Polyp segmentation U-Net (ResNet-34, Kvasir-SEG)
Binary colon-polyp segmentation: a U-Net with an ImageNet-pretrained ResNet-34 encoder and
Dropout2d(0.2) after each decoder block, trained on 800 Kvasir-SEG frames. Training code,
evaluation and error analysis: github.com/haroonkhan02/polyp-segmentation-unet.
Browser demo: spaces/haroonx02/polyp-segmentation-browser.
Not for clinical use. Research and portfolio model trained on one hospital's data, with no validation on other endoscopes or populations and no regulatory clearance.
Files
| File | What it is | Size | Test mean Dice |
|---|---|---|---|
unet_resnet34.pt |
PyTorch checkpoint (state_dict + training config) |
98 MB | 0.904 |
onnx/polyp_unet_fp16.onnx |
ONNX, FP16 weights, float32 input/output | 49 MB | 0.904 |
onnx/polyp_unet_int8.onnx |
ONNX, static INT8 (QDQ, per-channel, 128 train frames for calibration) | 25 MB | 0.904 |
onnx/polyp_unet_mc8_int8.onnx |
INT8 Monte-Carlo-dropout graph: 8 decoder samples, outputs mean and std |
25 MB | 0.904 (mean) |
ONNX input image: float32 (1, 3, 320, 320), RGB in [0, 1] (resize with area interpolation,
divide by 255; ImageNet normalisation is inside the graph). Output probs: (1, 1, 320, 320)
polyp probability (sigmoid inside the graph); resize back bilinearly and threshold at 0.5.
Evaluation
Held-out test split of 100 Kvasir-SEG images (committed seeded split), threshold 0.5, original resolution. Mean per-image Dice 0.904 and IoU 0.844 for the PyTorch model; the ONNX FP32/FP16 and INT8 files score the same to within 0.001 (INT8: 0.9043 vs 0.9036). By polyp size: small (< 5 % of the frame) 0.868, medium 0.938, large 0.867. The MC-dropout std map detects the model's pixel errors with an AUROC between 0.91 and 0.94 (8 samples; it varies by up to 0.02 between random seeds). It flags uncertain boundaries well but misses regions the model is confidently wrong about.
Limitations
Single source (one Norwegian hospital), every frame contains a polyp (false-positive rate on healthy mucosa unknown), 100 test images (about ±0.025 on mean Dice), errors are mostly under-segmentation of flat polyps.
Data
Kvasir-SEG: D. Jha et al., "Kvasir-SEG: A Segmented Polyp Dataset", MMM 2020. CC BY 4.0.