RDPA-UNet: Differential-Path Feature Learning and Multi-Scale Context Aggregation for Ultrasound Lesion Segmentation

Abstract

Accurate lesion segmentation in ultrasound images remains challenging because of severe speckle noise, low tissue contrast, ambiguous boundaries, and substantial variations in lesion morphology and scale. To address these limitations, we propose RDPA-UNet, an enhanced U-shaped network that integrates differential-path feature modeling with multi-scale contextual aggregation. First, a residual convolutional backbone equipped with Group Normalization is employed to facilitate feature propagation and stabilize optimization under small-batch training. Second, a Differential-Path Feature (DPF) block is introduced to jointly encode local-detail and dilated-context responses from different receptive fields. Their element-wise absolute difference is explicitly modeled to quantify cross-receptive-field response discrepancies and provide complementary structural information for subsequent feature fusion. Third, a Residual Multi-scale Atrous Spatial Pyramid Pooling (RMASPP) bottleneck aggregates high-level semantic information across multiple receptive fields and refines the fused representation, thereby improving the delineation of lesions with heterogeneous sizes and irregular contours. The proposed method was evaluated on the BUSI, DDTI, and Hemangioma datasets, as well as under a mixed-data setting, using five-fold cross-validation. RDPA-UNet achieved mean Dice scores of 78.30%, 77.05%, 83.58%, and 77.08%, together with mean IoU scores of 69.45%, 66.05%, 74.12%, and 67.10%, respectively. It achieved the highest mean Dice and IoU scores across all four dataset settings and obtained the lowest mean HD95 on BUSI, DDTI, and the mixed-data setting, while achieving the second-best HD95 on Hemangioma. Ablation experiments further demonstrated the effectiveness and complementarity of residual encoding, differential-path feature modeling, and multi-scale contextual aggregation. These results indicate that RDPA-UNet provides robust lesion segmentation performance across heterogeneous ultrasound datasets while maintaining competitive boundary-localization accuracy.