Performance Improvement of Automated Segmentation of Multiple Organs and Tissue Regions in Torso CT images:

Training and Performance Evaluation of CNN and Transformer using Multi-Directional Cross-Sectional Images

Abstract

The development of computer systems to assist radiologists to accomplish medical image diagnosis requires recognition of target organ regions on images. However, fully automated organ segmentation to medical images is desirable, as manual annotation of pixel-by-pixel target organ regions on images is tedious and error-prone. Recent works have focused on two types of segmentation methods. One is CNN, which tends to capture local features, and the other is Transformer, which tends to capture global context. In this study, we aim to improve the performance of CNN networks by integrating with Transformer for multi-organ and tissue region segmentations, which has not been previously explored. Previous studies used three orthogonal cross-sections, but this study uses more sections in non-orthogonal directions to validate their use. We also use pre-trained models to validate the variability of organ region extraction accuracy. We validated the accuracy of organ extraction using multiple cross-sectional orientations. The proposed method improved the extraction accuracy by 3.2% in terms of the Jaccard coefficient compared to the baseline using axial sections only.