Cross-pyramid consistency regularization for semi-supervised medical image segmentation

TL;DR

The paper addresses the challenge of semi-supervised medical image segmentation with limited labeled data by introducing CPCR, a network-level consistency regularization framework built on DBPNet. DBPNet uses two perturbed decoders to generate both final and auxiliary pyramid predictions, while CPCR imposes cross-pyramid, cross-decoder consistency and uncertainty minimization to leverage unlabeled data. The approach demonstrates superior performance over several state-of-the-art SSL methods on the ACDC dataset with only 10% labeled data, especially excelling in boundary-related metrics. This work advances practical SSL for medical imaging by combining multi-scale auxiliary supervision with robust knowledge distillation across decoders, improving segmentation accuracy and confidence.

Abstract

Semi-supervised learning (SSL) enables training of powerful models with the assumption of limited, carefully labelled data and a large amount of unlabeled data to support the learning. In this paper, we propose a hybrid consistency learning approach to effectively exploit unlabeled data for semi-supervised medical image segmentation by leveraging Cross-Pyramid Consistency Regularization (CPCR) between two decoders. First, we design a hybrid Dual Branch Pyramid Network (DBPNet), consisting of an encoder and two decoders that differ slightly, each producing a pyramid of perturbed auxiliary predictions across multiple resolution scales. Second, we present a learning strategy for this network named CPCR that combines existing consistency learning and uncertainty minimization approaches on the main output predictions of decoders with our novel regularization term. More specifically, in this term, we extend the soft-labeling setting to pyramid predictions across decoders to support knowledge distillation in deep hierarchical features. Experimental results show that DBPNet with CPCR outperforms five state-of-the-art self-supervised learning methods and has comparable performance with recent ones on a public benchmark dataset.

Figures (1)

• Figure 1: Overview of the proposed DBPNet and its learning framework via Cross Pyramid Consistency Regularization (CPCR). Note that we did not depicted average prediction-based uncertainty minimization, as it is not the contribution of our work.