Deep learning feature-based model on abdominal radiography outperforms experts for early necrotizing enterocolitis...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neonatal Care

Background:

• Plain abdominal radiography is standard for diagnosing neonatal necrotizing enterocolitis (NEC).
• Early-stage NEC diagnosis is challenging due to subtle radiographic features.
• Deep learning (DL) offers potential for enhancing diagnostic accuracy.

Purpose of the Study:

• To evaluate the efficacy of DL models in the early diagnosis of stage I NEC using plain abdominal radiography.
• To compare the diagnostic performance of DL models against human experts.

Main Methods:

• Retrospective analysis of 680 neonates across two centers (June 2016 - December 2023).
• DL features extracted using DenseNet121; radiomics models built with logistic regression (LR) and random forest (RF).
• Performance assessed via ROC curves in training and external validation cohorts; direct comparison with human expert performance.

Main Results:

• DL models achieved high accuracy in both training (LR AUC: 0.972, RF AUC: 0.961) and validation (LR AUC: 0.964, RF AUC: 0.951) cohorts.
• The DL models significantly outperformed human experts in diagnosing stage I NEC.
• 25 DL features were identified as key for model development.

Conclusions:

• DL models utilizing plain abdominal radiography are effective for identifying stage I NEC in neonates.
• This non-invasive approach enhances early NEC diagnosis and supports clinical decision-making.
• AI-powered imaging analysis shows promise for improving neonatal care outcomes.