Mutual Distillation of Dual-Foundation Models for Semi-Supervised PET/CT Segmentation

The article presents the MuDuo framework, a mutual distillation approach that integrates structural and functional foundation models, specifically SAM-Med3D for CT and SegAnyPET for PET, to enhance semi-supervised learning in PET/CT segmentation. This method enables a lightweight student network to achieve state-of-the-art performance on the AutoPET dataset using only five labeled cases, significantly reducing annotation costs while leveraging unlabeled data. This advancement is crucial for practitioners, as it improves segmentation accuracy and efficiency in clinical applications with limited labeled datasets.