Development and Validation of a CT-based Deep Transfer Learning Radiomic Model for Predicting Post-COVID-19 Pulmonary Fibrosis

objectiveTo assess the efficacy of a deep transfer learning radiomic (DLR) model in predicting the 12-month risk of pulmonary fibrosis after COVID-19, which is crucial for early clinical intervention and treatment.

methodsRetrospective analysis of 260 COVID-19 patients (Dec 2022-Jan 2023) included chest CT and clinical data collection during hospitalization, with 1-year radiological follow-up. Final follow-up CT determined fibrosis status. ResNet-50-based DLR automatically segmented lesions from initial CTs. Radiomic features were extracted, filtered via Pearson’s correlation, and LASSO. Clinical predictors were identified through univariate/multivariate analyses. Seven classifiers built clinical, radiomic, DL, DLR, and nomogram models. Evaluation of performance was based on the AUC, calibration curves, and DCA, and model comparisons were made using DeLong’s test.

resultsAge and hospital stay duration were independent fibrosis predictors. Logistic regression outperformed other classifiers. The nomogram achieved test-set AUC=0.868 (95%CI:0.783–0.952), significantly surpassing the clinical feature-based model (p<0.05) but not the DLR model. DCA indicated higher clinical net benefits for DLR and the nomogram. DISCUSSION: The DLR model, based on automated segmentation technology, has advanced the prediction of post-COVID-19 fibrosis and provided an effective predictive tool for clinical radiology.

conclusionThe DLR model, leveraging automated CT lesion segmentation, effectively predicts post-COVID-19 pulmonary fibrosis, offering robust clinical utility. The nomogram integrating clinical and radiomic data enhances risk stratification but does not significantly improve upon the standalone DLR model.