ABSTRACT Medical image segmentation plays a critical role in computer‐aided diagnosis and clinical decision‐making; however, the performance of U‐Net and its variants is often degraded when dealing with images exhibiting blurred lesion boundaries and pronounced imaging heterogeneity. To address these challenges, this paper proposes DFABU‐Net (Double Frame Attention‐Based U‐Net), a dual‐channel segmentation framework designed to enhance boundary perception and robust feature representation. Specifically, a dual‐channel encoder architecture is constructed by incorporating a Large‐Scale Feature Transmission (LSFT) encoder to capture global contextual and morphological information, while an Attention Contrast Fusion Module (ACFM) is introduced to perform comparative analysis and adaptive fusion of dual‐channel features, thereby emphasizing boundary‐related information and reducing ambiguity in edge pixel classification. Extensive experiments are conducted on a public benchmark dataset (LiTS) and an internal clinical dataset. On the LiTS dataset, DFABU‐Net achieves a Dice coefficient of 98.34% and an mIoU of 96.20%, outperforming representative state‐of‐the‐art methods by 0.44% in Dice and 0.45% in mIoU, respectively. On the internal dataset, DFABU‐Net attains a Dice score of 97.64% and an mIoU of 95.88%, achieving the best performance among all compared methods and improving the Dice score by 0.49% over the strongest competing approach. Qualitative comparisons further demonstrate that DFABU‐Net produces more accurate and complete lesion boundaries. In addition, independent prognostic analyses based on model‐generated segmentations and expert annotations validate the clinical relevance and practical reliability of the proposed framework.
Zhang et al. (Mon,) studied this question.