Deep learning algorithms for detecting and visualising intussusception on plain abdominal radiography in children: a...

Area of Science:

• Pediatric Radiology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Intussusception is a common surgical emergency in young children.
• Accurate and timely diagnosis of intussusception is crucial for effective treatment.
• Current diagnostic methods can be challenging, necessitating improved screening tools.

Purpose of the Study:

• To validate a deep convolutional neural network (CNN) algorithm for detecting intussusception.
• To assess the diagnostic performance of the CNN using a large dataset of pediatric abdominal X-rays.
• To evaluate the algorithm's efficacy in screening for intussusception in children.

Main Methods:

• Utilized a dataset of 1449 pediatric abdominal X-rays with intussusception and 9935 without.
• Employed Single Shot MultiBox Detector for abdominal detection and ResNet for classification.
• Performed internal and external validation to analyze diagnostic performance.

Main Results:

• Achieved high internal validation performance with Area Under the Curve (AUC) of 0.935 and Youden Index of 0.743.
• Internal test AUC ranged from 0.946 to 0.971, with accuracy up to 0.952.
• External validation showed lower but significant diagnostic values, indicating algorithm generalizability.

Conclusions:

• Deep CNN algorithm demonstrates significant potential for detecting intussusception in children.
• Plain abdominal X-rays analyzed by deep CNN can serve as an effective screening tool.
• This AI-driven approach may aid in earlier and more accurate diagnosis of pediatric intussusception.