RAIL: Region-Aware Instructive Learning for Semi-Supervised Tooth Segmentation in CBCT

TL;DR

RAIL tackles the challenge of limited annotations in CBCT tooth segmentation by introducing Region-Aware Instructive Learning, a dual-group, dual-student Mean Teacher framework that alternates training to promote inter-group knowledge transfer. It couples Disagreement-Focused Supervision (DFS), which concentrates learning on structurally ambiguous or mislabeled regions, with Confidence-Aware Learning (CAL), which stabilizes pseudo-labels by down-weighting low-confidence regions. The method formalizes region-aware losses within a progressive mean-teacher setting, leveraging DiffMask, MisMask, and DivMask to guide supervision and pseudo-label refinement. evaluated on four CBCT datasets, RAIL consistently outperforms state-of-the-art SSL methods under sparse supervision, demonstrating improved Dice and Jaccard scores as well as reduced boundary errors, with practical implications for efficient clinical segmentation. The approach offers robust performance and shows promise for broader application in semi-supervised medical image segmentation tasks.

Abstract

Semi-supervised learning has become a compelling approach for 3D tooth segmentation from CBCT scans, where labeled data is minimal. However, existing methods still face two persistent challenges: limited corrective supervision in structurally ambiguous or mislabeled regions during supervised training and performance degradation caused by unreliable pseudo-labels on unlabeled data. To address these problems, we propose Region-Aware Instructive Learning (RAIL), a dual-group dual-student, semi-supervised framework. Each group contains two student models guided by a shared teacher network. By alternating training between the two groups, RAIL promotes intergroup knowledge transfer and collaborative region-aware instruction while reducing overfitting to the characteristics of any single model. Specifically, RAIL introduces two instructive mechanisms. Disagreement-Focused Supervision (DFS) Controller improves supervised learning by instructing predictions only within areas where student outputs diverge from both ground truth and the best student, thereby concentrating supervision on structurally ambiguous or mislabeled areas. In the unsupervised phase, Confidence-Aware Learning (CAL) Modulator reinforces agreement in regions with high model certainty while reducing the effect of low-confidence predictions during training. This helps prevent our model from learning unstable patterns and improves the overall reliability of pseudo-labels. Extensive experiments on four CBCT tooth segmentation datasets show that RAIL surpasses state-of-the-art methods under limited annotation. Our code will be available at https://github.com/Tournesol-Saturday/RAIL.

Figures (3)

• Figure 1: Pipeline of our Region-Aware Instructive Learning (RAIL) framework in Mean Teacher architecture. The total loss function for every student network in the training phase includes supervised losses $\mathcal{L}_s$, $\mathcal{L}_{DFS}$, and unsupervised losses $\mathcal{L}_U$, $\mathcal{L}_T$, $\mathcal{L}_{CAL}$.
• Figure 2: 2D segmentation visualization of different semi-supervised methods on FDDI+ (first line), FDDI-E (second line), 3D CBCT Tooth (third line) and CTooth (last line) dataset under 14%, 10%, 10% and 10% labeled, respectively.
• Figure 3: 3D segmentation visualization of different semi-supervised methods on FDDI+ (first line), FDDI-E (second line), 3D CBCT Tooth (third line) and CTooth (last line) dataset under 14%, 10%, 10% and 10% labeled, respectively.