🧬 Early Cancer Detection Using Multi-Scale CNN and Transformer-Based Deep Learning Approaches

> Author: Xinyu Luo

> Institution: Oxford Brookes University × Chengdu University of Technology

> Supervisor: Dr. Grace U. Nneji

> Module Code: CHC6096

📌 Overview

This repository presents an advanced hybrid deep learning framework for early breast cancer detection based on histopathological image analysis. The model combines EfficientNetV2 for local feature extraction with Vision Transformers (ViT) enhanced by Shifted Patch Tokenization (SPT) and Learned-Scale Attention (LSA) mechanisms.

Two public histopathology datasets are used:

• BreakHis: Breast Cancer Histopathological Image dataset (7,909 samples across 4 magnification levels: 40x, 100x, 200x, 400x)
• BACH: BACH Grand Challenge dataset, for cross-dataset validation.

---

💡 Key Features

• ✅ Hybrid CNN-ViT architecture for combining local and global image understanding
• ✅ Shifted Patch Tokenization (SPT): enables spatial context enhancement for ViT
• ✅ Learned-Scale Attention (LSA): dynamic attention scaling for better interpretability
• ✅ Custom data split and augmentation strategy for training robustness
• ✅ ROC, confusion matrix, AUC, and other metric visualizations
• ✅ Integrated Streamlit GUI for user-friendly model interaction
• ✅ Modular TensorFlow 2.13 implementation with GPU memory control

---

🧠 Architecture Highlights

Module	Description
EfficientNetV2-B0	Extracts multi-scale local features from histopathological images
Vision Transformer	Captures global context via multi-head self-attention
SPT	Improves ViT patch representation via local shifts
LSA	Learnable scaling in attention to improve focus on relevant regions
Fusion Strategy	Concatenation of CNN and ViT branches followed by dense classifier head

---

🖥️ GUI with Streamlit

To launch the Streamlit-based GUI interface:

cd gui
streamlit run main.py

The GUI supports:

• Image upload and preprocessing
• Real-time inference using trained hybrid model
• Visual display of prediction (Benign vs. Malignant)
• Confidence scores and Grad-CAM support (optional)

⚙️ Installation

Requirements

pip install -r requirements.txt

Ensure the following Python libraries are installed:

• tensorflow>=2.13
• scikit-learn
• opencv-python
• streamlit
• matplotlib, seaborn
• pandas, numpy

Clone Repository

git clone https://github.com/CarsonLLuo/FinalProject.git
cd FinalProject

👩‍⚕️ Target Audience

• Medical professionals (pathologists, oncologists)
• Medical AI researchers
• Developers building AI-assisted diagnostic tools
• Institutions seeking to deploy lightweight, accurate cancer screening models

🙏 Acknowledgment

• Supervisor: Dr. Grace U. Nneji
• Module Leader: Dr. Joojo Walker
• Dataset providers: BreakHis, BACH Grand Challenge
• Special thanks to teammates, dormmates, and collaborators in 一緒Impactですよ for their unwavering support

📜 License

This project is distributed for educational and research purposes. Please contact the author for any commercial inquiries.