Automatic Segmentation of Hepatic and Portal Veins using SwinUNETR and Multi-Task Learning

Abstract

Accurate segmentation of the hepatic and portal veins plays a vital role in planning and guiding liver surgeries. This paper presents a novel approach using multi-task learning(MTL) within SwinUNETR architecture to segment both the hepatic and portal veins at the same time. The MTL framework is trained using Dice-Focal loss and designed with two decoder branches each for segmenting the hepatic and portal vein branches. The results from the clinical CT data have shown significant performance for both the hepatic and portal veins compared to the base model (SwinUNETR), especially at the early stages of training. Notably, the MTL model achieved statistically significant results for the portal vein segmentation compared to the base model after 100 epochs. Our proposed MTL model (SwinUNETR_MTL) achieved a dice similarity coefficient (DSC) of 0.8404 for the hepatic vein and a DSC of 0.8120 for the portal vein segmentation. Our findings suggest that the MTL model attains faster convergence and increased segmentation accuracy, making it a promising approach for segmenting complex structures in the clinical setup.