Alternate Diverse Teaching for Semi-supervised Medical Image Segmentation

TL;DR

Semi-supervised medical image segmentation often suffers from confirmation bias when using pseudo-labels. The authors propose AD-MT, an alternate diverse teaching framework with a single trainable student and two non-trainable EMA teachers updated in Random Periodic Alternate fashion, plus a Conflict-Combating Module that learns from both agreements and disagreements via entropy-based ensembling. The two novel components, RPA and CCM, enlarge supervision diversity and explicitly leverage conflicting predictions through entropy-based ensembling, achieving state-of-the-art results on 2D and 3D benchmarks without extra pre-training. This approach improves segmentation accuracy under limited labeled data and offers a practical, scalable strategy for leveraging unlabeled medical images.

Abstract

Semi-supervised medical image segmentation studies have shown promise in training models with limited labeled data. However, current dominant teacher-student based approaches can suffer from the confirmation bias. To address this challenge, we propose AD-MT, an alternate diverse teaching approach in a teacher-student framework. It involves a single student model and two non-trainable teacher models that are momentum-updated periodically and randomly in an alternate fashion. To mitigate the confirmation bias from the diverse supervision, the core of AD-MT lies in two proposed modules: the Random Periodic Alternate (RPA) Updating Module and the Conflict-Combating Module (CCM). The RPA schedules the alternating diverse updating process with complementary data batches, distinct data augmentation, and random switching periods to encourage diverse reasoning from different teaching perspectives. The CCM employs an entropy-based ensembling strategy to encourage the model to learn from both the consistent and conflicting predictions between the teachers. Experimental results demonstrate the effectiveness and superiority of our AD-MT on the 2D and 3D medical segmentation benchmarks across various semi-supervised settings.

Figures (5)

• Figure 1: Frameworks of different SSMIS methods. Crucial distinctions arise from how the unlabeled data is leveraged. a) the plain teacher-student framework, b) a two-student co-training paradigm that enforces mutual learning, c) a two-teacher ensemble framework where two differently initialized and updated teacher models supervise the training of the student model in a average manner, d) our proposed alternate diverse mean-teacher (AD-MT) framework. Our framework involves two teacher models that are updated periodically and randomly, using complementary unlabeled data batches, distinct data augmentation strategies, and randomized switchable periods, to enlarge their disagreement. Additionally, our Conflict-Combating Module (CCM) encourages the student model to learn from the conflict predictions of the teacher models rather than dropping conflicts directly. "sg" denotes "stop gradient".
• Figure 2: We compare our proposed AD-MT with recent SSIMS methods in terms of the Dice score on 2D ACDC, 3D LA and Pancreas datasets with 3, 4, and 6 labeled instances, respectively. Our end-to-end AD-MT can consistently outperform the current state-of-the-art BCP (which requires an additional pre-training stage).
• Figure 3: The diagram of our proposed AD-MT. Our method consists of two main modules: the Random Periodic Alternate Updating Module (RPA) and the Conflict-combating Module (CCM). Specifically, two teacher models T1 and T2 are updated in turn periodically and randomly. At each iteration, only one certain teacher model $\hbox{T}_m$, ($m=1,2$) will be updated, using complementary unlabeled data batches and different strong data augmentation strategies $A_m$ accordingly. Furthermore, the switchable period of two teachers is randomly generated by the RPA module, aiming to increase the disagreement between the two teacher models. Meanwhile, the CCM module separates the consistent and conflicting predictions of two teacher models, and encourage the model to learning from instead of dropping the conflicts. $q_i^{s}, q_i^{t_1}, q_i^{t_2}$ represent the generated pseudo-labels from the student and two teachers models, respectively.
• Figure 4: Impact of different alternating periodic updating strategies in the RPA with varying values of $\mathcal{T}_{max}$ on the ACDC with 5% labeled data. By default, we adopt the random switching periods and set $\mathcal{T}_{max}$ as the half-epoch iterations.
• Figure 5: Qualitative results from the 3D LA (top 2 rows) and 2D ACDC (bottom 2 rows). a) the ground-truth, b) UA-MT, c) MC-Net, d) SS-Net, and e) AD-MT.