Adaptive Fusion and Edge‐Oriented Enhancement for Brain Tumor Segmentation With Missing Modalities

ABSTRACT Magnetic resonance imaging (MRI) offers comprehensive information about brain structures, enabling excellent performance in brain tumor segmentation using multimodal MRI in many methods. Nonetheless, missing modalities are common in clinical practice, which can significantly degrade segmentation performance. Current brain tumor segmentation methods often struggle to maintain feature consistency and robustness in multimodal feature fusion when modalities are missing and face difficulties in accurately capturing tumor boundaries. In this study, we propose an adaptive fusion and edge‐oriented enhancement method to address these challenges. Our approach introduces learnable parameters and a masked attention mechanism in the transformer model to achieve cross‐modal adaptive fusion, ensuring consistent feature representation even with missing data. To aggregate more information, we integrate multimodal and multi‐level features through a hierarchical context integration module. Additionally, to tackle the complex morphology of brain tumor regions, we design an edge‐enhanced deformable convolution module that captures deformation information and edge features from incomplete multimodal images, enabling precise tumor localization. Evaluations on the widely recognized BRATS2018 and BRATS2020 datasets demonstrate that our approach significantly surpasses existing brain tumor segmentation techniques in scenarios with missing modalities.