RL-U$^2$Net: A Dual-Branch UNet with Reinforcement Learning-Assisted Multimodal Feature Fusion for Accurate 3D Whole-Heart Segmentation

On November 13, 2025, the RL-U$^2$Net model was proposed to improve 3D whole-heart segmentation, a vital process for diagnosing cardiovascular diseases. Traditional multi-modal segmentation methods struggle with spatial inconsistencies and lack adaptability in feature fusion. RL-U$^2$Net employs a dual-branch U-Net architecture, processing CT and MRI data in parallel, and introduces the RL-XAlign module, which uses a cross-modal attention mechanism to enhance feature alignment. This approach allows for a more dynamic integration of information, significantly boosting segmentation accuracy and robustness. The model's reinforcement learning component learns optimal strategies for feature alignment, addressing the inefficiencies of previous methods. By integrating predictions from individual patches through an ensemble-learning-based decision module, RL-U$^2$Net represents a significant advancement in the field, promising better outcomes in cardiovascular diagnostics and treatment plannin…