Abstract
Diabetic retinopathy lesion segmentation in fundus images is challenging, due to the existence of the scale variation, similar lesion appearances, and foreground-background imbalance. The previous semantic segmentation approaches primarily address scale change without sufficiently tackling the other two specific challenges of fundus image segmentation. To solve the aforementioned problems, this paper presents a novel approach, named the HFM-Seg network, aimed to enhance lesion segmentation using a hybrid foreground modeling network. It consists the relation-based and the optimization-based foreground modeling. The relation-based part of the HFM-Seg network uses the cene-context enhancement module to learn the foreground-background relationship to associate the foreground-related multi-scale context, so as to enhance the discrimination of foreground features. And the relation-based part of the HFM-Seg network adopts the adaptive foreground enhancement optimization, which pays attention to both foreground samples and background difficult samples in the training process, achieving balanced optimization. The mAUC values of HFM-Seg on the IDRiD, DDR, and E-optha datasets surpass the average of six other advanced methods by 7.76%, 15.09%, and 13.51% respectively, demonstrating HFM-Seg's effectiveness.

This publication has 13 references indexed in Scilit: