Dual-Cross-Attention for Medical Image Segmentation

Official Pytorch implementation of the paper Dual-Cross-Attention for Medical Image Segmentation

We propose Dual Cross-Attention (DCA), a simple yet effective attention module that is able to enhance skip-connections in U-Net-based architectures for medical image segmentation. Our proposed module addresses the semantic gap between encoder and decoder features by sequentially capturing channel and spatial dependencies across multi-scale encoder features.

Benchmark Results

We test our method on GlaS, MoNuSeg, Kvasir-Seg, CVC-ClinicDB and Synapse datasets.

	Params	GlaS	GlaS	MoNuSeg	MoNuSeg	CVC-ClinicDB	CVC-ClinicDB	Kvasir-Seg	Kvasir-Seg	Synapse	Synapse
		DSC	IoU	DSC	IoU	DSC	IoU	DSC	IoU	DSC	IoU
U-net	8.64M	0.8887	0.7998	0.7714	0.6279	0.8963	0.8143	0.8299	0.7101	0.7855	0.6737
U-Net(DCA)	8.75M	0.8966	0.8129	0.7813	0.6411	0.8953	0.8128	0.8403	0.7253	0.7898	0.6797
ResUnet++	13.1M	0.8543	0.7462	0.7568	0.6087	0.8946	0.8114	0.8226	0.6993	0.7591	0.6461
ResUnet++(DCA)	13.1M	0.8735	0.7756	0.7740	0.6313	0.9019	0.8232	0.8207	0.6974	0.7735	0.6643
MultiResUnet	7.24M	0.8899	0.8018	0.7699	0.6259	0.8952	0.8135	0.8134	0.6866	0.7812	0.6730
MultiResUnet (DCA)	7.35M	0.8886	0.7998	0.7852	0.6463	0.8995	0.8191	0.8232	0.7000	0.7950	0.6865
R2Unet	9.78M	0.8516	0.7426	0.7820	0.6420	0.8812	0.7888	0.8107	0.6828	0.7586	0.6394
R2Unet(DCA)	9.89M	0.8721	0.7737	0.7852	0.6464	0.8839	0.7928	0.8219	0.6989	0.7590	0.6485
V-Net	35.97M	0.8878	0.7985	0.7479	0.5974	0.8809	0.7902	0.8079	0.6807	0.7927	0.6858
V-Net(DCA)	36.08M	0.8903	0.8027	0.7753	0.6331	0.8946	0.8107	0.8192	0.6953	0.7958	0.6900
DoubleUnet	29.68M	0.8907	0.8030	0.7716	0.6282	0.9020	0.8235	0.8440	0.7308	0.7976	0.6931
DoubleUnet (DCA)	30.68M	0.8990	0.8168	0.7950	0.6597	0.9086	0.8347	0.8516	0.7434	0.8022	0.6980

Citations

If you find this repo useful, please cite:

@misc{ates2023dual,
      title={Dual Cross-Attention for Medical Image Segmentation}, 
      author={Gorkem Can Ates and Prasoon Mohan and Emrah Celik},
      year={2023},
      eprint={2303.17696},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}

Contact

Gorkem Can Ates (g.canates@gmail.com)