Uncertainty-Guided Hybrid Decoder for Medical Image Segmentation

Medical image segmentation, a cornerstone of clinical diagnostics, faces persistent challenges in balancing global contextual understanding with precise local detail preservation. While conventional CNN-based methods excel at hierarchical feature extraction, their inherent locality constraints limit their ability to capture long-range anatomical dependencies—a critical shortcoming in complex scenarios such as vascular boundary delineation. Conversely, pure Transformer architectures, despite their global context modeling strengths, struggle with computational efficiency on high-resolution medical data and often produce suboptimal boundary localization due to the absence of spatial inductive biases. This paper introduces a novel hybrid segmentation framework designed to address these complementary limitations. Our method integrates a dual-path decoder architecture that synergistically combines Transformer-based global reasoning with CNN-based local refinement, enabling effective multi-scale feature fusion while maintaining computational feasibility. To further enhance clinical reliability, we propose an uncertainty-aware attention mechanism that dynamically prioritizes ambiguous anatomical regions during inference, mitigating overconfident predictions in low-contrast or noise-prone areas. Experimental evaluations on multi-organ CT and retinal vessel segmentation tasks demonstrate that our method achieves SOTA performance in boundary localization accuracy.