aceA CT-based deep learning model for the automated risk stratification of refractory Mycoplasma pneumoniae pneumonia in children
The accurate identification of children with refractory Mycoplasma pneumoniae pneumonia (RMPP) remains challenging. This study aimed to develop a transformer-based model utilizing clinically indicated chest computed tomography (CT) to stratify pediatric RMPP risk at a critical decision point. Non-contrast chest CT data from a multicenter retrospective cohort of 1224 pediatric patients with Mycoplasma pneumoniae pneumonia who underwent clinically indicated CT were used to develop a transformer-based deep learning…
The evolution of pediatric literature in the United States by Adams, Samuel Shugert, 1853-1928, author American Pediatric Society Cleveland Public Library, former owner. Public domain
A multicenter retrospective study developed a transformer-based deep learning framework to stratify risk of refractory Mycoplasma pneumoniae pneumonia in children using non-contrast chest CTs obtained for clinical indications. The model was trained on 506 cases and tested internally on 139 and externally on 331 and 108 cases.
The approach matters because early identification of refractory disease can guide timely treatment decisions without extra testing, leveraging imaging already acquired. Uncertainty remains about prospective performance, applicability to children without a clinical CT indication, and integration into routine pediatric workflows.
- Study used 1224 pediatric patients with Mycoplasma pneumoniae pneumonia across primary cohort of 785 and two external cohorts of 331 and 108.
- Median age was 6.83 years and 609 (49.8%) were male.
- Model was compared against a 3D-CNN, a clinical model, and a multimodal nomogram, significantly outperforming the clinical model.
- Interpretability via Grad-CAM suggested predictions were influenced by consolidations.
A transformer-based deep learning framework using clinically indicated non-contrast chest CT stratified risk of refractory Mycoplasma pneumoniae pneumonia in children with AUCs around 0.89-0.90 on internal and external test cohorts.
The rundown
Researchers built trans-DLF from non-contrast chest CTs in 1224 children with Mycoplasma pneumoniae pneumonia, splitting a primary cohort into training (n=506), validation (n=140), and internal testing (n=139) plus two external test sets. Performance was measured by AUC and benchmarked against a 3D-CNN, clinical model, and multimodal nomogram.
Results showed AUC 0.97 in training, 0.91 in validation, 0.90 in internal testing, and 0.89 in both external cohorts, with maintained performance in outpatient settings (AUC 0.87), good calibration and net clinical benefit. Grad-CAM analysis pointed to consolidations as influential features.
Sources
- Peer-reviewedBMC Medical Imaging2026-07-17
How should this claim be treated?
The debate