Baseline chest CT-based deep learning for predicting subsequent spinal tuberculosis in patients with pulmonary tuberculosis: a two-center retrospective study

To develop and internally evaluate a deep learning model using pretreatment baseline chest CT to classify documented subsequent spinal tuberculosis (STB) status during available follow-up in patients with pulmonary tuberculosis (PTB). This two-center retrospective study included 1,668 consecutive eligible patients with PTB and no documented STB at baseline. Records from January 2010 to October 2025 provided up to 15 years of follow-up. All eligible patients were retained without outcome-based sampling. The observed 25% STB fraction was not imposed by the study design. Patients were stratified by center and outcome into training (n = 1,168), validation (n = 252), and test (n = 248) sets. ResNet18, AlexNet, DenseNet121, and ResNet50 were compared using single-slice CT regions of interest. Architecture and classification threshold selection used validation data. The cut-off was fixed using the validation-set Youden index before test evaluation. No clinical-model comparison was performed. Calibration and decision curve analysis (DCA) were exploratory because follow-up varied and no prediction horizon was prespecified. The cohort included 417 patients with documented subsequent STB and 1,251 without. ResNet18 achieved the highest validation-set area under the ROC curve (AUC) and was selected for final evaluation. Its training, validation, and test AUCs were 0.858 (95% CI 0.833–0.883), 0.733 (95% CI 0.656–0.811), and 0.784 (95% CI 0.720–0.848), respectively. Test-set accuracy, sensitivity, specificity, precision, and F1-score were 0.746, 0.694, 0.763, 0.494, and 0.577. Test AUC was significantly higher than ResNet50, but not AlexNet or DenseNet121, after multiplicity correction. Calibration was imperfect. Qualitative Grad-CAM examples showed activation in pulmonary lesion-related and adjacent extrapulmonary regions. Baseline chest CT-based deep learning showed moderate discrimination for documented subsequent STB during available follow-up, rather than risk at a defined horizon. External validation with standardized outcome ascertainment and complete event and censoring dates is required before clinical application.

Authors

Institutions

Publication Details

Journal
European journal of medical research
Published
2026-11-21
DOI
https://doi.org/10.1186/s40001-026-05519-x
Primary Topic
Infectious Diseases and Tuberculosis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Baseline chest CT-based deep learning for predicting subsequent spinal tuberculosis in patients with pulmonary tuberculosis: a two-center retrospective study

Saimaitikari Abudoubari, Abudouresuli Tuersun, Mireayi Tudi, Xin Qiao et al.
1 links
European journal of medical research
Infectious Diseases and Tuberculosis
article

Baseline chest CT-based deep learning for predicting subsequent spinal tuberculosis in patients with pulmonary tuberculosis: a two-center retrospective study

Saimaitikari Abudoubari, Abudouresuli Tuersun, Mireayi Tudi, Xin Qiao, Nuermaimaijiang Abudouwufu, Lijuan Da, Yi You, Ya Qiu
article en

Abstract

To develop and internally evaluate a deep learning model using pretreatment baseline chest CT to classify documented subsequent spinal tuberculosis (STB) status during available follow-up in patients with pulmonary tuberculosis (PTB). This two-center retrospective study included 1,668 consecutive eligible patients with PTB and no documented STB at baseline. Records from January 2010 to October 2025 provided up to 15 years of follow-up. All eligible patients were retained without outcome-based sampling. The observed 25% STB fraction was not imposed by the study design. Patients were stratified by center and outcome into training (n = 1,168), validation (n = 252), and test (n = 248) sets. ResNet18, AlexNet, DenseNet121, and ResNet50 were compared using single-slice CT regions of interest. Architecture and classification threshold selection used validation data. The cut-off was fixed using the validation-set Youden index before test evaluation. No clinical-model comparison was performed. Calibration and decision curve analysis (DCA) were exploratory because follow-up varied and no prediction horizon was prespecified. The cohort included 417 patients with documented subsequent STB and 1,251 without. ResNet18 achieved the highest validation-set area under the ROC curve (AUC) and was selected for final evaluation. Its training, validation, and test AUCs were 0.858 (95% CI 0.833–0.883), 0.733 (95% CI 0.656–0.811), and 0.784 (95% CI 0.720–0.848), respectively. Test-set accuracy, sensitivity, specificity, precision, and F1-score were 0.746, 0.694, 0.763, 0.494, and 0.577. Test AUC was significantly higher than ResNet50, but not AlexNet or DenseNet121, after multiplicity correction. Calibration was imperfect. Qualitative Grad-CAM examples showed activation in pulmonary lesion-related and adjacent extrapulmonary regions. Baseline chest CT-based deep learning showed moderate discrimination for documented subsequent STB during available follow-up, rather than risk at a defined horizon. External validation with standardized outcome ascertainment and complete event and censoring dates is required before clinical application.

European journal of medical research
Kashi University (CN)
Natural Science Foundation of Xinjiang Province
Good health and well-being
Infectious Diseases and Tuberculosis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon