DPSeg: Differentiated Perception Medical Image Segmentation via Context-Aware Compensation and Stratified Uncertainty Attention.
Authors
Abstract
In medical image segmentation, accurate recognition of regions hindered by blurry boundaries, low contrast, and complex backgrounds is crucial. These issues lead to significant differences in segmentation difficulty within a single image. However, existing methods typically adopt a uniform modeling strategy across the entire feature map, overlooking the mismatch between segmentation difficulty and attention intensity. Consequently, they fail to adequately focus on key regions, diluting the representations of challenging regions. To address these challenges, we propose DPSeg, a differentiated perception medical image segmentation framework via context-aware compensation and stratified uncertainty attention. Specifically, DPSeg comprises three key components. The Context-Aware Compensation (CAC) module refines encoder features by integrating key region-aware prompts and contextual information, compensating for encoding-induced distortions and enabling differentiated perception, thereby providing higher-quality representations for decoding. Additionally, the Stratified Uncertainty Transformer (SUTrans) module explicitly modulates the attention intensity of different regions based on their segmentation uncertainty, thereby possessing differentiated perception ability while effectively suppressing redundant feature interactions. Moreover, we propose a Foreground-Uncertainty Adaptive Weighted Loss (FUAWL) to enable differentiated supervision across decoding stages, applying stronger guidance to uncertain regions in shallow decoding and emphasizing foreground regions in deep decoding to improve boundary delineation and mitigate foreground-background imbalance. Extensive experiments on four challenging datasets demonstrate that DPSeg achieves competitive segmentation performance.