Scholarlink

<p>To address diagnostic delays in pediatric abdominal emergencies, this study aimed to develop and validate multi-institutional deep learning models for detecting intussusception and splenomegaly on abdominal radiographs, thereby evaluating their potential to enhance clinical triage. This retrospective study included 26,552 radiographs from seven tertiary hospitals (2012–2022). Models were trained with EfficientDet-B2, selected after an ablation study against ResNet-50, DenseNet-121, and other EfficientDet variants. Three strategies were compared: independent binary classifiers, a multiclass model, and transfer learning (fine-tuning from intussusception to splenomegaly). External performance was primarily assessed using a locked external test set, with leave-one-institution-out validation performed as a secondary robustness analysis. Subgroup analyses were performed by institution and body weight (&lt; 10&#xa0;kg, 10– &lt; 20&#xa0;kg, 20- &lt; 30&#xa0;kg), and visualized using area under curve (AUC) with 95% confidence intervals. An ablation study identified EfficientDet-B2 as the best-performing backbone among ResNet-50, DenseNet-121, and other EfficientDet variants. The EfficientDet-B2 models achieved internal AUCs of 0.851 for intussusception and 0.834 for splenomegaly, with locked external AUCs of 0.818 and 0.806, respectively. Secondary leave-one-institution-out (LOIO) AUCs ranged from 0.832 to 0.912, while multiclass training and transfer learning improved intussusception and splenomegaly performance, respectively. Subgroup analyses likewise showed stable AUC across institutions, with reduced sensitivity in the &lt; 10&#xa0;kg group but maintained performance in those weighing 10– &lt; 30&#xa0;kg. Multiclass training improved intussusception detection, whereas transfer learning enhanced splenomegaly recognition. With robust performance across institutions and clinically interpretable subgroup analyses, these models show promise for supporting triage of pediatric abdominal radiographs. Prospective validation in real-world workflows is warranted to confirm utility.</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

Transfer learning for Multi-institutional Classification of Intussusception and Splenomegaly in Pediatric Abdominal Radiographs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Minsoo Shin,
  • Sungwon Ham,
  • Yoon Lee,
  • Jung Ok Shim

摘要

To address diagnostic delays in pediatric abdominal emergencies, this study aimed to develop and validate multi-institutional deep learning models for detecting intussusception and splenomegaly on abdominal radiographs, thereby evaluating their potential to enhance clinical triage. This retrospective study included 26,552 radiographs from seven tertiary hospitals (2012–2022). Models were trained with EfficientDet-B2, selected after an ablation study against ResNet-50, DenseNet-121, and other EfficientDet variants. Three strategies were compared: independent binary classifiers, a multiclass model, and transfer learning (fine-tuning from intussusception to splenomegaly). External performance was primarily assessed using a locked external test set, with leave-one-institution-out validation performed as a secondary robustness analysis. Subgroup analyses were performed by institution and body weight (< 10 kg, 10– < 20 kg, 20- < 30 kg), and visualized using area under curve (AUC) with 95% confidence intervals. An ablation study identified EfficientDet-B2 as the best-performing backbone among ResNet-50, DenseNet-121, and other EfficientDet variants. The EfficientDet-B2 models achieved internal AUCs of 0.851 for intussusception and 0.834 for splenomegaly, with locked external AUCs of 0.818 and 0.806, respectively. Secondary leave-one-institution-out (LOIO) AUCs ranged from 0.832 to 0.912, while multiclass training and transfer learning improved intussusception and splenomegaly performance, respectively. Subgroup analyses likewise showed stable AUC across institutions, with reduced sensitivity in the < 10 kg group but maintained performance in those weighing 10– < 30 kg. Multiclass training improved intussusception detection, whereas transfer learning enhanced splenomegaly recognition. With robust performance across institutions and clinically interpretable subgroup analyses, these models show promise for supporting triage of pediatric abdominal radiographs. Prospective validation in real-world workflows is warranted to confirm utility.