Scholarlink

<p>Acute Ischemic Stroke (AIS) necessitates precise and timely clinical decision-making to optimize patient outcomes. Current multimodal deep learning models often struggle with small or multifocal ischemic lesions due to faint features and cross-modal misalignments. To address these challenges, we introduce MMG-SiamNet, an advanced deep learning framework that integrates 2D and 3D data from Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. MMG-SiamNet leverages multi-scale convolutional layers to capture ischemic lesions across varying severities. It employs cross-modal contrastive learning to align CT and MRI images, reducing semantic drift and spatial misalignment. Furthermore, a Path Aggregation Network (PAN) fuses fine- and coarse-level features across modalities, minimizing feature redundancy and enhancing detection accuracy. Rigorous negative mining minimizes false positives, particularly in visually similar regions, while an attention-enhanced fusion layer prioritizes stroke-relevant features. Here we show that MMG-SiamNet achieves state-of-the-art performance in AIS detection, with an accuracy of 97.0%, specificity of 96.0%, and sensitivity of 96.5%. The Youden’s J statistic of 0.92, DICE coefficient of 97.0%, Matthews’ correlation coefficient of 0.950, Hausdorff distance of 9.00&#xa0;mm, and Average Symmetric Surface Distance of 8.90&#xa0;mm highlight its superior diagnostic precision. These results underscore MMG-SiamNet’s potential to improve clinical outcomes by enabling more accurate and reliable detection of ischemic strokes.</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

MMG-SiamNet: multi-modal granular siamese network for robust ischemic stroke detection via cross-modal learning

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Majid Rahman Lafta Alkaraawi,
  • Xiong Shengwu

摘要

Acute Ischemic Stroke (AIS) necessitates precise and timely clinical decision-making to optimize patient outcomes. Current multimodal deep learning models often struggle with small or multifocal ischemic lesions due to faint features and cross-modal misalignments. To address these challenges, we introduce MMG-SiamNet, an advanced deep learning framework that integrates 2D and 3D data from Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) scans. MMG-SiamNet leverages multi-scale convolutional layers to capture ischemic lesions across varying severities. It employs cross-modal contrastive learning to align CT and MRI images, reducing semantic drift and spatial misalignment. Furthermore, a Path Aggregation Network (PAN) fuses fine- and coarse-level features across modalities, minimizing feature redundancy and enhancing detection accuracy. Rigorous negative mining minimizes false positives, particularly in visually similar regions, while an attention-enhanced fusion layer prioritizes stroke-relevant features. Here we show that MMG-SiamNet achieves state-of-the-art performance in AIS detection, with an accuracy of 97.0%, specificity of 96.0%, and sensitivity of 96.5%. The Youden’s J statistic of 0.92, DICE coefficient of 97.0%, Matthews’ correlation coefficient of 0.950, Hausdorff distance of 9.00 mm, and Average Symmetric Surface Distance of 8.90 mm highlight its superior diagnostic precision. These results underscore MMG-SiamNet’s potential to improve clinical outcomes by enabling more accurate and reliable detection of ischemic strokes.