Cross modal reliable pixel contrastive learning for incomplete modal brain tumor segmentation

Abstract In clinical practice, Magnetic Resonance Imaging (MRI) data frequently suffer from the absence of certain imaging modalities, which inevitably degrades predictive performance. Existing approaches typically treat different modalities as independent and non-interacting during modal feature extraction, despite the presence of rich pixel-wise semantic correlations across modalities. To address this limitation, we propose Cross Modal Reliable Pixel Contrastive Learning (CMR-PCL), a novel framework for incomplete-modality brain tumor segmentation that aims to compensate for inter-modality information loss. Specifically, we introduce a label-inaccuracy–guided sampling strategy for each modality to identify and preserve reliable regions, thereby reducing the risk of noisy sample selection. We then enforce that reliable pixel embeddings belonging to the same semantic class are more similar to each other than to those from different classes. CMR-PCL adopts a standard training paradigm and imposes no additional architectural constraints, allowing it to be seamlessly integrated into existing incomplete-modality brain tumor segmentation frameworks. Extensive experiments on the BraTS2020, BraTS2018, and BraTS2015 datasets demonstrate that CMR-PCL consistently improves the performance of state-of-the-art methods.