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

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pediatric Radiology
• Deep Learning Applications

Background:

• Diagnostic delays in pediatric abdominal emergencies can lead to adverse outcomes.
• Accurate and timely detection of conditions like intussusception and splenomegaly is crucial for effective triage.

Purpose of the Study:

• To develop and validate multi-institutional deep learning models for detecting intussusception and splenomegaly on pediatric abdominal radiographs.
• To evaluate the potential of these models in enhancing clinical triage for pediatric abdominal emergencies.

Main Methods:

• A retrospective study of 26,552 radiographs from seven tertiary hospitals (2012-2022).
• EfficientDet-B2 models were trained and validated using strategies including independent classifiers, multiclass models, and transfer learning.
• Performance was assessed using area under curve (AUC) on internal, locked external, and leave-one-institution-out test sets, with subgroup analyses by institution and body weight.

Main Results:

• EfficientDet-B2 models achieved strong internal AUCs (0.851 for intussusception, 0.834 for splenomegaly) and locked external AUCs (0.818 and 0.806, respectively).
• Leave-one-institution-out validation demonstrated robust performance (AUCs 0.832-0.912).
• Multiclass training improved intussusception detection, while transfer learning enhanced splenomegaly recognition, with stable performance across institutions and weight groups (except <10kg).

Conclusions:

• Multi-institutional deep learning models demonstrate robust performance in detecting intussusception and splenomegaly on pediatric abdominal radiographs.
• These AI models show significant potential to support and improve the triage of pediatric abdominal emergencies.
• Further prospective validation in real-world clinical workflows is recommended to confirm their utility.