Transformer-based deep learning on chest CT stratifies risk of refractory Mycoplasma pneumonia in children

A transformer deep learning model analyzing non-contrast chest CT achieved external test AUCs of 0.89 (0.84-0.94) and 0.89 (0.82-0.95) for identifying children with refractory Mycoplasma pneumoniae pneumonia, outperforming a clinical model (p<0.001).

• The trans-DLF significantly outperformed a clinical model (p<0.001) and was non-inferior to a multimodal nomogram, suggesting a streamlined imaging-only approach.
• In an outpatient subcohort, the model achieved an AUC of 0.87 with good calibration and net clinical benefit.
• Gradient-weighted class activation mapping indicated consolidations were key imaging features driving predictions.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.