Establishment and evaluation of a CT-based radiomic model for AIDS-associated pulmonary cryptococcosis.

BACKGROUND

Establish a CT-based diagnostic radiomic model for AIDS complicated with pulmonary cryptococcosis and evaluate the diagnostic efficacy of this model.

METHODS

This retrospective study enrolled 98 AIDS patients with pulmonary cryptococcosis and 103 AIDS patients with other infections or neoplastic lesions, comprising a total of 699 lesions. Patients were randomly divided into a training group and test group at a ratio of 2.75:1. Features from all lesions, cavity lesions and solid nodule lesions were extracted, and two kinds of radiomic models (6 types) were established. ROC curves were drawn, and the sensitivity and specificity were calculated to compare the SVM model and LR model, radiologists' empirical diagnoses and the combination of these empirical diagnoses with the radiomic model.

RESULTS

The AUCs of senior radiologist for all lesions and cavity lesions were lower than those of the SVM and LR models. The diagnostic efficacy of primary radiologist was lower than that of both of the other model types. The diagnostic efficacy of the LR model was relatively stable, with the highest diagnostic efficiency of the 3 model/radiologist groups. The AUCs of intermediate radiologist in combination with the LR radiomic model for all lesions, nodular lesions and cavity lesions were 0.88, 0.84, and 0.9, respectively, which were the highest among all models and radiologists.

CONCLUSIONS

The CT-based radiomic LR model of AIDS-associated pulmonary cryptococcosis exhibits good diagnostic performance, which was similar to that of senior radiologists and higher than that of the primary radiologist. With the help of a radiomic model, radiologists can achieve improved diagnostic accuracy compared to that when only an empirical diagnosis is used.