WaveU3S: A Lightweight Wavelet Dual-Attention Unet for 3D Medical Image Segmentation

Tianzhao Zhong, Huaishui Yang, Jihao Li, Mengye Lyu, Shaojun Liu

Abstract: Medical image segmentation is a key step in medical image analysis. Compared with the convolutional network where the receptive field is usually limited, the transformer model is able to extract the long-range dependencies between various locations and thus suitable for this task. However, for 3D medical image segmentation, the computation cost increases dramatically with the spatial resolution. In this paper, we propose a lightweight hybrid model (WaveU3S) for 3D medical segmentation based on Unet framework and 3D wavelet transform with dual attention. With 3D wavelet transform, the input features are compressed to half spatial resolution without losing information, reducing the computation burden while maintaining the performance. Experiments on FLARE22 and ACDC datasets demonstrate that the proposed method can produce satisfied segmentation result at a low cost.

External IDs:dblp:conf/isbi/ZhongYLLL24