PAC-UNet: Parallel Dual Self-Attention with Convolution for Meniscal MRI Image Segmentation

Abstract

Meniscal tears are common sports-injury disorders, and thus accurate segmentation of the meniscus is a worthwhile means of medical diagnosis. However, due to the limited research on image segmentation of the meniscus, most previous works have poor segmentation effects on the meniscus, which affects the adoption of assisted medical diagnosis of meniscus. To address this issue, we propose an improved neural network model, termed as the PAC-UNet model, which introduces a parallel dual self-attention with depth-wise convolution for the meniscus image segmentation. Firstly, in the encoder stage, the parallel block is used to eliminate the negative impact of weak modeling ability caused by weight sharing on shared dimensions, so as to better learn the semantic information. Then, in the decoder stage, the parallel block and patch expansion are used to restore the predicted results of meniscus segmentation by performing self-attention and up-sampling on the features. By analyzing the outcomes of experiments on the meniscus dataset, which were collected from the hospital, it is found that the proposed network performed better than other medical image segmentation networks, where the average Dice Similarity Coefficient (DSC), Intersection Over Union (IOU), precision, and Hausdorff_95 of the medial meniscus and lateral meniscus are up to 89.01%, 80.81%, 90.5%, and 2.43, respectively.