NeoCLIP: a self-supervised foundation model for the interpretation of neonatal radiographs

Area of Science:

• Neonatal medicine
• Artificial intelligence in healthcare
• Medical imaging analysis

Background:

• Deep learning excels in adult and pediatric medical image interpretation.
• Current deep learning applications for neonates remain limited.
• Neonatal radiographs require specialized interpretation due to unique anatomical and physiological considerations.

Purpose of the Study:

• To develop and evaluate NeoCLIP, a novel deep contrastive learning model for neonatal radiograph interpretation.
• To detect pathologies and medical devices in neonatal intensive care unit (NICU) patients.
• To establish a benchmark for AI-driven analysis of neonatal medical imaging.

Main Methods:

• Retrospective study of 4629 infants from a NICU (2008-2023) with 20,154 radiographs and 15,795 reports.
• Development of NeoCLIP, a deep contrastive learning model trained on neonatal radiographs.
• Model trained to identify 15 radiological features and 5 medical devices, with data split into training, validation, and test sets.

Main Results:

• NeoCLIP demonstrated superior performance (higher AUROC) compared to controls across most labels, excluding portal venous gas.
• The model successfully identified key radiological features and medical devices relevant to neonatal care.
• Inclusion of demographic data showed a trend towards improved performance, though not statistically significant.

Conclusions:

• NeoCLIP is the first deep learning model specifically designed for neonatal radiograph interpretation.
• The model shows promising efficacy, outperforming comparable models developed for adult populations.
• NeoCLIP represents a significant advancement in applying AI to critical neonatal medical imaging analysis.