End-to-end breast cancer detection system using thermal imaging.
Built with YOLOv8 for classification, enhanced with Grad-CAM explainability, and deployed on a
Raspberry Pi 5 + MLX90640 thermal camera.
This project targeted early, non-invasive detection of breast cancer using thermal imaging. Infrared thermography captures heat patterns, which can highlight vascular and metabolic activity linked to tumors. I trained deep learning models to classify benign vs malignant conditions and later optimized deployment on low-power embedded hardware.
The final YOLOv8 model achieved over 90% accuracy with balanced class performance across benign and malignant cases. Grad-CAM visualizations validated that the model focused on biologically relevant thermal patterns. The embedded Raspberry Pi deployment ran in real time, proving the feasibility of portable low-cost screening systems.
The dataset suffered from class imbalance (more benign samples). I addressed this by applying data augmentation and class rebalancing during training. Running YOLOv8 on Raspberry Pi required quantization and hardware-aware optimization. The addition of Grad-CAM addressed the issue of clinical trust by making the AI’s decisions interpretable to medical practitioners.