A deep learning model trained on pretreatment baseline chest computed tomography scans demonstrated moderate accuracy in identifying subsequent spinal tuberculosis among patients initially diagnosed with pulmonary tuberculosis, according to a two-center retrospective study published in the European Journal of Medical Research. The researchers sought to determine whether computational image analysis could classify subsequent spinal involvement during available follow-up rather than estimating risk across a prespecified prospective timeline.
To examine potential radiological markers at diagnosis, the investigative team evaluated pretreatment baseline chest scans from eligible individuals who presented with active pulmonary infection but had no documented spinal involvement at their initial clinical assessment. The primary goal focused on developing and internally validating an artificial intelligence system capable of distinguishing patients who later developed documented spinal complications from those who did not.
The retrospective cohort comprised 1,668 consecutive eligible patients gathered across two medical centers, utilizing institutional records spanning from January 2010 to October 2025. Across this observation interval, clinical follow-up extended up to 15 years. The dataset included 417 patients with documented subsequent spinal tuberculosis and 1,251 patients without documented spinal involvement, reflecting an observed 25 percent disease fraction that emerged naturally from consecutive enrollment rather than outcome-based sampling.
To structure the experimental pipeline, the cohort was stratified by clinical center and outcome status into a training set of 1,168 patients, an internal validation set of 252 patients, and an independent test set of 248 patients. The investigators compared four established convolutional neural network architectures—ResNet18, AlexNet, DenseNet121, and ResNet50—using single-slice regions of interest extracted from the pretreatment chest scans. Model selection and classification threshold determination were conducted strictly using validation data, where the decision cut-off was fixed via the Youden index prior to evaluating the held-out test cohort.
Among the evaluated architectures, ResNet18 achieved the highest discrimination during validation and was selected for primary evaluation. ResNet18 attained an area under the receiver operating characteristic curve of 0.858 with a 95 percent confidence interval of 0.833 to 0.883 in the training cohort, 0.733 with a 95 percent confidence interval of 0.656 to 0.811 in the validation cohort, and 0.784 with a 95 percent confidence interval of 0.720 to 0.848 in the final test set.
Within the test evaluation, ResNet18 achieved an overall accuracy of 0.746, a sensitivity of 0.694, a specificity of 0.763, a precision of 0.494, and an F1-score of 0.577. After applying multiplicity corrections, the test area under the curve for ResNet18 was significantly superior to ResNet50, though it did not demonstrate statistically significant superiority when compared directly against AlexNet or DenseNet121.
Exploratory interpretability assessments using qualitative Gradient-weighted Class Activation Mapping revealed that model attention localized to pulmonary lesion-related regions as well as adjacent extrapulmonary thoracic areas. However, the authors noted several important constraints, including imperfect statistical calibration and the exploratory nature of decision curve analyses arising from variable follow-up durations without a prespecified prediction horizon. Furthermore, the protocol did not include a comparison against standard clinical prediction models, leading the authors to conclude that rigorous external validation with standardized outcome ascertainment and precise event dates remains necessary before clinical deployment.
Subject of Research: Deep learning prediction of subsequent spinal tuberculosis from baseline chest CT scans
Article Title: Baseline chest CT-based deep learning for predicting subsequent spinal tuberculosis in patients with pulmonary tuberculosis: a two-center retrospective study
Article References: Tuersun, A., Tudi, M., Da, L., Abudoubari, S., You, Y., Qiao, X., Abudouwufu, N., & Qiu, Y. (2026). Baseline chest CT-based deep learning for predicting subsequent spinal tuberculosis in patients with pulmonary tuberculosis: a two-center retrospective study. European Journal of Medical Research. https://doi.org/10.1186/s40001-026-05519-x
Image Credits: AI Generated
DOI: 10.1186/s40001-026-05519-x
Keywords: pulmonary tuberculosis, spinal tuberculosis, deep learning, chest CT, ResNet18, medical imaging, retrospective study, receiver operating characteristic, extrapulmonary tuberculosis, convolutional neural networks, internal validation, Grad-CAM, diagnostic radiology, image classification, clinical decision support, computed tomography
Cite Scienmag News
Ophelia Keating. (November 21, 2026). Deep Learning on Baseline Chest CT Predicts Subsequent Spinal Tuberculosis. Scienmag. https://scienmag.com/deep-learning-on-baseline-chest-ct-predicts-subsequent-spinal-tuberculosis/
Ophelia Keating. "Deep Learning on Baseline Chest CT Predicts Subsequent Spinal Tuberculosis." Scienmag, 21 November 2026, https://scienmag.com/deep-learning-on-baseline-chest-ct-predicts-subsequent-spinal-tuberculosis/. Accessed 21 November 2026.
Ophelia Keating. "Deep Learning on Baseline Chest CT Predicts Subsequent Spinal Tuberculosis." Scienmag. November 21, 2026. https://scienmag.com/deep-learning-on-baseline-chest-ct-predicts-subsequent-spinal-tuberculosis/








