Abstract
Training medical image segmentation models typically requires massive amounts of pixel-perfect labels, which are expensive and slow to produce. This paper introduces a dual-branch "Teacher-Student" framework that thrives on mixed supervision—using just a few fully labeled images alongside many weakly labeled ones (like scribbles). Remarkably, the study finds that the "student" branch eventually outperforms its "teacher." Evaluated on the ACDC cardiac MRI dataset, the proposed method demonstrates significant improvements in Dice Similarity Scores and boundary accuracy over existing mixed-supervision baselines.
Introduction
Deep learning has revolutionized medical imaging, but its hunger for data is a major bottleneck. In clinical settings, asking a radiologist to manually outline every pixel of an organ in hundreds of MRI scans is impractical. While researchers have explored "mixed supervision" (combining a few full labels with many weak ones), these models often underuse the weak data or fall into "trivial solutions" where the model becomes confidently wrong. This paper aims to bridge the performance gap between these data-starved models and their fully supervised counterparts by better integrating information from unlabeled or partially labeled pixels.
Methods
The authors propose a dual-branch architecture designed to decouple the learning of strong and weak features:
- The Teacher (Top Branch): Trained on the small set of fully labeled images using standard cross-entropy loss.
- The Student (Bottom Branch): Trained on the weakly labeled images.
The synergy between these branches is governed by a joint objective function:
- Knowledge Distillation ($L_{kd}$): The teacher "teaches" the student by forcing the student’s predictions on fully labeled images to match the teacher’s predictions using Kullback-Leibler (KL) divergence.
- Shannon Entropy Minimization ($L_{ent}$): To handle pixels without labels, the model minimizes Shannon entropy. This encourages the student to be highly confident in its predictions (pushing pixel probabilities toward 0 or 1).
- The Synergy: Entropy minimization alone is risky as the model might become confidently wrong. However, when guided by the Teacher via KL divergence, the entropy term is "reined in," leading to much more accurate segmentations.
Experimental Results
The study utilized the 2017 ACDC Challenge dataset, containing 100 cine Magnetic Resonance (MR) exams, with the task of segmenting the left-ventricle (LV) endocardium. The authors tested three scenarios (Set-3, Set-5, and Set-10), where only 3, 5, or 10 images were fully labeled.
One of the most striking findings is that the student branch consistently outperforms the teacher branch. In the most challenging scenario (Set-3), the "Student" achieved a Dice Similarity Score (DSC) of 75.92, significantly higher than the "Teacher" at 69.25. Compared to other state-of-the-art mixed-supervision strategies like the "Decoupled" approach, this method improved Dice scores by 10-20% and reduced the Hausdorff Distance (HD-95) by nearly 30%.
Discussion and Conclusion
The results suggest that the student's exposure to a wider variety of images—even if weakly labeled—allows it to generalize better than a teacher trained on a limited set of perfect labels. Furthermore, the authors show that Shannon Entropy is superior to standard "pseudo-masking" because it is "softer," preventing the model from locking in early mistakes. While currently focused on cardiac MRI, future work includes validating the method on more complex multi-organ tasks and exploring automated hyperparameter balancing for the various loss terms.
Summary updated