BiV-ICL: Bilateral Visual In-Context Learning for Enhancing Semi-Supervised Medical Image Segmentation

The efficacy of semi-supervised segmentation models is contingent upon the quantity and quality of labeled data. Specifically, in pseudo-label based semi supervised segmentation paradigms, limited labeled data often hampers the generalization capacity of teacher models, resulting in suboptimal utilization of unlabeled data and diminished student model performance. To address this issue, inspired by the ability of in-context learning (ICL) models to generalize to new tasks without retraining through paired in-context examples, we propose a novel semi supervised segmentation framework based on bilateral visual in-context learning (BiV-ICL). Our BiV-ICL incorporates a selective in-context learning (S-ICL) module to produce high-quality pseudo-labels from minimal labeled data. Additionally, to further enhance the performance of the student model, BiV-ICL introduces an innovative inverse in-context learning (I-ICL) module to effectively screen out unreliable pseudo-labels. Extensive experiments demonstrate that BiV-ICL consistently exhibits outstanding performance and generalization capabilities across various publicly available datasets with varying modalities and data splits, surpassing state-of-the-art semi-supervised alternatives.