Multi-task Learning Network for CT Whole Heart Segmentation

Abstract

Accurate whole heart segmentation from CT is important for the adjuvant treatment of cardiovascular diseases. Considering the complex anatomical structure of the heart, single task is difficult to provide abundant segmentation information. In this work, we propose a novel multi-task learning network for CT whole heart segmentation. The framework contains two task branches, coarse-grained segmentation task and fine-grained segmentation task. To learn more useful information, two decoders are designed independently, and they predict fine-grained segmentation results and coarse-grained segmentation results, separately. Coarse-grained segmentation as an auxiliary task, its label contains the prior information of the anatomical structure to assist the learning of fine-grained segmentation tasks. Meanwhile, a shared encoder is used to extract features to realize knowledge reuse for two tasks. Besides, according to the characteristics of CT images, we design a set of suitable image pre-processing method. We evaluated our approach on the MM-WHS CT dataset, the experimental results show that our method is superior to other methods, verifying the effectiveness of the proposed method.