Scholarlink

<p>Fractures are among the most common presentations to emergency and orthopedic services, yet radiograph interpretation remains inconsistent, particularly in resource-limited and off-hours settings. Weakly supervised learning offers a scalable alternative to costly pixel-level annotation but often fails to capture subtle, focal fracture cues. We propose FAD-MIL (Fracture-Aware Dual-stream Multiple-Instance Learning), which addresses these limitations through three design choices: (1) a global–local dual-stream architecture that captures both whole-image context and fine-grained tile features, (2) a fracture-aware gating mechanism that re-weights instance representations toward fracture-discriminative patterns, and (3) Top-K instance selection that focuses learning on the most informative regions. On FracAtlas (4,068 radiographs), FAD-MIL achieves an AUC of 0.833 (95% CI 0.797–0.871), average precision of 0.619, and F1 of 0.541, outperforming Mean-Pool MIL and Tile-Vote MIL and performing comparably to ABMIL while offering more interpretable instance-level attribution. Transferability was assessed on a retrospective single-center positive-only cohort (distal radius, <i>n</i> = 975; ankle, <i>n</i> = 350); because contemporaneous non-fracture controls were unavailable, recall across decision thresholds is reported as a preliminary sensitivity analysis. Gradient-based feature-attribution heatmaps provide a qualitative visualization of regions associated with fracture predictions. Future validation with matched non-fracture controls is required to further evaluate specificity and false-positive rates.</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

FAD-MIL: a weakly supervised fracture detection model based on X-ray images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • Feng Xue,
  • Yuan Zhang,
  • Wen Zhao,
  • Zexi Wang,
  • Abulikemu Maimaiti,
  • Jie Zhang,
  • Yingting Li

摘要

Fractures are among the most common presentations to emergency and orthopedic services, yet radiograph interpretation remains inconsistent, particularly in resource-limited and off-hours settings. Weakly supervised learning offers a scalable alternative to costly pixel-level annotation but often fails to capture subtle, focal fracture cues. We propose FAD-MIL (Fracture-Aware Dual-stream Multiple-Instance Learning), which addresses these limitations through three design choices: (1) a global–local dual-stream architecture that captures both whole-image context and fine-grained tile features, (2) a fracture-aware gating mechanism that re-weights instance representations toward fracture-discriminative patterns, and (3) Top-K instance selection that focuses learning on the most informative regions. On FracAtlas (4,068 radiographs), FAD-MIL achieves an AUC of 0.833 (95% CI 0.797–0.871), average precision of 0.619, and F1 of 0.541, outperforming Mean-Pool MIL and Tile-Vote MIL and performing comparably to ABMIL while offering more interpretable instance-level attribution. Transferability was assessed on a retrospective single-center positive-only cohort (distal radius, n = 975; ankle, n = 350); because contemporaneous non-fracture controls were unavailable, recall across decision thresholds is reported as a preliminary sensitivity analysis. Gradient-based feature-attribution heatmaps provide a qualitative visualization of regions associated with fracture predictions. Future validation with matched non-fracture controls is required to further evaluate specificity and false-positive rates.