Anatomy-guided Multi-View Fusion Framework for Abdominal CT Multi-Organ Segmentation

Zhongwei Yang, Haopeng Kuang, Xukun Zhang, Yang Liu, Peng Zhai, Lubin Chen, Lihua Zhang

Abstract: Multi-organ segmentation from abdominal CT images plays a vital role in clinical practice. However, due to the low contrast of soft tissues in CT images and the significant differences in the shape and appearance of organs, this is a challenging task. In this paper, we propose a two-stage framework based on multi-view fusion to solve this challenge. Specifically, the first stage is to segment the organs in the original abdominal CT image quickly. Based on this, we introduce anatomical knowledge to robustly extract the image region of a single organ. Then, inspired by the clinician's image reading, the organ image blocks from three views are used as the input of the second stage network, and the features from different views are adaptively fused to output accurate segmentation results. We conduct extensive experiments on a public CT dataset, and the experimental results show that our method is accurate and robust to this challenging segmentation task.