AI Model Reads Chest Scans to Tell Fungal Infection and Tuberculosis Apart in HIV Patients

Doctors treating people with HIV/AIDS in Southeast Asia and southern China face a diagnostic dilemma that can cost weeks of precious time: the lung-dwelling fungus Talaromyces marneffei and Mycobacterium tuberculosis, the bacterium behind pulmonary tuberculosis, produce symptoms that overlap so heavily — prolonged fever, weight loss, cough, swollen lymph nodes — that clinicians frequently cannot tell them apart until laboratory cultures come back. A new study from Nanning, in China’s Guangxi region, reports that an artificial intelligence system combining clinical data, quantitative image features and deep learning on chest CT scans can make that distinction with substantially higher accuracy than any single data source alone, achieving an area under the receiver operating characteristic curve of 0.931 and a specificity of 0.981 on an independent test set.

The research, published in Frontiers in Cellular and Infection Microbiology, was conducted by a team at the Fourth People’s Hospital of Nanning working with Guangxi Medical University. It enrolled 380 HIV/AIDS patients with pulmonary infections admitted between January 2018 and August 2025, of whom 203 had confirmed pulmonary tuberculosis and 177 had confirmed pulmonary T. marneffei infection. Every diagnosis was etiologically verified: tuberculosis cases required a positive Xpert MTB/RIF assay, a positive culture, or matching histopathology and microbiological evidence, while fungal cases required microscopic detection of characteristic yeast cells, culture growth, or histopathology confirmed by culture or molecular testing. Patients with other opportunistic coinfections, non-infectious lung disease, or prior targeted anti-infective therapy before imaging were excluded, ensuring a clean two-way comparison.

The rationale for turning to computation lies in the weaknesses of current practice. Culture remains the gold standard for T. marneffei, but fungal growth can take up to 14 days, and the method may miss as many as half of infections. Microscopic and histopathologic examination of skin lesions, lymph nodes or bone marrow can give a preliminary answer, but in patients without localized lesions these approaches are invasive and impractical. In the interval, physicians often start empirical anti-tuberculosis or antifungal therapy and judge success by treatment response — a strategy that increases diagnostic uncertainty and exposes patients to drugs they may not need. Chest CT, meanwhile, is widely available and non-invasive, but the imaging appearance of the two diseases — nodules, consolidation, cavities — overlaps considerably, and visual interpretation depends heavily on radiologist experience.

The team built three kinds of models and then fused them. The clinical model used variables that emerged as independent predictors in multivariate regression: fever, CD4+ T lymphocyte count, platelet count, hemoglobin level, diabetes and hypertension. Notably, patients with the fungal infection had far lower CD4 counts (median 17 versus 127 cells/μL), lower platelets (137 versus 219 ×10⁹/L) and lower hemoglobin (81 versus 102 g/L) than the tuberculosis group. A Random Forest classifier on these clinical features reached a test-set AUC of 0.856, the best among five machine learning algorithms tested.

The second model, radiomics, extracted 1,561 quantitative features from three-dimensional volumes of interest drawn by two thoracic radiologists blinded to the diagnosis. These features captured lesion morphology, first-order intensity statistics and texture heterogeneity through matrices such as the Gray Level Co-occurrence Matrix and Gray Level Run Length Matrix. After a hierarchical selection pipeline — statistical filtering, removal of redundant features with correlations above 0.9, minimum Redundancy Maximum Relevance ranking, and LASSO-regularized logistic regression with 10-fold cross-validation — 11 features remained. An XGBoost model built on them achieved a test-set AUC of 0.866, with the smallest gap between training and test performance among the algorithms compared.

The third and most technically novel component was a 2.5-dimensional deep learning model. Rather than analyzing a single CT slice in isolation or running a computationally expensive 3D convolutional network, the approach extracted seven slices per patient: the axial slice with the largest lesion cross-section, plus slices at offsets of ±1, ±2 and ±4. Each slice was passed through a ResNet-18 network pretrained on ImageNet — a choice that outperformed heavier architectures, since ResNet-50, despite a training-set AUC of 0.937, dropped to 0.782 on the test set, a signature of overfitting. ResNet-18 achieved a slice-level AUC of 0.865 with a specificity of 0.918. The slice-level outputs were then aggregated into a patient-level prediction using multi-instance learning, employing two strategies: a Predict Likelihood Histogram approach that compiled slice probabilities into normalized distributions, and a Bag of Words approach that encoded recurring prediction patterns weighted by term frequency–inverse document frequency. A support vector machine operating on these aggregated features reached a test-set AUC of 0.885, the best patient-level performance, with strong generalization from a training AUC of 0.944.

To visualize where the network was looking, the researchers applied Grad-CAM to the final convolutional layer. In representative cases, activation concentrated within and around a cavitary lesion in a tuberculosis patient and over nodular infiltrates in a fungal patient, suggesting the model attends to clinically relevant regions rather than background anatomy — though the authors caution this does not prove the network recognizes specific pathological features such as cavity-wall irregularity or satellite nodules.

The final multimodal model concatenated the clinical variables, the 11 radiomic features and the deep learning signature into a single feature vector fed to a linear support vector machine, with probability calibration fitted exclusively on training-cohort predictions. On the locked test set of 114 patients, it achieved an AUC of 0.931 with 95% confidence intervals of 0.888 to 0.975, sensitivity of 0.717 and specificity of 0.981. DeLong tests showed the combined model significantly outperformed the clinical model (ΔAUC = 0.075, P = 0.020) and the radiomics model (ΔAUC = 0.065, P = 0.008), but the improvement over the deep learning model alone (ΔAUC = 0.046, P = 0.116) did not reach statistical significance. Calibration was strong, with the combined model posting the lowest Brier score on the test set (0.089), and decision curve analysis showed the highest net benefit across threshold probabilities from roughly 0.10 to 0.90.

The authors emphasize an important trade-off. The combined model’s very high specificity means that when it predicts tuberculosis, it is almost always right — sparing patients unnecessary antifungal therapy and its associated toxicities, which matters especially in resource-limited settings. But its moderate sensitivity of 0.717 means 17 fungal cases in the test set were missed, so clinicians must remain vigilant when the model indicates tuberculosis yet suspicion for T. marneffei persists. Threshold tuning, the authors suggest, could shift this balance depending on the clinical context: lower thresholds favor sensitivity to avoid missing fungal disease, higher ones favor specificity to avoid needless treatment. A sex subgroup analysis found no significant performance difference between men and women (P = 0.672), but with only 25 women in the test set the analysis was too underpowered to draw firm conclusions.

Limitations frame how far these results can travel. The study was single-center and retrospective, with internal validation only; generalizability to other regions, populations or CT protocols remains unproven, and prospective external validation will be needed before clinical deployment. The binary design — excluding coinfections, mixed infections and other opportunistic pathogens — ensured a homogeneous cohort for algorithm development but means conclusions cannot be extended to the messier reality of polymicrobial lung disease. The model also demonstrates discrimination, not outcomes: whether it actually shortens time to diagnosis or improves survival is untested. And the manual segmentation of lesion volumes by radiologists, while reproducible, is time-consuming; the team proposes integrating automated segmentation networks such as nnU-Net in future work. Even so, the study offers a concrete demonstration that fusing host clinical data, handcrafted image texture features and learned spatial representations can push a hard diagnostic boundary in HIV medicine measurably closer to being resolved.

Subject of Research: A CT-based multimodal 2.5D deep learning model for differentiating pulmonary Talaromyces marneffei infection from pulmonary tuberculosis in HIV/AIDS patients

Article Title: Differential diagnosis of pulmonary Talaromyces marneffei infection and pulmonary tuberculosis in HIV/AIDS patients using a CT-based multimodal 2.5D deep learning approach

Article References: Song, C., Zhao, C.-Y., Lin, X.-S., Huang, Z.-T., Huang, X.-W., Song, S.-L., Xie, Z.-H., & Zhu, Q.-D. (2026). Differential diagnosis of pulmonary Talaromyces marneffei infection and pulmonary tuberculosis in HIV/AIDS patients using a CT-based multimodal 2.5D deep learning approach. Frontiers in Cellular and Infection Microbiology, 16, Article 1927286. https://doi.org/10.3389/fcimb.2026.1927286

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1927286

Keywords: Talaromyces marneffei, pulmonary tuberculosis, HIV/AIDS, chest CT, deep learning, radiomics, multi-instance learning, ResNet18, differential diagnosis, opportunistic infection, machine learning, Guangxi