UADC: Uncertainty-Aware Diverse Co-Training for Semi-Supervised Medical Image Segmentation

Abstract

In semi-supervised segmentation, deep co-training has emerged as a promising paradigm by utilizing prediction agreement and unlabelled-data-driven supervisory signals. Despite its effectiveness, insufficient diversity between the collaborating models often causes them to become highly consistent at an early stage, weakening the benefit of mutual learning and eventually making the framework behave similarly to self-training. In an effort to postpone this premature convergence, different initial perturbations were incorporated, but unfortunately, this resulted in a decrease in the quality of the model's pseudo-labels, which further deteriorated the model's performance. To tackle this issue, we propose an Uncertainty Aware and Diverse Co-training model (UADC) which is in line with the essential postulates of co-training. The model features a novel co-training framework working in hybrid spatial-frequency domain. Specifically, after revisiting the basic hypotheses of co-training, we devise a novel multi-view training approach that addresses the issues of dependency and role inequality inherent in traditional methods. Targeting the misconceptions in previous co-training architectures, we incorporate Segformer and frequency domain training to decouple and thereby counteract the tendency towards self-training degeneration during later training stages. Furthermore, to enhance the quality of pseudo-labels, we delve into the impact of uncertainty on model performance. We alter the structure of the model, designing a method grounded in the epistemic uncertainty of the segmentation model to optimize the quality of pseudo labels. We conduct extensive experiments on three public medical datasets of ISIC, Kvasir and CVC-ClinicDB, and compared our model with other state-of-the-art approaches, demonstrating its effectiveness and competitive performance.