Fully Perturbed Self-Ensemble Framework using Cascaded Parallel CNN-Transformer for Semi-Supervised Medical Image Segmentation

Semi-supervised learning (SSL) has achieved remarkable progress in the field of medical image segmentation (MIS), but it still faces two main challenges. First, the consistency learning employed in existing semi-supervised MIS (SSMIS) methods is mainly restricted to a single perturbation or a simple combination of multiple perturbations, which limits their abilities to effectively exploit the potential information from unlabeled medical images. Second, although some SSMIS methods make full use of the significant structural differences between CNN and Transformer networks for model perturbation, which causes a new conflict in that a Transformer network usually requires a large amount of labeled data for model training but only a small amount of labeled data is provided to SSMIS. In this paper, a novel fully perturbed self-ensemble framework (FPSE) using cascaded parallel CNN-Transformer is proposed to address the aforementioned problems. First, we present a fully perturbed consistency learning strategy that empowers the framework to handle complex variations through the skillful synergy of data, model, and feature perturbations, effectively exploring the potential information from unlabeled medical images. Second, we design a novel strategy of model perturbation based on the cascaded parallel CNN-Transformer structure, which maximizes the efficacy of Transformer under limited labeled data since Transformer is operated on the shallow local features extracted by CNN, thus effectively alleviating the requirement of a large number of labeled data for the proposed network architecture. Experiments demonstrate that our FPSE framework achieves remarkable results, outperforming the existing state-of-the-art (SOTA) methods by 11.5%, 0.7% and 2.4% in Dice score when labeled data accounts for only 5% (ACDC), 5% (AbdomenCT-1K) and 1% (ISIC2018), respectively.