Scholarlink

<p>This study proposes a wrist radiography-based deep learning model for identifying radiographic features of metabolic bone disease (MBD) of prematurity and evaluate the impact of its decision support. This retrospective study included preterm infants with birth weights under 1500&#xa0;g, born at Seoul National University Hospital (internal dataset: 814 subjects) and Seoul National University Bundang Hospital (external dataset: 261 subjects). Demographic and clinical information and wrist radiographs (postnatal ages: 4–8 weeks) were collected. An internal dataset was used to develop and train identification models and an external dataset was used for validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and performance quality was compared based on paired <i>t</i>- and Wilcoxon signed-rank tests. The DenseNet-based model exhibited the highest performance quality with AUROC of 0.961 (sensitivity: 94.4%, specificity: 91.2%, accuracy: 92.0%). The external validation study yielded an AUROC of 0.927, indicating its potential applicability in a broad clinical setting. The reader study using external data demonstrated an improved reading performance, especially for non-radiologists (65.4% to 78.7% accuracy; <i>P</i> = 0.008, among pediatricians). The developed model can assist clinicians, especially non-radiologists, in identifying radiographic signs suggestive of MBD, enabling timely diagnosis and treatment to prevent disease progression.</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

Deep learning model for identification of metabolic bone disease of prematurity using wrist radiographs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月9日

查看全文
  • Seul Gi Park,
  • Seoi Jeong,
  • Minwoo Cho,
  • Mi Jin Kim,
  • Chang Won Choi,
  • Ee-Kyung Kim,
  • Jung-Eun Cheon,
  • Hyoun-Joong Kong

摘要

This study proposes a wrist radiography-based deep learning model for identifying radiographic features of metabolic bone disease (MBD) of prematurity and evaluate the impact of its decision support. This retrospective study included preterm infants with birth weights under 1500 g, born at Seoul National University Hospital (internal dataset: 814 subjects) and Seoul National University Bundang Hospital (external dataset: 261 subjects). Demographic and clinical information and wrist radiographs (postnatal ages: 4–8 weeks) were collected. An internal dataset was used to develop and train identification models and an external dataset was used for validation. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and performance quality was compared based on paired t- and Wilcoxon signed-rank tests. The DenseNet-based model exhibited the highest performance quality with AUROC of 0.961 (sensitivity: 94.4%, specificity: 91.2%, accuracy: 92.0%). The external validation study yielded an AUROC of 0.927, indicating its potential applicability in a broad clinical setting. The reader study using external data demonstrated an improved reading performance, especially for non-radiologists (65.4% to 78.7% accuracy; P = 0.008, among pediatricians). The developed model can assist clinicians, especially non-radiologists, in identifying radiographic signs suggestive of MBD, enabling timely diagnosis and treatment to prevent disease progression.