Automatic Segmentation of Kidney Volume Using Multi-Module Hybrid Based U-Shape in Polycystic Kidney Disease

Abstract

Polycystic kidney disease (Autosomal Dominant Polycystic Kidney Disease, ADPKD) is the most common genetic disease of the kidney, and the measurement of Total Kidney Volume (TKV) in clinical research of this disease is essential to study the progression of ADPKD. At present, the volume segmentation of polycystic kidneys mainly relies on doctors to manually outline the kidney boundary on the radiological image. This process is time-consuming, labor-intensive, inefficient, subjective, and difficult to guarantee consistency. In the research of this paper, A multi-module hybrid U-shape segmentation method is proposed (HUNet), which introduces wavelet pooling, cascade residual, and efficient multi-head self-attention into the U-shape structure. We use wavelet pooling instead of traditional down-sampling to reduce the loss of detailed features, the use of cascaded residual modules can improve the ability of model feature reuse, and the use of efficient multi-head self-attention modules can effectively capture global multi-scale information. In the decoding process of the U-shape, the corresponding loss value of each decoder will be calculated, and finally, the total loss value of the model will be obtained by weighted average. The method was trained and tested on the polycystic kidney dataset provided by Shanghai Changzheng Hospital. We automatically segmented the ADPKD in MRI images using the proposed method with a remarkably high Dice similarity coefficient relative to the manual segmentation (mean=0.915). The percentage difference between the total kidney volume values using manual and HUNet methods was only 0.4%. The proposed approach enables fast and accurate TKV measurement.