Dual Teacher-Student Learning for Semi-supervised Medical Image Segmentation

TL;DR

Semi-supervised medical image segmentation suffers from annotation costs; this paper recasts mean teacher learning as self-paced curriculum regulation and introduces Dual Teacher-Student Learning (DTSL). DTSL uses two teacher-student groups with distinct architectures and a Consensus Label Generator (CLG) that constructs pseudo-labels from the agreement between a temporally averaged in-group teacher and a cross-group student, formalized through a pace loss $\mathcal{L}_{\mathrm{pace}} = \alpha \mathcal{L}_{\mathrm{semi}} + \beta \mathcal{L}_{\mathrm{URL}}$. The CLG relies on pixel-wise Jensen-Shannon divergence to separate easy and difficult regions, enabling curriculum-like learning from simple to complex structures. Across four benchmark datasets, DTSL achieves state-of-the-art results, and in three of four cases, semi-supervised learning with limited labels outperforms fully supervised baselines, demonstrating the efficacy of explicit self-paced curriculum via cross-signal consensus.

Abstract

Semi-supervised learning reduces the costly manual annotation burden in medical image segmentation. A popular approach is the mean teacher (MT) strategy, which applies consistency regularization using a temporally averaged teacher model. In this work, the MT strategy is reinterpreted as a form of self-paced learning in the context of supervised learning, where agreement between the teacher's predictions and the ground truth implicitly guides the model from easy to hard. Extending this insight to semi-supervised learning, we propose dual teacher-student learning (DTSL). It regulates the learning pace on unlabeled data using two signals: a temporally averaged signal from an in-group teacher and a cross-architectural signal from a student in a second, distinct model group. Specifically, a novel consensus label generator (CLG) creates the pseudo-labels from the agreement between these two signals, establishing an effective learning curriculum. Extensive experiments on four benchmark datasets demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches. Remarkably, on three of the four datasets, our semi-supervised method with limited labeled data surpasses its fully supervised counterparts, validating the effectiveness of our self-paced learning design.

Figures (3)

• Figure 1: Agreement between teacher-generated labels and ground truth labels across training iterations in a supervised setting with the MT strategy. The regions of agreement expand from the object's center to its borders as training progresses. This suggests that the MT strategy may implicitly introduce a curriculum, guiding the model to learn from easier to more complex regions.
• Figure 2: Overview of the proposed DTSL framework on an unlabeled data sample. The framework utilizes two teacher-student groups with distinct architectures. The core component, the CLG, creates a pseudo-label based on the consensus between the in-group teacher's prediction and the cross-group student's prediction. This pseudo-label guides the student model via a semi-supervised loss ($\mathcal{L}_{\mathrm{semi}}$). For regions of disagreement, a uniform regularization ($\mathcal{L}_{\mathrm{URL}}$) is applied to handle prediction uncertainty.
• Figure 3: Representative segmentation results on the ACDC dataset using the 10% labeled data. The proposed method produces segmentations that align more closely with the ground truth than competing approaches, demonstrating superior performance in capturing complex, fine-grained structures. Notably, in the red-highlighted region, only DTSL achieves an accurate segmentation.