Boundary-Aware Lightweight TFAB-Net for Liver Tumor Segmentation in Contrast-Enhanced CT

Abstract

Segmentation of the tumor areas in contrast-enhanced computed tomography (CT) images is a vital process involved in diagnoses, therapy planning, and follow-ups; however, it is challenging because of fuzzy edges, lesions heterogeneity, and extreme imbalance of the tumor class and background one. We present a light-weighted approach called Tiny Focal Attention Boundary Network (TFAB-Net) for tumor regions segmentation. The proposed architecture makes use of light-weighted convolutional blocks along with the focal attention technique and incorporates the boundary-specific information for the joint optimization of a target area and its outline. The efficiency of the TFAB-Net was assessed on the liver tumors in the form of CT images and compared to four other networks: UNet, TinySegNet, UNetLite, and SegNetLite. The evaluation of these models was carried out according to the Dice coefficient, IoU, precision, and recall criteria. The results showed that our TFAB-Net achieved Dice score 0.8450 and IoU 0.7300 with the number of parameters equal to 0.006 million.