LK-UNet: Large Kernel Design for 3D Medical Image Segmentation

Jiang Shang, Sifan Zhou

Abstract

Recently, the medical image segmentation have made rapid progress. Specifically, the precision of medical image segmentation play a pivotal role in the realm of disease diagnosis and treatment. Therefore, it is vital to improve the segmentation performance. Generally, Transformer-based methods exhibit superior performance compared to CNN-based methods on 3D medical image segmentation tasks due to their inherent capability to capture global-aware context. However, the existing transformer-based models are still unsatisfactory in accuracy. In this paper, we propose a novel fully convolution architecture for medical image segmentation tasks, called LK-UNet. Specifically, the key of LK-UNet lies in its incorporation of a large kernel module, which can achieve comparable receptive fields to transformer module. Besides, we introduce Depth-wise Convolution Layer (DCL) and Point-wise Convolution Layer (PCL) to substitute the vanilla convolution layer to reduce the number of model parameters and enhance the feature representation. Extensive experiment shows that our method achieves state-of-the-art performance on the public BTCV dataset, which even outperforms hybrid transformer-based and CNN-based networks.
BibTeX
@inproceedings{shang2024lk,
  author = {Jiang Shang and Sifan Zhou},
  title = {{LK-UNet: Large Kernel Design for 3D Medical Image Segmentation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10446818},
}

Loading related content...

Figure