A self-supervised representation learning paradigm with global content perception and peritumoral context restoration for MRI breast tumor segmentation

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is highly sensitive in breast cancer screening and treatment assessment, with breast tumor segmentation being a pivotal step in comprehensive analysis. However, developing reliable automated segmentation models remains challenging, as generating sufficient high-quality annotations requires significant time and effort from physicians. In this study, we propose a novel self-supervised learning paradigm tailored for breast DCE-MRI, aimed at improving downstream breast tumor segmentation by utilizing knowledge extracted from unlabeled data. Specifically, we design a peritumoral context restoration task to learn local detail features from unlabeled data. Notably, we replace the typical random masking strategy with a Peritumoral Masking Strategy (PMS), leveraging contrast differences to preserve tumor semantics. Additionally, a global content perception module is proposed to enhance the network's ability to capture global features by contrasting inter-individual differences and predicting contrast agent states. We validated our method using a clinical dataset comprising 229 breast DCE-MRI cases. When transferring self-supervised knowledge to tumor segmentation tasks, our approach achieved an 87.8% Dice score and a 7.469 mm 95HD, surpassing both the model trained from scratch (Dice score: 84.1%, 95HD: 19.331 mm) and nine state-of-the-art self-supervised benchmarks. Moreover, the results also demonstrate the superiority of our method with limited annotated data. With only 50% annotated data, our method outperformed the model trained from scratch using complete annotations and exceeded advanced semi-supervised approaches under the same annotation conditions.