Most cases in colorectal cancer begin as polyps — small growths on the colon wall that are easily removed if caught early. Detecting these polyps can be missed out by the doctors. A real-time AI system can act as a second pair of eyes to detect the polyps that a doctor might miss.
PolyDetect is a deep learning system that:
- Detects polyps in colonoscopy images
- Draws a pixel-level segmentation mask around each polyp
- Outputs a confidence score per detection
- Runs as a live web demo anyone can use
YOLOv8m-seg
Kvasir--SEG
- Computer vision — instance segmentation, object detection
- Deep learning — YOLOv8, transfer learning
- Data pipeline — mask to polygon conversion, YOLO format
- Model evaluation — mAP, Dice coefficient, IoU, precision/recall
- Deployment — Gradio web demo
Evaluated on 100 held-out test images from Kvasir-SEG that the model never saw during training.
| Metric | Score |
|---|---|
| mAP@50 | 0.920 |
| mAP@50-95 | 0.751 |
| Dice coefficient | 0.882 ± 0.170 |
| IoU | 0.818 ± 0.200 |
| Recall | 0.941 |
| Precision | 0.865 |