Bootstrapping Semi-supervised Medical Image Segmentation with Anatomical-aware Contrastive Distillation (2206.02307v4)

Abstract: Contrastive learning has shown great promise over annotation scarcity problems in the context of medical image segmentation. Existing approaches typically assume a balanced class distribution for both labeled and unlabeled medical images. However, medical image data in reality is commonly imbalanced (i.e., multi-class label imbalance), which naturally yields blurry contours and usually incorrectly labels rare objects. Moreover, it remains unclear whether all negative samples are equally negative. In this work, we present ACTION, an Anatomical-aware ConTrastive dIstillatiON framework, for semi-supervised medical image segmentation. Specifically, we first develop an iterative contrastive distillation algorithm by softly labeling the negatives rather than binary supervision between positive and negative pairs. We also capture more semantically similar features from the randomly chosen negative set compared to the positives to enforce the diversity of the sampled data. Second, we raise a more important question: Can we really handle imbalanced samples to yield better performance? Hence, the key innovation in ACTION is to learn global semantic relationship across the entire dataset and local anatomical features among the neighbouring pixels with minimal additional memory footprint. During the training, we introduce anatomical contrast by actively sampling a sparse set of hard negative pixels, which can generate smoother segmentation boundaries and more accurate predictions. Extensive experiments across two benchmark datasets and different unlabeled settings show that ACTION significantly outperforms the current state-of-the-art semi-supervised methods.

• The paper introduces the ACTION framework utilizing iterative contrastive distillation to capture nuanced semantic similarities between image features.
• It addresses class imbalance by integrating global semantic relationships with local anatomical feature learning to boost segmentation accuracy.
• Extensive experiments on benchmark datasets show that ACTION outperforms existing methods with smoother boundaries and more accurate predictions.

The paper "Bootstrapping Semi-supervised Medical Image Segmentation with Anatomical-aware Contrastive Distillation" introduces the ACTION framework, a novel approach to enhancing medical image segmentation through semi-supervised learning. The authors tackle the challenges posed by class imbalance and the need for more nuanced handling of negative samples in contrastive learning.

Key contributions of the paper include:

1. Iterative Contrastive Distillation: This approach refines the process of labeling negative samples. Instead of using binary labels to distinguish between positive and negative pairs, the method involves soft labeling. This allows the model to capture semantic similarities between features of negative samples, fostering diversity in the sampled data.
2. Handling Imbalanced Samples: The paper raises the question of effectively managing imbalanced data to improve segmentation performance. The authors propose learning both global semantic relationships across the entire dataset and local anatomical features among neighboring pixels, achieving this with minimal additional memory.
3. Anatomical Contrast Sampling: By sampling a sparse set of hard negative pixels, the method introduces anatomical contrast, which helps produce smoother segmentation boundaries and more accurate predictions.
4. Empirical Validation: The authors validate ACTION through extensive experiments on two benchmark datasets within various unlabeled settings. The results demonstrate that ACTION significantly outperforms existing state-of-the-art semi-supervised methods, highlighting its effectiveness in addressing annotation scarcity and class imbalance.

Overall, the paper provides a comprehensive solution to enhancing the performance of semi-supervised medical image segmentation, particularly in contexts with imbalanced data distributions, by integrating innovative contrastive learning techniques.