach not only improves diagnostic accuracy but also enhances clinical trust and
usability, marking A significant step forward in ai-assisted lung cancer detection.
LITERATURE REVIEW
EXISTING SYSTEM
• The existing system for lung cancer detection primarily relies on CONVOLUTIONAL NEURAL
NETWORKS (CNN) and RESIDUAL NEURAL NETWORKS (RESNET) as baseline models. These deep
learning architectures are widely used in medical image analysis due to their ability to extract hierarchical
features from complex visual data.
• CONVOLUTIONAL NEURAL NETWORK (CNN)
• Functionality: CNNS automatically learn spatial features from lung images through layers of
convolutions and pooling.
• Architecture highlights:
• Convolutional layers detect edges, textures, and patterns.
• Pooling layers reduce dimensionality while preserving important features.
• Dropout layers prevent overfitting.
• Final dense layer with sigmoid activation performs binary classification (cancerous vs. Non-cancerous).
• LIMITATIONS:
Moderate precision leads to false positives, and it lacks interpretability and uncertainty
quantification.
• RESIDUAL NEURAL NETWORK (RESNET)
• FUNCTIONALITY: ResNet Enhances CNNS By Introducing Skip Connections That Allow Deeper Networks
To Train Effectively Without Vanishing Gradients.
• ARCHITECTURE HIGHLIGHTS:
• Pretrained Resnet-50 Is Fine-tuned On Lung Cancer Data.
• Captures Subtle And Complex Patterns In Histopathological Images.
LIMITATIONS:
Slightly Lower Accuracy And Recall Than CNN; Still Lacks Spatial Modeling And Uncertainty
Handling.
PROPOSED SYSTEM
• The Proposed System Introduces A Hybrid Machine Learning Framework For Early And Accurate Lung
Cancer Detection, Combining The Strengths Of Bayesian Networks (BN), And Graph Neural Networks
(GNN). It Is Deployed Via A Secure And User-friendly Django Web Platform, Making It Accessible To
Clinicians In Real-time.
• 1. Data Acquisition & Preprocessing
• Uses The LC25000 Histopathological Image Dataset With 25,000 Labeled Images Across Three Categories:
LUAD, LUSC, And Benign Tissue.
• Preprocessing Steps Include:
• Histogram Equalization: Enhances Contrast.
• Threshold Segmentation: Isolates Lung Regions.
• Normalization & Augmentation: Standardizes Inputs And Expands Dataset Diversity.
• Baseline Feature Extraction
• CNN And Resnet Are Used To Extract High-dimensional Features From Images.
• CNN Captures Local Spatial Patterns.
• Resnet Uses Skip Connections For Deeper Feature Learning.
• These Models Serve As Both Classifiers And Feature Generators For The Hybrid System.
• Bayesian Network (BN) For Probabilistic Reasoning
• Models Conditional Dependencies Among Features Like Nodule Size, Texture, And
Demographics.
• Provides Probability Scores Instead Of Binary Outputs, Helping Clinicians Assess
Diagnostic Confidence.
• Enhances Interpretability And Manages Diagnostic Uncertainty.