Performance of chest CT-based artificial intelligence models in distinguishing pulmonary mucormycosis, invasive pulmonary aspergillosis, and pulmonary tuberculosis

Abstract In clinical practice, differentiating among pulmonary mucormycosis (PM), invasive pulmonary aspergillosis (IPA), and pulmonary tuberculosis (PTB) can be challenging. We aimed to evaluate the performance of chest CT-based artificial intelligence (AI) models in distinguishing among these three diseases. Patients with confirmed PM, IPA, or PTB were retrospectively recruited from three tertiary hospitals. Two models were developed: the unannotated supervised training (UST) model, trained with original CT images and the annotated supervised training (AST) model, trained with manually annotated lesion images. A network questionnaire with 20 cases was designed to assess the performance of clinicians. Sensitivity, specificity, and accuracy were calculated for both models and clinicians. A total of 61 PM cases, 136 IPA cases, and 155 PTB cases were included in the study. In the internal validation set, both models had an accuracy of 66.1%. The UST model had sensitivities of 27.3%, 73.9%, and 76.0% for PM, IPA, and PTB, while the AST model had sensitivities of 9.1%, 69.6%, and 88.0% for the same conditions. In the external validation set, both models had an accuracy of 57.6%. The UST model had sensitivities of 0, 85.7%, and 53.3% for PM, IPA, and PTB, respectively, while AST model had sensitivities of 0, 42.9% and 83.3%. A total of 112 clinicians had an accuracy of 42.9%, with sensitivities of 31.5%, 43.4%, and 48.0% for PM, IPA, and PTB. We demonstrated that two AI models showed comparable performance in diagnosing three diseases. Both models effectively detected IPA and PTB but struggled to identify PM.