📌 Official PyTorch implementation of GroupKAN: Rethinking Nonlinearity with Grouped Spline-based KAN Modeling for Efficient Medical Image Segmentation

[Paper (arXiv)] [Model Zoo]

GroupKAN revisits the design of nonlinear representations in medical image segmentation by introducing grouped spline-based functional modeling.
It integrates two novel components — Grouped KAN Transform (GKT) and Grouped KAN Activation (GKA) — into a lightweight and interpretable U-shaped backbone.

Figure 1. Overall architecture of GroupKAN. The Grouped KAN Transform and Activation efficiently model intra-group dependencies with learnable spline functions.

• 🔹 Efficient Grouped Nonlinearity:
  Splits features into groups and applies spline-based transformations, reducing full-channel complexity from O(C²) to O(C²/G).

• 🔹 Interpretable Functional Mapping:
  Each group learns localized nonlinearities, providing clearer feature attribution in medical segmentation.

• 🔹 Compact Yet Powerful:
  Achieves +1.1% IoU gain over U-KAN with 47.6% fewer parameters (3.02M vs. 6.35M).

Figure 2. Accuracy–complexity comparison across models. GroupKAN achieves better accuracy with nearly half the parameters of U-KAN.

Figure 3. GroupKAN produces more accurate and sharper segmentation results across BUSI, GlaS, and CVC datasets.

GroupKAN exhibits improved activation–mask alignment and interpretable feature distributions compared with U-KAN.

Figure 4. Activation maps align more closely with anatomical boundaries, showing improved interpretability.

git clone https://github.com/GroupKAN/GroupKAN.git
cd GroupKAN
conda create -n groupkan python=3.10
conda activate groupkan
pip install -r requirements.txt

python train.py --dataset busi --input_size 256 --epochs 400 --batch_size 8

python val.py --weights checkpoints/groupkan_best.pth --dataset glas

• Release segmentation code
• Release pre-trained checkpoints
• Add demo and visualization scripts

This work is inspired by U-KAN
and the Kolmogorov–Arnold Network (KAN).
We thank the open-source community for their valuable contributions.

For questions and collaborations, please reach out to
Guojie Li ([email protected]).