| --- |
| license: apache-2.0 |
| tags: |
| - image-segmentation |
| - medical-imaging |
| - polyp-segmentation |
| - pytorch |
| - malunet |
| datasets: |
| - cvc-clinicdb |
| library_name: pytorch |
| pipeline_tag: image-segmentation |
| --- |
| |
| # MALUNet · CVC-ClinicDB (Polyp Segmentation) |
|
|
| Lightweight U-shape segmentation network adapted from |
| [jcruan519/MALUNet](https://github.com/jcruan519/MALUNet) and trained on |
| [CVC-ClinicDB](https://www.kaggle.com/datasets/balraj98/cvcclinicdb) for |
| binary polyp segmentation in colonoscopy frames. |
|
|
| ## Model |
|
|
| - Architecture: MALUNet (DGA + IEA + CAB + SAB) |
| - Channels: `[8, 16, 24, 32, 48, 64]`, `split_att="fc"`, `bridge=True` |
| - Input: RGB, 256×256 |
| - Output: single-channel sigmoid mask (1 = polyp) |
| - Parameters: ~0.18 M |
|
|
| ## Training |
|
|
| - Dataset: CVC-ClinicDB (612 paired image/mask frames) |
| - Split: 80% train / 20% val (seeded by filename, `seed=42`) |
| - Loss: BCE + Dice |
| - Optimizer: AdamW, `lr=1e-3`, `weight_decay=1e-2` |
| - Schedule: CosineAnnealingLR, `T_max=50`, `eta_min=1e-5` |
| - Augmentations: random h/v flip, random rotation |
| - Epochs: 150 |
|
|
| ## Usage |
|
|
| ```python |
| import torch |
| from huggingface_hub import hf_hub_download |
| from infer import load_model, predict_mask # infer.py from this repo |
| from PIL import Image |
| |
| model = load_model("YOUR_USERNAME/malunet-cvc") |
| mask = predict_mask(model, Image.open("polyp.png")) |
| Image.fromarray(mask).save("mask.png") |
| ``` |
|
|
| `infer.py` and `models/malunet.py` are bundled in this repo so you can |
| also clone it and run inference without the original training code. |
|
|
| ## Limitations |
|
|
| - Trained on CVC-ClinicDB only (612 frames, single source). Generalization |
| to other colonoscopy systems / patient populations is unverified. |
| - Not a medical device. Research / demo use only. |
|
|
