RA V-Net: deep learning network for automated liver segmentation

Abstract

Objective. Segmenting liver from CT images is the first step for doctors to diagnose a patient’s disease. Processing medical images with deep learning models has become a current research trend. Although it can automate segmenting region of interest of medical images, the inability to achieve the required segmentation accuracy is an urgent problem to be solved. Approach. Residual Attention V-Net (RA V-Net) based on U-Net is proposed to improve the performance of medical image segmentation. Composite Original Feature Residual Module is proposed to achieve a higher level of image feature extraction capability and prevent gradient disappearance or explosion. Attention Recovery Module is proposed to add spatial attention to the model. Channel Attention Module is introduced to extract relevant channels with dependencies and strengthen them by matrix dot product. Main results. Through test, evaluation index has improved significantly. Lits2017 and 3Dircadb are chosen as our experimental datasets. On the Dice Similarity Coefficient, RA V-Net exceeds U-Net 0.1107 in Lits2017, and 0.0754 in 3Dircadb. On the Jaccard Similarity Coefficient, RA V-Net exceeds U-Net 0.1214 in Lits2017, and 0.13 in 3Dircadb. Significance. Combined with all the innovations, the model performs brightly in liver segmentation without clear over-segmentation and under-segmentation. The edges of organs are sharpened considerably with high precision. The model we proposed provides a reliable basis for the surgeon to design the surgical plans.