Breast-Cancer-Classification

A Comparative Study of SVM, CNN, and Transformer Models on the CBIS-DDSM Dataset

Overview

This project focuses on classifying breast cancer using mammogram images from the CBIS-DDSM dataset. The study implements:

• Support Vector Machines (SVM) with texture-based feature extraction (Gray-Level Co-Occurrence Matrix, GLCM)

• EfficientNet-Based Convolutional Neural Network (CNN) with transfer learning for two-view mammogram classification

• Residual Neural Network (ResNet) for improved feature extraction and classification

Dataset

The CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography) dataset includes:

• 2,620 mammogram images categorized as Normal, Benign, and Malignant

• Region of Interest (ROI) segmentation with verified pathology reports

• DICOM format images, converted to PNG for preprocessing

Methodology

1. Data Preprocessing

• Convert DICOM to PNG
• Grayscale conversion
• Contrast enhancement using histogram equalization
• GLCM-based feature extraction with optimal angle selection

2. Feature Normalization Techniques

• Baseline (No Normalization): Poor stability
• Standard Scaling: Mean = 0, Std Dev = 1
• Power Transformer: Maps features to a Gaussian-like distribution
• Yeo-Johnson Transformation: Normalizes both positive and negative values

3. Model Implementation

Model 1: Support Vector Machines (SVM)

• Uses GLCM features
• Radial Basis Function (RBF) Kernel
• Cross-validation for hyperparameter tuning

Model 2: EfficientNet-Based CNN

• Two-view classification (Craniocaudal & Mediolateral Oblique views)
• Pretrained EfficientNet-B0 weights with fine-tuning
• Data Augmentation: Rotation, zoom, shear, intensity shift

Model 3: Residual Neural Network (ResNet)

• ResNet architecture for robust feature extraction
• Residual connections improve gradient flow and learning stability

Workflow

This is the workflow diagram for the breast cancer classification process:

Evaluation Metrics

• Accuracy, Precision, Recall, F1-Score
• Cross-validation for SVM (varying GLCM angles)
• Hyperparameter tuning using Optuna for CNN and ResNet