Scholarlink

<p>ELU-Net is an efficient, lightweight U-Net architecture for skin lesion segmentation, optimized to meet clinical computational constraints while addressing three core challenges: lesion scale variability, blurred boundary delineation, and sample feature heterogeneity. The network resolves these issues using three specific mechanisms: First, the Adaptive Scale Feature Module (ASFModule) addresses scale variability by processing features through parallel depthwise and dilated convolution paths. A learned gating mechanism dynamically aggregates multi-scale information, yielding compact descriptors that capture both local details and global context. Second, the Adaptive Receptive Field Convolution (ARFConv) optimizes receptive-field feature extraction. By employing intra-patch interaction, per-patch attention weighting, and response reprojection, ARFConv enhances boundary and texture distinction, mitigating the insufficient spatial capture inherent in standard parameter-sharing convolutions. Finally, the Cross-Sample Feature Enhancement Module (CSFEModule) improves generalization across diverse samples. It introduces dataset-level global context via memory interaction space modeling, integrating cross-sample commonalities and differences into local representations through residual fusion. Combining this architecture, ELU-Net exhibits both compactness (approximately 0.27M parameters, 0.153 GFLOPs) and high performance, achieving mean Intersection over Union (mIoU) of 79.84%/80.58% and Dice coefficients of 88.79%/89.25% on the ISIC2017 and ISIC2018 benchmark datasets, respectively. Comparative results demonstrate that ELU-Net simultaneously improves segmentation accuracy and robustness at minimal computational cost, offering an efficient solution for rapid clinical screening. The code is available at <a href="https://github.com/1502GaoYi/ELU-Net.">https://github.com/1502GaoYi/ELU-Net.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ELU-Net: An Efficient Lightweight U-Net Model for Skin Lesion Segmentation

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Yi Gao,
  • Haicheng Qu,
  • Yang Yu,
  • Jinshi Bao,
  • Junyi Yue,
  • Zhenxing Hui

摘要

ELU-Net is an efficient, lightweight U-Net architecture for skin lesion segmentation, optimized to meet clinical computational constraints while addressing three core challenges: lesion scale variability, blurred boundary delineation, and sample feature heterogeneity. The network resolves these issues using three specific mechanisms: First, the Adaptive Scale Feature Module (ASFModule) addresses scale variability by processing features through parallel depthwise and dilated convolution paths. A learned gating mechanism dynamically aggregates multi-scale information, yielding compact descriptors that capture both local details and global context. Second, the Adaptive Receptive Field Convolution (ARFConv) optimizes receptive-field feature extraction. By employing intra-patch interaction, per-patch attention weighting, and response reprojection, ARFConv enhances boundary and texture distinction, mitigating the insufficient spatial capture inherent in standard parameter-sharing convolutions. Finally, the Cross-Sample Feature Enhancement Module (CSFEModule) improves generalization across diverse samples. It introduces dataset-level global context via memory interaction space modeling, integrating cross-sample commonalities and differences into local representations through residual fusion. Combining this architecture, ELU-Net exhibits both compactness (approximately 0.27M parameters, 0.153 GFLOPs) and high performance, achieving mean Intersection over Union (mIoU) of 79.84%/80.58% and Dice coefficients of 88.79%/89.25% on the ISIC2017 and ISIC2018 benchmark datasets, respectively. Comparative results demonstrate that ELU-Net simultaneously improves segmentation accuracy and robustness at minimal computational cost, offering an efficient solution for rapid clinical screening. The code is available at https://github.com/1502GaoYi/ELU-Net.