Addressing data annotation scarcity in Brain Tumor Segmentation on 3D MRI scan Using a Semi-Supervised Teacher-Student Framework

TL;DR

This work tackles annotation scarcity in brain tumor segmentation from 3D MRI by introducing a semi-supervised teacher–student framework. The teacher emits probabilistic masks with per-voxel uncertainty, enabling a confidence-guided, progressive sampling curriculum for the student, who is trained with a dual-loss that emphasizes high-confidence regions and unlearns low-confidence ones; a feedback loop refines pseudo-labels via teacher–student agreement. The architecture, TransASPP-UNet, combines ASPP with a Transformer bottleneck to capture multi-scale context and uncertainty; the approach achieves strong data efficiency on BraTS 2021, with the student surpassing the teacher on key tumor subregions (NCR/NET and Edema) and recovering the Enhancing class. Practically, this framework reduces reliance on large annotated datasets while improving robustness to domain shift across institutions, aided by uncertainty estimates for clinical decision support. Future work includes external multi-site validation and exploring active learning and federated training to further reduce labeling burden.

Abstract

Accurate brain tumor segmentation from MRI is limited by expensive annotations and data heterogeneity across scanners and sites. We propose a semi-supervised teacher-student framework that combines an uncertainty-aware pseudo-labeling teacher with a progressive, confidence-based curriculum for the student. The teacher produces probabilistic masks and per-pixel uncertainty; unlabeled scans are ranked by image-level confidence and introduced in stages, while a dual-loss objective trains the student to learn from high-confidence regions and unlearn low-confidence ones. Agreement-based refinement further improves pseudo-label quality. On BraTS 2021, validation DSC increased from 0.393 (10% data) to 0.872 (100%), with the largest gains in early stages, demonstrating data efficiency. The teacher reached a validation DSC of 0.922, and the student surpassed the teacher on tumor subregions (e.g., NCR/NET 0.797 and Edema 0.980); notably, the student recovered the Enhancing class (DSC 0.620) where the teacher failed. These results show that confidence-driven curricula and selective unlearning provide robust segmentation under limited supervision and noisy pseudo-labels.

Figures (10)

• Figure 1: Progressive pseudo-labeling framework for semi-supervised brain tumor segmentation. The Teacher U-Net is first trained on labeled MRI data to generate probabilistic segmentation masks and uncertainty estimates. Pseudo-labels are then generated for unlabeled target-domain images and ranked by confidence scores . A progressive sampling strategy selects high-confidence samples to train the Student model in stages, with feedback refinement comparing teacher-student predictions to improve pseudo-label quality iteratively.
• Figure 2: Overlay of tumor subregions (enhancing tumor, tumor core, and edema) on T1, T1ce, T2, and FLAIR MRI sequences.
• Figure 3: Schematic of the proposed U-Net–based segmentation model integrating Atrous Spatial Pyramid Pooling (ASPP) and attention transformer to enhance multiscale feature learning and focus on salient regions.
• Figure 4: Learning Curves of the Teacher Model: Training vs. Validation Performance (Accuracy, Loss, Dice Coefficient over Epochs).
• Figure 5: Progressive training dynamics across six curriculum stages (red dashed lines) over 60 epochs. Training Dice (top-left) improves from 0.3 to 0.87; validation Dice (top-right) increases from 0.393 to 0.872 with temporary Stage 5 dip. Training loss (bottom-left) decreases from 2.2 to $<$0.8; validation loss (bottom-right) declines from 1.8 to 0.88 with transient Stage 5 spike. Curves demonstrate effective progressive learning with diminishing later-stage gains.