We have introduced a novel approach called Deformable Large Kernel Attention (D-LKA Attention) to enhance medical image segmentation. This method efficiently captures volumetric context using large convolution kernels, avoiding excessive computational demands. D-LKA Attention also benefits from deformable convolutions to adapt to diverse data patterns. We've created both 2D and 3D versions, with the 3D version excelling in cross-depth data understanding. This forms the basis of our new hierarchical Vision Transformer architecture, the D-LKA Net, which outperforms existing methods on popular medical segmentation datasets (Synapse, NIH Pancreas, and Skin lesion).

Reza Azad, Leon Niggemeier, Michael Hüttemann, Amirhossein Kazerouni, Ehsan Khodapanah Aghdam, Yury Velichko, Ulas Bagci and Dorit Merhof

• 24.10.2023 | Accepted in WACV 2024! 🥳

• 16.10.2023 | Release of 2D, 3D Synapse weights and 3D Pancreas weights.

State-of-the-art comparison on the abdominal multi-organ Synapse dataset for 2D methods. The model complexity and performance (DSC, HD95) are shown for all models. The proposed 2D D-LKA Net achieves superior segmentation performance. Abbreviations stand for: Spl: spleen, RKid: right kidney, LKid: left kidney, Gal: gallbladder, Liv: liver, Sto: stomach, Aor: aorta, Pan: pancreas. Best results are shown in blue, second best in red.

State-of-the-art comparison on the abdominal multi-organ Synapse dataset for 3D methods. The model complexity and performance (DSC, HD95) are shown for all models. The proposed 3D D-LKA Net achieves superior segmentation performance. Our model also is considerably small with the lowest number of parameters. Abbreviations stand for: Spl: spleen, RKid: right kidney, LKid: left kidney, Gal: gallbladder, Liv: liver, Sto: stomach, Aor: aorta, Pan: pancreas. Best results are shown in blue, second best in red.

While the 2D version achieves great segmentation results in comparison to other 2D models, the main limitation is the lack of inter-slice connections. Here, the 3D models achieve favorable segmentations.

For detailed instructions for the 2D methods, please refer to the Readme in the 2D folder.

For detailed instructions for the 3D methods, please refer to the Readme in the 3D folder.

This repository is built based on nnFormer, UNETR++, transnorm, MCF, D3D. We thank the authors for their code repositories.

All implementations were done by Leon Niggemeier. For any query please contact us for more information.

leon.niggemeier@rwth-aachen.de

@article{azad2023beyond,
  title={Beyond Self-Attention: Deformable Large Kernel Attention for Medical Image Segmentation},
  author={Azad, Reza and Niggemeier, Leon and Huttemann, Michael and Kazerouni, Amirhossein and Aghdam, Ehsan Khodapanah and Velichko, Yury and Bagci, Ulas and Merhof, Dorit},
  journal={arXiv preprint arXiv:2309.00121},
  year={2023}
}