Multimodal Rebalanced Network with bidirectional interaction for tumor segmentation in PET/CT images

Accurate tumor segmentation is essential for computer-aided cancer diagnosis and treatment planning. Positron emission tomography (PET) and computed tomography (CT) are widely used in deep medical image segmentation for their complementary diagnostic information. However, existing multimodal approaches suffer from optimization imbalance across modalities and from limited interaction between local and global context. We address these issues with BGASGM-Net, a multimodal rebalancing network that integrates bidirectional global–local attention and Shapley dynamic gradient modulation. The attention module comprehensively captures multi-stage local and global contextual features, promoting bidirectional interaction between them; while the gradient-modulation module dynamically adjusts the gradient update intensity for PET and CT modality learning, enhancing the model’s focus on under-optimized modalities. Extensive quantitative and qualitative experiments verify the contribution of each component, and BGASGM-Net achieves significantly better segmentation performance than state-of-the-art (SOTA) baselines.