Lung Cancer Detection using CNN Transfer Learning

📋 Project Overview

This project implements a deep learning-based system for detecting lung cancer from CT scan images using state-of-the-art Convolutional Neural Networks (CNNs) with transfer learning. The system classifies CT images into three categories: Benign, Malignant, and Normal. After evaluating multiple pre-trained architectures, MobileNetV2 emerged as the best-performing model with fine-tuning, achieving high accuracy and robust generalization.

The project includes comprehensive exploratory data analysis, model comparison, performance evaluation, and interpretability using Grad-CAM visualizations to provide explainable AI insights for medical diagnosis.

---

📊 Dataset

Dataset Link: https://www.kaggle.com/datasets/rayhankhan831/lung-cancer-ct-scan-image-for-federated-learning

The dataset consists of 3,206 lung CT scan images distributed across three classes:

Class	Count	Percentage
Malignant	1,308	40.8%
Normal	986	30.8%
Benign	912	28.4%

Sample Images

Sample CT scan images from each class (Benign, Malignant, Normal)

The dataset is sourced from multiple clients (Client1-Client4), simulating a federated learning scenario, though this implementation uses centralized training for model comparison.

Dataset Characteristics

• Total Images: 3,206
• Image Format: CT scan images (grayscale converted to RGB)
• Input Size: 224x224 pixels (resized for model compatibility)
• Classes: 3 (Benign, Malignant, Normal)
• Split: Train (80%), Validation (10%), Test (10%)

---

🏗️ Methodology

Data Preprocessing & Augmentation

To enhance model generalization and prevent overfitting, the following augmentation techniques were applied to the training data:

• Rescaling: Pixel values normalized to [0,1]
• Random Horizontal Flip: 50% probability
• Random Rotation: Up to 15 degrees
• Random Zoom: Up to 15%
• Width/Height Shift: Up to 10%

Transfer Learning Models

Four pre-trained CNN architectures were evaluated using transfer learning:

1. MobileNetV2 - Lightweight and efficient architecture
2. EfficientNetV2B0 - State-of-the-art efficiency
3. DenseNet121 - Dense connectivity architecture
4. Xception - Depthwise separable convolutions

Each model was initialized with ImageNet weights, and the top classification layer was replaced with a custom head:

• Global Average Pooling
• Dense Layer (256 units, ReLU activation)
• Dropout (0.5) for regularization
• Output Layer (3 units, Softmax activation)

Training Configuration

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Batch Size: 32
• Initial Epochs: 10
• Fine-tuning Epochs: 10
• Callbacks: Early Stopping, ReduceLROnPlateau, ModelCheckpoint

Fine-tuning Strategy

After initial training with frozen base layers, the best model (MobileNetV2) was fine-tuned by:

1. Unfreezing the top 30 layers
2. Using a lower learning rate (1e-5)
3. Training for an additional 10 epochs

---

📈 Results

Model Comparison

Model	Accuracy	Precision	Recall	F1 Score	AUC
MobileNetV2	87.85%	87.89%	87.85%	87.26%	95.65%
Xception	77.88%	77.47%	77.88%	75.28%	90.88%
DenseNet121	74.14%	75.73%	74.14%	67.46%	91.46%
EfficientNetV2B0	40.81%	16.65%	40.81%	23.66%	59.67%

Performance Visualization

Comparison of accuracy across different CNN architectures

MobileNetV2 significantly outperformed other models, demonstrating excellent classification capabilities for lung CT scan analysis. The high AUC score (95.65%) indicates strong discrimination ability between classes.

Best Model: MobileNetV2 (Fine-tuned)

Confusion Matrix:

	Benign	Malignant	Normal
Benign	61	20	11
Malignant	8	123	0
Normal	0	0	98

Performance Metrics:

• Overall Accuracy: 87.85%
• Precision (Weighted): 87.89%
• Recall (Weighted): 87.85%
• F1 Score (Weighted): 87.26%
• AUC (Macro): 95.65%

Class-wise Performance:

• Benign: Precision 88%, Recall 66%, F1 76%
• Malignant: Precision 86%, Recall 94%, F1 90%
• Normal: Precision 90%, Recall 100%, F1 95%

The model shows exceptional performance in detecting malignant and normal cases, with slightly lower recall for benign cases, which is clinically acceptable as the priority is to identify malignant cases.

---

🔍 Grad-CAM Visualizations

Gradient-weighted Class Activation Mapping (Grad-CAM) provides visual explanations of the model's decision-making process by highlighting the regions in CT images that most influenced the prediction.

Grad-CAM visualization showing the model's focus regions for lung cancer detection

Random Predictions with Grad-CAM

Random test samples with actual labels, predictions, and confidence scores

The Grad-CAM visualizations show that the model focuses on relevant anatomical regions (lung nodules, tissue abnormalities) when making predictions, providing interpretability crucial for medical applications.

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• TensorFlow 2.19.0
• CUDA-capable GPU (recommended)

Installation

1. Clone the repository:

git clone https://github.com/stephinjacob007/Lung-Cancer-Detection-Transfer-Learning.git
cd Lung-Cancer-Detection-Transfer-Learning

2. Install dependencies:

pip install -r requirements.txt

3. Download the dataset:
   • Place the dataset in the appropriate directory structure
   • Dataset should have the following structure:

LungData/
├── Client1/
│   ├── Benign/
│   ├── Malignant/
│   └── Normal/
├── Client2/
│   ├── Benign/
│   ├── Malignant/
│   └── Normal/
├── Client3/
│   ├── Benign/
│   ├── Malignant/
│   └── Normal/
└── Client4/
    ├── Benign/
    ├── Malignant/
    └── Normal/

Running the Code

1. Open the Jupyter Notebook:

jupyter notebook Lung_Cancer_Detection_CNNs.ipynb

2. Execute cells sequentially to:
   • Load and preprocess data
   • Perform exploratory data analysis
   • Train models
   • Evaluate performance
   • Generate visualizations

---

📁 Project Structure

Lung-Cancer-Detection-Transfer-Learning/
├── Lung_Cancer_Detection_CNNs.ipynb   # Main Jupyter notebook
├── requirements.txt                    # Python dependencies
├── Results/                            # README images
|   ├── Confusion-Matrices/             # Confusion matrices of all models
│   ├── Models/                         # Accuracy and Loss curves of models
|   ├── Class_Distribution.png
│   ├── Grad-CAM_Example.png
│   ├── Model_Comparison.png
│   ├── Prediction_With_GRAD-CAM.png
│   └── Sample_Images.png
|
└── README.md

---

🧪 Model Architectures

MobileNetV2 Architecture (Best Model)

Layer Type	Output Shape	Parameters
MobileNetV2 Base (Frozen)	(7, 7, 1280)	2.2M
Global Average Pooling	(1280)	0
Dense (ReLU)	(256)	327,936
Dropout (0.5)	(256)	0
Dense (Softmax)	(3)	771

Total Parameters: ~2.6M (after fine-tuning)

Why MobileNetV2?

• Lightweight: Fewer parameters suitable for deployment
• Efficiency: Depthwise separable convolutions reduce computational cost
• Performance: Best accuracy among tested models
• Generalization: High AUC score indicates robust discrimination

---

📊 Evaluation Metrics

• Accuracy: Overall correctness of predictions
• Precision: Proportion of correct positive predictions
• Recall: Proportion of actual positives correctly identified
• F1 Score: Harmonic mean of precision and recall
• AUC-ROC: Area under the Receiver Operating Characteristic curve

---

🩺 Clinical Implications

1. Early Detection: The model can assist radiologists in early detection of lung cancer, potentially improving patient outcomes through timely intervention.

2. Reduced Subjectivity: AI-based analysis provides consistent and reproducible results, reducing inter-observer variability.

3. Triage Support: Can prioritize high-risk cases for immediate review by specialists.

4. Educational Tool: Grad-CAM visualizations help in understanding radiological features associated with different classes.

---

🔧 Fine-Tuning Results

Fine-tuning the MobileNetV2 model improved performance:

Metric	Before Fine-tuning	After Fine-tuning
Training Accuracy	87.85%	88.18%
Validation Accuracy	86.29%	86.60%
Test Accuracy	87.85%	~88%

Training and validation accuracy/loss curves during fine-tuning

---

🙏 Acknowledgments

• Dataset sourced from Kaggle: Lung Cancer CT Scan Image for Federated Learning
• TensorFlow and Keras teams for providing pre-trained models
• Research community for advancing medical imaging AI

---

📧 Contact

Project Link: https://github.com/stephinjacob007/Lung-Cancer-Detection-Transfer-Learning

---

⚠️ Disclaimer

This project is for research and educational purposes only. The model should not be used as a standalone diagnostic tool. Always consult qualified medical professionals for clinical decisions. The "Not for diagnostic use" label on predictions emphasizes this limitation.

---

🔮 Future Work

• Implement federated learning approach
• Integrate with a web-based interface for clinical use
• Expand dataset with more diverse CT scans
• Implement multi-modal analysis (combine with patient history)
• Deploy as a mobile application using TensorFlow Lite
• Ensemble methods for improved accuracy
• 3D CNN for volumetric analysis of CT scans