DCCE-UNet: a difference and context-aware contrast enhanced framework for ultrasound image segmentation.
Authors
Affiliations (6)
Affiliations (6)
- Department of Neurosurgery, Huantai County People's Hospital, Huantai Avenue, Zibo, Shandong, 256400, China.
- Department of Neurology, Huantai County People's Hospital, Huantai Avenue, Zibo, Shandong, 256400, China.
- Blood Purification Center, Huantai County People's Hospital, Huantai Avenue, Zibo, Shandong, 256400, China.
- Department of Head and Neck Radiotherapy, Shandong Second Provincial General Hospital, Duanxing West Road, Jinan, Shandong, 250021, China.
- Department of Neurosurgery, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jingwuweiqi Road, Jinan, Shandong, 250021, China.
- Department of Neurosurgery, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jingwuweiqi Road, Jinan, Shandong, 250021, China. [email protected].
Abstract
Accurate segmentation of lesion regions in ultrasound images (US) is critical for computer-aided diagnosis and clinical decision-making. However, the inherent challenges of US imaging—including blurred lesion boundaries, low tissue contrast, and irregular morphological variations—significantly hinder the reliability and generalizability of existing segmentation methods. To overcome these limitations, we propose Difference and Context-aware Contrast Enhanced UNet (DCCE-UNet), a novel structure-aware and context-enhanced segmentation framework that extends the U-Net backbone with three specialized modules: (1) the multi-source attention-guided semantic generation block, which improves multi-scale semantic extraction by adaptively aggregating complementary feature cues, (2) the bidirectional difference-aware attention module, which enhances boundary modeling by capturing directional feature differences, and (3) the context-aware guidance module, which integrates spatial contextual dependencies to refine feature representations. These components work in synergy to improve structural perception and spatial discrimination in US image segmentation. Extensive experiments were conducted on two publicly available breast ultrasound datasets, BUSI and TN3K. The proposed DCCE-UNet consistently outperformed state-of-the-art models across key evaluation metrics, achieving a Dice coefficient of 79.85% and an IoU of 66.46% on BUSI, and reaching 89.29% Dice and 88.83% IoU on TN3K. Ablation studies further confirmed the effectiveness and complementary of each module in improving segmentation accuracy and boundary integrity, particularly in challenging cases involving small lesions, heterogeneous textures, and ambiguous contours. These results demonstrate that DCCE-UNet provides a robust and generalizable solution for automated ultrasound image segmentation. The proposed framework exhibits strong potential for real-world clinical deployment, with implications for improving diagnostic efficiency, reducing manual workload, and enabling reliable lesion analysis in ultrasound imaging workflows. Not applicable.