Scholarlink

<p>The accurate diagnosis of large vessel occlusion in acute ischemic stroke using Non-Contrast Computed Tomography (NCCT) is often impeded by the subtle nature of early radiological signs and the intrinsic heterogeneity between imaging and clinical data. Current multimodal fusion approaches frequently suffer from domain shift when applying pre-trained networks to medical scans and fail to effectively model the non-linear synergy between high-dimensional visual features and low-dimensional clinical indicators. To overcome these challenges, we propose the Multimodal Medical Fusion Network (M<sup>2</sup>FNet), a unified framework designed for robust and interpretable occlusion detection. M<sup>2</sup>FNet introduces three core architectural innovations: (1) a Hierarchical Visual Encoder equipped with a trainable Domain-Adaptive Stem, which aligns the grayscale medical manifold with the feature space of frozen pre-trained backbones to mitigate covariate shift; (2) a Transformer-based Tabular Encoder that treats clinical variables as a pseudo-sequence to capture intra-modal feature dependencies; and (3) a Tri-Pathway Synergistic Fusion Mechanism that orchestrates bidirectional cross-attention and gated modality weighting to dynamically recalibrate feature importance. Validated on a multi-center cohort of 599 patients (1,542 NCCT slices) from two Grade 3A hospitals, M<sup>2</sup>FNet significantly outperforms state-of-the-art methods, achieving an AUC of 0.956 on the testing set and exhibiting exceptional robustness with an AUC of 0.958 and specificity of 0.972 on independent external validation cohorts. These results confirm that M<sup>2</sup>FNet effectively resolves modality misalignment and provides a reliable, generalizable tool for computer-aided stroke diagnosis, demonstrating significant potential as a rapid triage adjunct to prioritize HAS-negative cases in real clinical workflows.</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

M2FNet: A multimodal fusion network for HAS-negative large vessel occlusion in acute ischemic stroke

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月8日

查看全文
  • Kun Zhu,
  • Hang Qu,
  • Yi Zhou

摘要

The accurate diagnosis of large vessel occlusion in acute ischemic stroke using Non-Contrast Computed Tomography (NCCT) is often impeded by the subtle nature of early radiological signs and the intrinsic heterogeneity between imaging and clinical data. Current multimodal fusion approaches frequently suffer from domain shift when applying pre-trained networks to medical scans and fail to effectively model the non-linear synergy between high-dimensional visual features and low-dimensional clinical indicators. To overcome these challenges, we propose the Multimodal Medical Fusion Network (M2FNet), a unified framework designed for robust and interpretable occlusion detection. M2FNet introduces three core architectural innovations: (1) a Hierarchical Visual Encoder equipped with a trainable Domain-Adaptive Stem, which aligns the grayscale medical manifold with the feature space of frozen pre-trained backbones to mitigate covariate shift; (2) a Transformer-based Tabular Encoder that treats clinical variables as a pseudo-sequence to capture intra-modal feature dependencies; and (3) a Tri-Pathway Synergistic Fusion Mechanism that orchestrates bidirectional cross-attention and gated modality weighting to dynamically recalibrate feature importance. Validated on a multi-center cohort of 599 patients (1,542 NCCT slices) from two Grade 3A hospitals, M2FNet significantly outperforms state-of-the-art methods, achieving an AUC of 0.956 on the testing set and exhibiting exceptional robustness with an AUC of 0.958 and specificity of 0.972 on independent external validation cohorts. These results confirm that M2FNet effectively resolves modality misalignment and provides a reliable, generalizable tool for computer-aided stroke diagnosis, demonstrating significant potential as a rapid triage adjunct to prioritize HAS-negative cases in real clinical workflows.