Scholarlink

<p>To address the clinical challenges of interpreting subtle radiographic signs of high altitude pulmonary edema and the limited accessibility of invasive EVLWI monitoring, this study validates a deep learning-based framework for noninvasive, quantitative prediction of EVLWI from chest X-ray images. Utilizing a Vision Transformer architecture adapted for medical image regression, the model integrates lung field segmentation, hybrid local–global feature extraction, and relative position encoding to enhance sensitivity to early pathological changes. Trained and evaluated on data from a high altitude medical center in Qinghai (<i>n</i> = 1,280), the model achieves a Pearson correlation coefficient of 0.921 and an RMSE of 1.324 mL/m². External validation on an independent cohort from Tibet (<i>n</i> = 1,646) confirms robust generalizability, with high accuracy maintained in subclinical stages (R²=0.832). This study demonstrates the clinical feasibility of translating imaging-based deep learning into a noninvasive, interpretable tool for early warning and decision support in high altitude pulmonary edema.</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

Construction and validation of high altitude pulmonary edema prediction models based on deep learning and quantitative analysis of X-ray images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月27日

查看全文
  • Xuelong Geng,
  • Ziyi Wang,
  • Yi Liu,
  • Jing Zhu,
  • Qijie Xiang,
  • Feizhou Du,
  • Peng Wang,
  • Rui Jiang

摘要

To address the clinical challenges of interpreting subtle radiographic signs of high altitude pulmonary edema and the limited accessibility of invasive EVLWI monitoring, this study validates a deep learning-based framework for noninvasive, quantitative prediction of EVLWI from chest X-ray images. Utilizing a Vision Transformer architecture adapted for medical image regression, the model integrates lung field segmentation, hybrid local–global feature extraction, and relative position encoding to enhance sensitivity to early pathological changes. Trained and evaluated on data from a high altitude medical center in Qinghai (n = 1,280), the model achieves a Pearson correlation coefficient of 0.921 and an RMSE of 1.324 mL/m². External validation on an independent cohort from Tibet (n = 1,646) confirms robust generalizability, with high accuracy maintained in subclinical stages (R²=0.832). This study demonstrates the clinical feasibility of translating imaging-based deep learning into a noninvasive, interpretable tool for early warning and decision support in high altitude pulmonary edema.