Abstract
Training deep learning models for medical image segmentation typically requires extensive expert-annotated data, which is both costly and time-consuming to obtain. This paper introduces a novel framework that integrates Semi-supervised Learning (SemiSL) with Contrastive Learning (CL) by utilizing automatically generated "pseudo-labels" to guide the contrastive process. By ensuring the model "pulls" similar tissues together and "pushes" different ones apart in the feature space, the proposed method achieves state-of-the-art results using as little as 2.5% of the original labeled data.
Introduction
In medical imaging, pixel-perfect labels are the gold standard but are incredibly difficult to produce at scale. Semi-supervised learning (SemiSL) attempts to mitigate this by training on a small labeled set and a large unlabeled set, often using "pseudo-labels" (the model's own predictions) to fill the gaps.
However, standard SemiSL often fails to create compact feature clusters for the same tissue type or clear separation between different types. Contrastive Learning (CL) is effective at organizing feature spaces but suffers from class collision: without labels, a CL algorithm might mistakenly "push apart" two patches of the same tissue type simply because they originate from different images. This paper bridges this gap by using pseudo-labels to inform the contrastive algorithm which patches should be considered similar.
Methodology
The proposed framework architecture (A) illustrates the Pseudo-label Guided Contrastive Loss (PLGCL) workflow, (B) shows the supervised learning component, and (C) details the student-teacher consistency training. Source, p. 4
The authors propose a patch-based, end-to-end framework consisting of four main components:
- Student-Teacher Architecture: A "Mean Teacher" setup where a student model learns from labeled data and a teacher model (an exponential moving average of the student) provides stable targets for unlabeled data.
- Pseudo-label Guided Contrastive Loss (PLGCL): Instead of picking random patches, the model uses the teacher’s pseudo-labels to identify "positives" (patches of the same class) and "negatives" (patches of different classes), optimizing for intra-class compactness and inter-class separability.
- Entropy-based Patch Sampling: The authors use an average patch entropy metric to select the most informative patches, focusing on those with high uncertainty and intensity variation rather than redundant background noise.
- Warm-up Phase: The model undergoes a 50-epoch warm-up using only standard supervised and consistency losses to ensure the initial pseudo-labels are accurate enough to guide the subsequent contrastive learning phase.
Experimental Results
The framework was evaluated on three diverse datasets: ACDC (Cardiac MRI), KiTS19 (Kidney CT), and CRAG (Histopathology). The proposed method consistently outperformed state-of-the-art semi-supervised methods across all metrics.
Key quantitative findings include:
- ACDC: With only 10% labeled data, the model achieved a Dice Similarity Score (DSC) of 0.891, nearly matching the fully supervised performance of 0.923.
- KiTS19: Using a tiny fraction of labels (2.5%), it achieved a DSC of 0.905, significantly beating other methods in low-data regimes.
- CRAG: On complex histopathology slides, it maintained the highest DSC and lowest Hausdorff Distance (HD95), proving its robustness to intricate textures.
Discussion and Conclusion
The results demonstrate a virtuous cycle: SemiSL provides the pseudo-labels that prevent class collision in CL, while the discriminative features learned by CL make the SemiSL segmentation more accurate. t-SNE visualizations confirm that with PLGCL, feature clusters become much tighter and more distinct compared to standard approaches.
While patch-based contrastive learning introduces some computational overhead, the significant reduction in required labeled data makes it a highly practical approach for medical domains. Future work may extend this pseudo-label guidance to other tasks like object detection or cross-modality synthesis.
Summary updated