Detection of neonatal pneumoperitoneum on radiographs using deep multi-task learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neonatal Care

Background:

• Neonatal pneumoperitoneum presents subtle radiographic signs, challenging emergency diagnosis.
• Prompt diagnosis is critical for this life-threatening condition.

Purpose of the Study:

• Develop and validate a deep multi-task learning model for neonatal pneumoperitoneum detection.
• Assess the model's clinical utility and impact on clinician diagnostic performance.

Main Methods:

• Retrospective diagnostic study with internal and external datasets from multiple neonatal intensive care units.
• A deep multi-task learning model combining classification and segmentation for pneumoperitoneum detection.
• Reader study comparing AI-assisted and unassisted performance among physicians with varying experience levels.

Main Results:

• The AI model achieved high diagnostic performance: AUC of 0.98 (internal) and 0.89 (external).
• AI assistance significantly improved reader accuracy (82.5% to 86.6%) and inter-reader agreement (kappa from 0.33 to 0.71).
• The model demonstrated excellent diagnostic performance and enhanced clinician diagnostic capabilities.

Conclusions:

• The developed deep multi-task learning model shows excellent diagnostic performance for neonatal pneumoperitoneum.
• AI assistance improves diagnostic accuracy and agreement among clinicians in neonatal intensive care settings.
• This AI tool has the potential to enhance the clinical management of neonatal pneumoperitoneum.