Scholarlink

<p>Alzheimer’s Disease (AD) presents subtle and diffuse brain changes that complicate both diagnosis and segmentation. To address this challenge, we introduce Dynamic Swin– UNet, a 3D Transformer-based framework that combines adaptive windowing, depth control, dropout scheduling, multi-resolution fusion, and boundary-aware patch-based inference. This design allows the model to spatially allocate computational resources according to local anatomical complexity, capturing fine details in small structures such as the hippocampus while preserving global cortical context. A boundary attention module, guided by continuity loss, further ensures spatial consistency across sliding-window patches. We evaluated our method comprehensively on three demographically diverse public datasets (ADNI- -4, AIBL, OASIS–3) across both region-of- interest and whole-brain segmentation tasks. Dynamic Swin–UNet consistently outperformed state-of-the-art baselines including Swin– UNETR, nnU-Net, 3D U-Net and TransBTS with statistically significant improvements across all metrics (<i>p</i> &lt; 0.01 ×, Wilcoxon signed- rank). Results demonstrate Dice scores exceeding 87% for hippocampal segmentation and 88% for whole-brain analysis, alongside substantial boundary improvements (Hausdorff distance reductions of 0.4–0.5 mm). The method achieved a 1.52 × inference speedup and 36.6% reduction in FLOPs while reducing memory usage by 22.9%. Sensitivity analysis revealed robust performance under clinically realistic conditions, maintaining stable accuracy (Dice drop <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\le\)</EquationSource> </InlineEquation>5%) under moderate noise (20 dB SNR) and motion artifacts. Ablation studies confirmed that each adaptive component contributes significantly to accuracy, efficiency, and training stability. Cross- dataset evaluation and harmonization robustness testing validated generalization across different scanner protocols and demographic profiles. Overall, Dynamic Swin–UNet establishes a new benchmark for computationally efficient and accurate AD neuroimaging analysis, demonstrating robust performance essential for early diagnosis, longitudinal monitoring, and scalable multi-center clinical studies.</p> Graphical Abstract <p></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

Dynamic Swin-UNet: a transformer-based adaptive framework for precise and efficient Alzheimer’s disease brain segmentation

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Marwa Ben Gara Ali,
  • Abir Smiti

摘要

Alzheimer’s Disease (AD) presents subtle and diffuse brain changes that complicate both diagnosis and segmentation. To address this challenge, we introduce Dynamic Swin– UNet, a 3D Transformer-based framework that combines adaptive windowing, depth control, dropout scheduling, multi-resolution fusion, and boundary-aware patch-based inference. This design allows the model to spatially allocate computational resources according to local anatomical complexity, capturing fine details in small structures such as the hippocampus while preserving global cortical context. A boundary attention module, guided by continuity loss, further ensures spatial consistency across sliding-window patches. We evaluated our method comprehensively on three demographically diverse public datasets (ADNI- -4, AIBL, OASIS–3) across both region-of- interest and whole-brain segmentation tasks. Dynamic Swin–UNet consistently outperformed state-of-the-art baselines including Swin– UNETR, nnU-Net, 3D U-Net and TransBTS with statistically significant improvements across all metrics (p < 0.01 ×, Wilcoxon signed- rank). Results demonstrate Dice scores exceeding 87% for hippocampal segmentation and 88% for whole-brain analysis, alongside substantial boundary improvements (Hausdorff distance reductions of 0.4–0.5 mm). The method achieved a 1.52 × inference speedup and 36.6% reduction in FLOPs while reducing memory usage by 22.9%. Sensitivity analysis revealed robust performance under clinically realistic conditions, maintaining stable accuracy (Dice drop \(\le\) 5%) under moderate noise (20 dB SNR) and motion artifacts. Ablation studies confirmed that each adaptive component contributes significantly to accuracy, efficiency, and training stability. Cross- dataset evaluation and harmonization robustness testing validated generalization across different scanner protocols and demographic profiles. Overall, Dynamic Swin–UNet establishes a new benchmark for computationally efficient and accurate AD neuroimaging analysis, demonstrating robust performance essential for early diagnosis, longitudinal monitoring, and scalable multi-center clinical studies.

Graphical Abstract