IDC Breast Cancer Classification 🧬

Overview

This project focuses on the classification of Invasive Ductal Carcinoma (IDC) in breast histopathology image patches. Using a dataset of 50×50 pixel images of tissue samples (either healthy or IDC-positive), we compare two machine learning pipelines:

1. Hybrid Approach - CNN-based feature extraction followed by traditional ML classifiers.
2. End-to-End Deep Learning - Fine-tuning pre-trained CNN models directly on labeled image patches.

---

📁 Project Files

📂 IDC_Breast_Cancer_Classification

├── IDC Breast Cancer Classification_Report.docx # Project report

├── idc-breast-cancer-classification_Final.ipynb # Jupyter notebook (main pipeline)

├── idc-breast-cancer-classification_Final.pdf # Notebook exported as PDF

├── idc-breast-cancer-classification_Final.html # Notebook exported as HTML

├── README.md # Project documentation (this file)

---

Dataset

• Source: Breast Histopathology Images – Kaggle
• Size: 277,524 image patches (50×50 pixels)
• Classes:
  • 0 – Healthy Tissue
  • 1 – IDC-Positive (Invasive Ductal Carcinoma)

---

Methodology

🔹 Data Preprocessing

• Balanced subset: 12,000 samples per class
• Images resized to 50×50 RGB
• Dataset split: 70% training, 30% testing

🔹 Hybrid Modeling (Feature Extraction + ML Classifiers)

• Used MobileNetV2 (ImageNet pretrained, top layers removed) for feature extraction
• Trained models:
  • Logistic Regression
  • Random Forest
  • K-Nearest Neighbors
  • Gradient Boosting
  • XGBoost
• Best performance: XGBoost with 79% accuracy

🔹 End-to-End Deep Learning

• Fine-tuned CNN models:
  • MobileNetV2: 84.07% test accuracy
  • ResNet50: 84.57% test accuracy
  • EfficientNetB0: 85.40% test accuracy
• Models trained with dropout, global average pooling, and a softmax head
• Evaluation metrics: Accuracy, Confusion Matrix, ROC-AUC

---

Key Results

• Best ML model (XGBoost): ~79% accuracy
• Best Deep Learning model (EfficientNetB0): 85.4% accuracy
• End-to-end fine-tuning significantly outperformed hybrid methods due to its ability to adapt feature extraction to task-specific patterns.

---

Future Enhancements

• Integrate with real-time pathology tools for deployment
• Evaluate generalizability on whole-slide histology images

---

👨‍💻 Contributor

• Gandhar Ravindra Pansare (Master’s in Data Science – Indiana University Bloomington)
• Guided by Professor Krista Li

---

⚖️ License

This project is licensed under the MIT License.

---

This project bridges computer vision and oncology using advanced machine learning to support faster, more accurate cancer diagnostics.