LR-Contrast: Accurate Multi-modal Brain Tumor Segmentation based on Local-Region Supervised Contrastive Learning

Early detection and accurate segmentation of brain tumors is of paramount importance for diagnosis and treatment. Due to the abundant information in multi-modal data, magnetic resonance imaging (MRI) has become the primary basis for brain tumor image segmentation. The existing deep learning-based segmentation methods mainly focus on refining the internal structure of deep networks, most of them employ Dice loss and Cross-Entropy loss to guide model learning. However, due to the limitations of these two losses in detailed areas, the discriminative ability and the segmentation accuracy of the model for these areas is not enough. To address this issue, we propose a multi-modal brain tumor segmentation method based on local-region supervised contrastive learning (LR-Contrast). The proposed method comprises an Anchor Selector Module (ASM), a Contrastive Sample Selector Module (CSSM) and a Supervised Contrastive Loss. ASM is designed to effectively pinpoint the detailed regions that are most suitable for contrastive learning. CSSM aims to identify efficient contrastive samples. Contrastive Loss is designed for improving the discriminative ability of features using contrastive samples in the detailed areas obtained by ASM. Extensive experiments on two public datasets (BraTS2020 and ISBI) indicate that the proposed method achieves superior performance in the Dice Score and 95% Hausdorff Distance (HD95) metrics compared to several state-of-the-art methods, making it more suitable for accurate segmentation of brain tumors.