---
license: apache-2.0
library_name: pytorch
pipeline_tag: image-segmentation
tags:
- medical-imaging
- polyp-segmentation
- colonoscopy
- kolmogorov-arnold-network
- kan
- illumination-robust
- miccai-2026
datasets:
- Kvasir-Sessile
- CVC-ColonDB
- ETIS-LaribPolypDB
- PolypGen-C6
metrics:
- dice
- iou
---

# KAN-AINet: Kolmogorov–Arnold Network with Adaptive Illumination Modulation for Generalizable Polyp Segmentation

## Model Description

KAN-AINet is a polyp segmentation architecture that leverages **Kolmogorov–Arnold Networks (KAN)** for adaptive illumination modulation and boundary-aware attention (MICCAI 2026). Unlike standard neural networks that use fixed activation functions, KAN learns optimal per-task activation functions, enabling more expressive feature transformations for challenging colonoscopy images.

It introduces two KAN-based modules:

- **KAN-IMM (Illumination Modulation Module):** adaptive illumination modulation that improves robustness under dark, medium, and bright conditions (largest gain under extreme lighting, p = 0.037).
- **KAN-BAM (Boundary Attention Module):** multi-scale edge-aware attention (3×3, 5×5, 7×7 receptive fields) that differentiates true polyp boundaries from illumination artifacts.

KAN-based activation functions are directly visualizable, providing interpretability into how the network adapts its feature transformations for segmentation.

## Training Details

- **Architecture:** KAN-AINet (KAN-IMM + KAN-BAM modules)
- **Training dataset:** Same as ESPNet, available from the [ESPNet Polyp Segmentation repository](https://github.com/Raneem-MT/ESPNet_Polyp_Segmentation)
- **Configuration:** Default settings or modifiable hyperparameters in `config.py`; trained via `train_threshold.py`
- **External benchmarks (unseen):** Kvasir-Sessile, CVC-ColonDB, ETIS-LaribPolypDB, PolypGen-C6

## Model Performance

Evaluated on unseen external validation datasets with segmentation-accuracy and boundary-based metrics (mDice, mIoU, Sα, Fβ^w, MAE, HD95, ASD, Precision, Recall, Specificity):

| Metric | Result |
|---|---|
| mDice | +4.99% over prior SOTA |
| mIoU | +5.07% over prior SOTA |
| HD95 (KAN-BAM) | −33.7% vs. variant without KAN |
| ASD (KAN-BAM) | −42.95% vs. variant without KAN |
| Prediction variance (Brown–Forsythe) | ratio 0.68, p < 0.001 |

- Improves mDice by 4.99% and mIoU by 5.07% over prior SOTA on external benchmarks
- KAN-IMM yields the largest gains under extreme lighting (p = 0.037)
- KAN-BAM reduces HD95 and ASD by 33.7% and 42.95% over the no-KAN variant
- Brown–Forsythe testing confirms significantly lower prediction variance across all illumination conditions, demonstrating stable, trustworthy performance

> Absolute per-dataset scores are reported in the comparison table in the [source repository](https://github.com/biodatlab/kanainet).

## Download & Use

Download the checkpoint from the [Hugging Face repo](https://huggingface.co/biodatlab/kan-ainet/tree/main):

```python
from huggingface_hub import hf_hub_download
from models.kan_acnet import KANACNet, visualize

model_path = hf_hub_download(repo_id="biodatlab/kan-ainet", filename="model.pth")

kan  = KANACNet(model_path)   # loads weights, eval mode, auto GPU/CPU
mask = kan("test.jpg")        # numpy uint8 array
visualize("test.jpg", mask)   # displays the result
```

`KANACNet` comes from the [source repo](https://github.com/biodatlab/kanainet) — clone it and `pip install -r requirements.txt` first.

## Intended Use

- Research on illumination-robust, generalizable polyp segmentation in colonoscopy images.
- Benchmarking against polyp segmentation baselines on external/unseen datasets.
- Support for boundary-accurate and interpretable segmentation in colonoscopy analysis pipelines.

## Limitations

- Research model — **not a medical device**; not for clinical diagnosis, screening, or treatment decisions.
- Trained on the ESPNet polyp segmentation data; performance on imaging modalities, scopes, or populations outside the evaluated benchmarks is not characterized.
- Outputs require expert clinical review before any patient-facing use.
- As with any deep learning system, risks include errors and domain shifts under conditions unlike the training/evaluation data.

## Acknowledgments

Developed by the [Biomedical and Data Lab (biodatlab)](https://biodatlab.github.io/) with the collaboartion with [Diagnostic Intelligence Group (DIG)](https://github.com/Lab-DIG) at University of Alabama at Birmingham. We acknowledge the broader open-source community whose tools and prior work on KAN, polyp segmentation, and the ESPNet dataset made this project possible.

**Code, training, and full results:** https://github.com/biodatlab/kanainet