CrossMatch: Enhance Semi-Supervised Medical Image Segmentation with Perturbation Strategies and Knowledge Distillation

TL;DR

CrossMatch tackles the scarcity of labeled medical images by marrying self-knowledge distillation with dual perturbation strategies in a single-model, multi-encoder/decoder framework. By generating diverse outputs through two image-level encoders and three feature-perturbed decoders, CrossMatch enforces cross- and decoder-distillation losses, with a total objective $L_{tot} = L_{sup} + L_{ip} + (1-\eta)L_{tkd} + \eta L_{dkd}$, to exploit unlabeled data effectively. Extensive experiments on LA, ACDC, Pancreas-CT, and ISIC-2018 demonstrate state-of-the-art performance, especially at low labeling ratios, while maintaining competitive computational efficiency. Ablation analyses validate the roles of $L_{dkd}$, $L_{tkd}$, and $L_{ip}$, and provide insights into perturbation types, hyperparameters, and failure cases, highlighting CrossMatch’s practical value for clinical-segmentation tasks.

Abstract

Semi-supervised learning for medical image segmentation presents a unique challenge of efficiently using limited labeled data while leveraging abundant unlabeled data. Despite advancements, existing methods often do not fully exploit the potential of the unlabeled data for enhancing model robustness and accuracy. In this paper, we introduce CrossMatch, a novel framework that integrates knowledge distillation with dual perturbation strategies-image-level and feature-level-to improve the model's learning from both labeled and unlabeled data. CrossMatch employs multiple encoders and decoders to generate diverse data streams, which undergo self-knowledge distillation to enhance consistency and reliability of predictions across varied perturbations. Our method significantly surpasses other state-of-the-art techniques in standard benchmarks by effectively minimizing the gap between training on labeled and unlabeled data and improving edge accuracy and generalization in medical image segmentation. The efficacy of CrossMatch is demonstrated through extensive experimental validations, showing remarkable performance improvements without increasing computational costs. Code for this implementation is made available at https://github.com/AiEson/CrossMatch.git.

Figures (8)

• Figure 1: Comparison of different types of KD and SSL methods. (a) Traditional KD requires pre-training of the teacher model. (b) Self-KD based on data augmentation. (c) Mean Teacher. (d) FixMatch.
• Figure 2: Overview of our proposed CrossMatch. CrossMatch integrates the core ideas of Self-KD and SSL by enhancing performance through the derivation and mutual distillation of multiple encoder-decoder architectures.
• Figure 3: (a) Our proposed CrossMatch method with Weak Drop denoting $\mathcal{P}^w$ and Strong Drop denoting $\mathcal{P}^s$. (b) FixMatch.
• Figure 4: Visualizations of several semi-supervised segmentation methods with 10% and 20% labeled data and ground truth on Pancreas-CT dataset.
• Figure 5: some visualization examples of several semi-supervised segmentation methods with 10% labeled data and ground truth on the LA dataset.