Deep learning detects retropharyngeal edema on MRI in patients with acute neck infections

PMCID: PMC12179047

PMID: 40536731

DOI: 10.1186/s41747-025-00599-6

Journal: European radiology experimental

Publication Date: 2025-6-19

Authors: Rainio O, Huhtanen H, Vierula JP, Nurminen J, Heikkinen J, et al.

Key Points

• Deep learning algorithm achieved 87.4% accuracy in detecting retropharyngeal edema at the patient level
• AUROC values of 94.8% (patient level) and 94.1% (slice level) indicate exceptional diagnostic performance
• Automated MRI interpretation could help radiologists quickly identify high-risk patients in emergency settings

Summary

This study developed a deep learning algorithm for automated detection of retropharyngeal edema (RPE) in acute neck infections using magnetic resonance imaging (MRI). The researchers created a convolutional neural network (CNN) that could accurately classify RPE at both slice and patient levels, achieving remarkable diagnostic performance with high accuracy, sensitivity, and specificity.

The deep learning model demonstrated exceptional diagnostic capabilities, with area under the receiver operating characteristic curve (AUROC) of 94.8% at the patient level and 94.1% at the slice level. Notably, the proposed CNN was computationally efficient, requiring only 3.3% of the training time compared to standard models like InceptionV3, while maintaining equivalent diagnostic accuracy. This approach offers a promising automated triage tool for emergency radiology departments, potentially improving early detection of patients at high risk for severe acute neck infections.