Differential diagnosis of pulmonary <i>Talaromyces marneffei</i> infection and pulmonary tuberculosis in HIV/AIDS patients using a CT-based multimodal 2.5D deep learning approach.

Authors

Abstract

To investigate the diagnostic value of a computed tomography (CT)-based multimodal 2.5-dimensional (2.5D) deep learning model for differentiating Talaromyces marneffei (TSM) from pulmonary tuberculosis (PTB) in patients with human immunodeficiency virus infection/acquired immunodeficiency syndrome (HIV/AIDS). A total of 380 HIV/AIDS patients with pulmonary infections who were admitted to the Fourth People's Hospital of Nanning from January 2018 to August 2025 were retrospectively enrolled, including 203 patients with PTB, 177 with TSM. The cohort was randomly divided into a training set and a test set at a ratio of 7:3. A clinical model, a radiomics model, and a 2.5D deep learning model were constructed separately. In the clinical feature model, Random Forest achieved the optimal performance, with a test set AUC of 0.856. In the radiomics model, XGBoost attained the highest AUC of 0.866. In the 2.5D deep learning model, ResNet18 performed best at the slice level with an AUC of 0.865, and the patient-level support vector machine (SVM) model based on MIL achieved an AUC of 0.885. The multimodal combined model integrating clinical, radiomics, and deep learning features achieved the highest AUC of 0.931 and a specificity of 0.981. It significantly outperformed the clinical model (P = 0.020) and the radiomics model (P = 0.008), but the improvement over DL25D was not statistically significant (ΔAUC = 0.046; P = 0.116). The high specificity of the combined model should be interpreted alongside its sensitivity (0.717) and the selected decision threshold. The CT-based multimodal 2.5D deep learning model effectively integrates clinical indicators, radiomics features, and deep learning representations, and substantially improves the differentiation of TSM from PTB in HIV/AIDS patients.

Tags

Topics