CSDA-MOSNet: Circular Sparse Dual Attention Multi Sparse Transformer for Multi Organ Segmentation

Abstract

Intelligent multi-organ segmentation is a decisive task in computer aided diagnosis for diagnosing disease in body regions. Numerous Deep Learning (DL) models have been widely utilized for multi-organ segmentation tasks whereas they hindered with locating the organs with varied sizes and sparse boundaries. To this end, we design a novel model named Circular Sparse Dual Attention Multi Sparse Transformer based Multi Organ Segmentation Network (CSDA-MOSNet). For testing and training, we collect data from the two benchmark multi organ datasets such as Beyond the Cranial Vault (BTCV), and Abdominal Multi-Organ Segmentation Challenge (AMOS). Different from the state of art works, the acquired images from the two datasets are simultaneously fed into our model from two ends individually. Initially, both of images are processed by the Circular Dilated Convolution Network (CDCN) for capturing the long-range dependencies and boundary details. Aftermath, the captured details are carried out by zig-zag connected Multi Sparse Transformer Network (MSTN) for extracting long range temporal details which results in global context feature map. At last, Dual Branch Spatial Attention Module (DBSAM) is utilized for processing the global context feature map to obtaining final segmented outcome. The proposed research is examined using two benchmark datasets in terms of Dice Similarity Score (DISC) with 98.96% and 99.70% respectively. The results shows that the proposed CSDA-MOSNet accomplishes improved performance over existing works.