Importance-Aware Transformer: Addressing Intra-Class Heterogeneity in Weakly Supervised Brain Tumor Segmentation

Weakly supervised brain tumor segmentation is a critical challenge in medical image analysis, characterized by inherent biological complexity and the scarcity of annotations. For segmentation on magnetic resonance imaging from image-level annotations, the intrinsic intra-class heterogeneity hampers accurate segmentation on distal images. In this paper, we propose a transformer-based model for weakly supervised brain tumor segmentation. Our approach tackles intra-class heterogeneity by utilizing an Importance-Aware Class Prediction strategy to improve segmentation accuracy on positive images. Additionally, we integrate Affine Transformation Equivariant Regularization to ensure consistent tumor localization. Finally, we introduce a Cross-Plane Refinement method to refine inaccurate segmentations using high-confidence images within the same volume. Evaluated on the Brain Tumor Segmentation 2023 dataset, our model demonstrates state-of-the-art performance on native Tl and T2-weighted sequences.