Scholarlink

<p>Medical image segmentation faces persistent challenges in anatomical heterogeneity, low-contrast boundaries, and multi-scale pathology presentations. Existing approaches compromise between convolutional networks' spatial precision and transformers' contextual awareness, while skip connections suffer from semantic discontinuities. To overcome these limitations, we propose DMSF-Net (Deformable MultiScale Fusion Net). Distinct from standard hybrids that passively combine modules, we introduce a 'Coupled Guidance' paradigm: (1) A Hybrid Convolution-Attention (HCA) mechanism where global attention maps actively modulate the offset generation of deformable convolutions. This solves the 'blind deformation' issue of standard CNNs by driving receptive field adjustments with global semantic topology, enabling pixel-level boundary refinement; (2) Multi-scale Dynamic Fusion (MDF) with channel attention-guided feature recalibration and deformable spatial refinement for scale-invariant lesion detection; and (3) Gated Hierarchical Skip Connections resolving semantic discontinuities through affine recalibration and anatomically informed spatial weighting. DMSF-Net achieves state-of-the-art performance on two clinically critical datasets: On the SLE dataset (94 images, irregular lesions), it attains 76.4% IoU, a 3.6% absolute improvement over TransUNet++; for ISIC2016 dermoscopy (1279 images), it reaches 85.7% mean IoU (+ 0.2% over DeeplabV3+) with 92.3% mean Dice accuracy. Crucially, these advancements are delivered with real-time inference (10.87&#xa0;ms/image), reducing latency by 39.9% versus transformer baselines while maintaining moderate parameter complexity (29.13&#xa0;M). Ablation studies validate each module’s contribution, with the full model consistently outperforming partial implementations, achieving up to 4.9% IoU improvement on the SLE dataset. This work establishes a new paradigm by unifying deformable operations, attention gating, and multi-scale fusion within an end-to-end trainable framework. DMSF-Net bridges critical clinical deployability gaps, providing unprecedented accuracy on historically intractable cases while meeting real-time operational constraints, and sets a foundational standard for next-generation medical imaging solutions.</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

DMSF-Net: Synergistic Integration of Deformable Convolution and Attention Gating for Clinically Efficient Medical Image Segmentation

  • 期刊论文

  • 发表日期：

    2026年4月26日

查看全文
  • Kun Zhu,
  • Nur Hana Samsudin

摘要

Medical image segmentation faces persistent challenges in anatomical heterogeneity, low-contrast boundaries, and multi-scale pathology presentations. Existing approaches compromise between convolutional networks' spatial precision and transformers' contextual awareness, while skip connections suffer from semantic discontinuities. To overcome these limitations, we propose DMSF-Net (Deformable MultiScale Fusion Net). Distinct from standard hybrids that passively combine modules, we introduce a 'Coupled Guidance' paradigm: (1) A Hybrid Convolution-Attention (HCA) mechanism where global attention maps actively modulate the offset generation of deformable convolutions. This solves the 'blind deformation' issue of standard CNNs by driving receptive field adjustments with global semantic topology, enabling pixel-level boundary refinement; (2) Multi-scale Dynamic Fusion (MDF) with channel attention-guided feature recalibration and deformable spatial refinement for scale-invariant lesion detection; and (3) Gated Hierarchical Skip Connections resolving semantic discontinuities through affine recalibration and anatomically informed spatial weighting. DMSF-Net achieves state-of-the-art performance on two clinically critical datasets: On the SLE dataset (94 images, irregular lesions), it attains 76.4% IoU, a 3.6% absolute improvement over TransUNet++; for ISIC2016 dermoscopy (1279 images), it reaches 85.7% mean IoU (+ 0.2% over DeeplabV3+) with 92.3% mean Dice accuracy. Crucially, these advancements are delivered with real-time inference (10.87 ms/image), reducing latency by 39.9% versus transformer baselines while maintaining moderate parameter complexity (29.13 M). Ablation studies validate each module’s contribution, with the full model consistently outperforming partial implementations, achieving up to 4.9% IoU improvement on the SLE dataset. This work establishes a new paradigm by unifying deformable operations, attention gating, and multi-scale fusion within an end-to-end trainable framework. DMSF-Net bridges critical clinical deployability gaps, providing unprecedented accuracy on historically intractable cases while meeting real-time operational constraints, and sets a foundational standard for next-generation medical imaging solutions.