EndoTector

Polyp segmentation pipeline: dataset → training → demo

Research tool — NOT a medical device.

The pipeline works end-to-end. Moreover, it achieves great results (mAP@50 ~0.95, median Dice ~0.96) despite being a light model (6 MB). It opens the door to real-time use.

YOLO-seg provides box + mask at the same time.

On the other hand, the test set has only 100 samples, so merging Kvasir-SEG with other datasets would help the model generalize.

Pipeline

data_prep.py          train.py                 app.py
download Kvasir-SEG   fine-tune YOLO26-seg     Gradio demo:
+ masks, polygons   -> 100 epochs          ->  image in, mask +
+ 80/10/10 split        models/best.pt         box

YOLO26-seg fine-tuned on Kvasir-SEG.

Results

Test split (100 images) on an RTX 3060 (6 GB):

Metric	Box	Mask
Precision	0.902	0.902
Recall	0.906	0.906
mAP@50	0.948	0.948
mAP@50-95	0.792	0.805

Mean Dice 0.888 (median 0.961) at conf 0.25. Reproduce the mAP with yolo segment val model=models/best.pt data=data/kvasir-seg/dataset.yaml split=test.

Test the project

The trained weights (models/best.pt) come with the repo, so the demo runs out of the box:

# Environment (PowerShell)
py -m venv .venv
./.venv/Scripts/python.exe -m pip install -r requirements.txt

# Demo, opens the Gradio UI in your browser
./.venv/Scripts/python.exe app.py

Reproduce the model (optional, needs a GPU for the training of YOLO-seg):

./.venv/Scripts/python.exe data_prep.py   # download Kvasir-SEG + build YOLO labels
./.venv/Scripts/python.exe train.py        # fine-tune, models/best.pt

License

Data, CC-BY-4.0 (Simula): training on Kvasir-SEG. Demo clip from Hyper-Kvasir.

---

_{Built and engineered by Iker Pacheco Herrero, Biomedical Engineer.}