Artificial Intelligence vs. Doctors: Diagnosing Necrotizing Enterocolitis on Abdominal Radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neonatal Surgery

Background:

• Necrotizing enterocolitis (NEC) diagnosis via radiography is challenging.
• Deep learning models show potential in improving diagnostic accuracy by detecting subtle imaging patterns.
• This study aimed to compare the diagnostic performance of a DCNN with senior surgical residents.

Purpose of the Study:

• To evaluate the accuracy of a deep convolutional neural network (DCNN) in diagnosing necrotizing enterocolitis (NEC) from neonatal abdominal radiographs.
• To compare the DCNN's performance against that of senior surgical residents in identifying pneumatosis, a key indicator of NEC.
• To assess the potential of AI in assisting clinical practice for prompt NEC identification.

Main Methods:

• A cohort of 494 neonatal abdominal radiographs was used, with 214 images showing NEC.
• A ResNet-50 DCNN was fine-tuned using transfer learning on a training/validation/test set.
• The DCNN's performance was evaluated using AUROC and compared to senior surgical residents using DeLong's method.

Main Results:

• The DCNN achieved an AUROC of 0.918 and 87.8% accuracy in identifying pneumatosis.
• Grad-CAM heatmaps indicated the DCNN focused on relevant image regions.
• The DCNN's performance was comparable to senior surgical residents (p > 0.05).

Conclusions:

• A DCNN trained for pneumatosis recognition can effectively assist clinicians in the rapid diagnosis of NEC.
• The AI model demonstrated comparable accuracy to experienced surgical residents.
• This technology offers a promising tool for improving NEC detection in clinical settings.