Weakly supervised brain tumour segmentation with label propagation and level set loss

Abstract Early diagnosis of brain tumors significantly enhances treatment success. However, accurate detection and segmentation of tumors, essential for diagnosis, rely heavily on costly manual annotation by experts. To mitigate these costs, weakly supervised methods have gained traction. This paper introduces a novel weakly supervised brain tumor segmentation approach utilizing point and scribble supervision. Experts annotate only the slice with the largest tumor area by marking a single point near the tumor center or drawing a scribble within the tumor region. The method operates in two phases. First, labels are propagated to unlabelled pixels, generating a pseudo‐ground‐truth with three labels: tumor, non‐tumor, and marginal pixels (unlabelled pixels surrounding the initial segmentation). Second, a segmentation model is trained using the pseudo‐ground‐truth and a loss function combining level‐set and binary cross‐entropy losses. Marginal pixels contribute to level‐set loss computation, refining the segmentation process. The approach is validated on 3D magnetic resonance imaging (MRI) volumes from BraTS2020, BraTS2021, and BraTS2023 benchmark datasets. Experimental results show Dice scores comparable to fully supervised methods for whole tumor segmentation, demonstrating the effectiveness of the proposed weakly supervised strategy. This method reduces annotation effort while maintaining competitive segmentation performance, making it valuable for clinical applications.