Angular Margin Constrained Loss for Automatic Liver Fibrosis Staging

Abstract

Automatic progression staging of liver fibrosis plays very important roles in the direct treatment and the evaluation of prognosis. In clinical site, liver biopsy is popularly used as the gold standard method of liver fibrosis staging, and has obvious drawbacks such as sampling error, heavy burden to patients and high inter-observer variability. Recently, non-invasive techniques as a diagnostic standard have attracted extensive attention. This study exploits a novel deep learning-based liver fibrosis staging framework using non-invasive MRI images. Since there exist large variance in both texture and shape of MRI liver images between patients and subtle distinctness among the progression stages of liver fibrosis, it is a challenge task for accurate progression staging of liver fibrosis. To enhance the discriminative power among the fibrosis stages with subtle difference, this study proposes to integrate angular margin penalty into the conventional softmax loss of the deep learning network, which is expected to enforce extra intra-class compactness and inter-class discrepancy simultaneously. Specifically, we explore the angular margin constrained loss in several classification neural network models such as VGG16, ResNet18, and ResNet50, and further incorporate the between-stage similarity of the training procedure to adaptively adjust the margin for boosting liver fibrosis classification performance. Experiments on the MRI image dataset provided by Shandong University, which includes three progression stages of liver fibrosis: early, middle and last stages, validate that the performance gain with the integration of the angular margin penalty are from 3% to 7% compared to the baseline models: VGG 16, ResNet18, and ResNet50.