Deep learning based detection of active pulmonary tuberculosis from chest radiographs in a retrospective prospective multicenter study

Summary

Active tuberculosis remains a major global health problem, and distinguishing active from non-active disease on chest radiographs could accelerate triage where radiology resources are limited. We developed ShuffleNetV2 using 4,613 radiographs from two centers and evaluated it in retrospective external (1,318 patients) and prospective (560 patients) cohorts. The model achieved AUCs of 0.917 and 0.896, respectively, and performance was stable across prespecified demographic and scanner subgroups. In the prospective cohort, artificial-intelligence assistance increased radiologist AUCs to 0.897–0.937 across experience levels. These findings support ShuffleNetV2 as a rapid decision-support tool for prioritizing confirmatory clinical and microbiological assessment; positive predictions should not be used alone to initiate treatment.

Graphical abstract

Highlights

• •ShuffleNetV2 distinguishes active from non-active tuberculosis on chest radiographs
• •A multicenter cohort of 6,491 patients underwent retrospective and prospective validation
• •ShuffleNetV2 achieved AUCs of 0.917 externally and 0.896 prospectively
• •AI assistance improved radiologist performance across experience levels

Artificial intelligence applications; Disease; Radiology

Introduction

Pulmonary tuberculosis (TB), caused by Mycobacterium tuberculosis, remains a major global public health challenge. According to the World Health Organization Global Tuberculosis Report 2025, an estimated 10.7 million people developed TB worldwide in 2024, including 5.8 million men, 3.7 million women, and 1.2 million children and young adolescents. TB caused approximately 1.23 million deaths in 2024, including 150,000 deaths among people living with human immunodeficiency virus (HIV).^1,2 Eight countries accounted for approximately two-thirds of all incident TB cases: India, Indonesia, the Philippines, China, Pakistan, Nigeria, the Democratic Republic of the Congo, and Bangladesh. Despite recent progress, the global reduction in TB incidence remains substantially below the World Health Organization End TB Strategy targets, highlighting the continued need for accessible and timely diagnostic approaches.^3,4,5 Furthermore, the rising prevalence of human immunodeficiency virus (HIV) infection and acquired immunodeficiency syndrome (AIDS) and immunocompromised populations have contributed to a resurgence of TB cases.⁶

Central to TB control efforts is the critical distinction between disease states. Non-active TB infection occurs when M. tuberculosis remains dormant within the host, neither replicating nor causing symptoms. Individuals with non-active TB are non-infectious, though they retain a lifelong risk of progression to active disease.⁷ However, active TB is characterized by the active replication of Mycobacterium TB bacteria, leading to symptomatic disease in infected individuals. Unlike non-active TB, active TB is highly transmissible, posing a significant challenge to public health containment strategies.⁸ Given these distinctions, accurate differentiation between active and non-active TB is essential for guiding treatment decisions and optimizing disease control, particularly in resource-limited settings where misdiagnosis can have severe consequences.

Traditional diagnosis of TB activity status relies on a synthetic assessment of medical history, physical examinations, chest imaging, sputum smear microscopy, mycobacterial culture, and molecular testing. In some cases, invasive procedures such as biopsy or bronchoscopy are required. While sputum smear microscopy and culture remain the diagnostic gold standard for active TB, these methods have limitations: smear microscopy suffers from poor sensitivity (typically 50–60%), while culture requires 2–8 weeks for conclusive results.^7,8,9 This underscores the urgent need for more accurate, rapid, and reliable detection methods in clinical practice.

Deep learning (DL), a subset of artificial intelligence (AI), shows great potential in medical image analysis, disease diagnosis, treatment planning, and prognosis evaluation. Distinct from conventional approaches, DL algorithms obviate the need for manual feature extraction while maintaining computational efficiency suitable for clinical use. Through automated feature learning directly from raw imaging data, DL systems demonstrate superior pattern recognition capabilities without requiring predefined rules or human-engineered features. Recent evidence further supports the use of artificial intelligence-assisted chest radiograph interpretation for TB detection. Han ZL et al.,¹⁰ found that artificial intelligence-based chest radiograph software generally achieved high sensitivity for TB detection and could support rapid screening in settings with limited radiological expertise. However, the study also identified variability in specificity across different systems, emphasizing the importance of external validation and context-specific operating thresholds.

In this study, we developed a deep learning system to distinguish active TB from non-active TB using chest radiographs (CXRs) (hereafter, ShuffleNetV2). Through rigorous validation with both retrospective and prospective cohorts, we assessed the model’s diagnostic performance and generalizability across diverse clinical settings. We also evaluated the model’s diagnostic performance and assessed its impact on radiologists' interpretation accuracy both with and without AI assistance.

Results

Baseline patient characteristics

Table 1 shows the baseline characteristics of patients. The study flow diagram detailing inclusion and exclusion criteria is presented in Figure 1. A total of 5,931 patients were included across the four retrospective centers, comprising the training and retrospective external validation cohorts, among which, 2755 (46.5%) patients were diagnosed with active TB and 3176 (53.5%) had non-active TB. This cohort comprised 2924 (49.3%) males and 3007 (50.7%) females. Key clinical manifestations included cough (3553 [59.9%] patients), expectoration (3385 [57.1%] patients), hemoptysis (671 [11.3%] patients), low-grade fever (2789 [47.0%] patients), fatigue (1984 [33.5%] patients), night sweats (1689 [28.5%] patients), dyspnea (286 [4.8%] patients), and weight loss (502 [8.5%] patients). The comprehensive model development and evaluation workflow is illustrated in Figure 2.

Table 1.

Baseline information

Parameter	Training cohort	External validation cohort	Prospective cohort	All	p Value
No. of patients	4613	1318	560	6491
Active	2067(44.8%)	688(52.2%)	159(28.4%)	2914	<0.001
Non-active	2546(55.2%)	630(47.8%)	401(71.6%)	3577
Mean age (y) mean ± SD	60.36 ± 16.44	61.69 ± 15.36	61.51 ± 17.40	60.73 ± 16.32	0.136
No. of men	2260 (49.0%)	664 (50.4%)	285(50.9%)	3209 (49.4%)	0.520
No. of women	2353 (51.0%)	654 (49.6%)	275(49.1%)	3282 (50.6%)
Region source
	Kashgar:2872	Yecheng county:307	Maigaiti county:560	–
	Jiashi county:1741	Shufu county:1011		–
Hemoptysis
No	4107 (89.0%)	1153 (87.5%)	495(88.4%)	5755(88.7%)	0.287
Yes	506 (11.0%)	165 (12.5%)	65(11.6%)	736(11.3%)
Cough
No	1818 (39.4%)	560 (42.5%)	232(41.4%)	2610(40.2%)	0.110
Yes	2795 (60.6%)	758 (57.5%)	328(58.6%)	3881(59.8%)
Low fever					0.228
No	2470 (53.5%)	672 (51.0%)	303(54.1%)	3445(53.1%)
Yes	2143 (46.5%)	646 (49.0%)	257(45.9%)	3046(46.9%)
Dyspnea					0.096
No	4404 (95.5%)	1241 (94.2%)	529(94.5%)	6176(95.1%)
Yes	209 (4.5%)	77 (5.8%)	31(5.5%)	315(4.9%)
Expectoration					0.366
No	2001 (43.4%)	545 (41.4%)	247(44.1%)	2793(43.0%)
Yes	2612 (56.6%)	773 (58.6%)	313(55.9%)	3691(57.0%)
Fatigue					0.307
No	3085 (66.9%)	862 (65.4%)	359(64.1%)	4306(66.3%)
Yes	1528 (33.1%)	456 (34.6%)	201(35.9%)	2185(33.7%)
Night sweats					0.325
No	3318 (71.9%)	924 (70.1%)	409(73.0%)	4651(71.7%)
Yes	1295 (28.1%)	394 (29.9%)	151(27.0%)	1840(28.3%)
Thin					0.437
No	4234 (91.8%)	1195 (90.7%)	512(91.4%)	5941(91.5%)
Yes	379 (8.2%)	123 (9.3%)	48(8.6%)	550(8.5%)

Values are shown as n (%) or mean ± SD.

Figure 1.

Inclusion and exclusion criteria for study participants

A total of 4613 patients (2067 active, 2546 non-active) from two centers underwent model training. An external validation set comprised 1318 patients from two additional centers. Furthermore, the model was prospectively validated in Maigaiti County Hospital using data collected between January 1, 2023, and January 1, 2024.

Figure 2.

Workflow depicts ShuffleNetV2 model development and evaluation

A lightweight ShuffleNet architecture was developed and evaluated using AUC, sensitivity, and specificity metrics. In the prospective cohort, diagnoses by senior, intermediate, and junior radiologists were compared against model predictions.

The prospective validation set comprised 560 patients from MCH, with 285 (50.9%) males and 275 (49.1%) females and a mean age of 60.73 ± 16.32 years. Among these, 159 (28.4%) were classified as active TB cases and 401 (71.6%) as non-active cases. Reported symptoms included cough (n = 328, 58.6%), expectoration (n = 313, 55.9%), low-grade fever (n = 257, 45.9%), hemoptysis (n = 65, 11.6%), fatigue (n = 201, 35.9%), night sweats (n = 157, 28.0%), dyspnea (n = 31, 5.5%), and weight loss (n = 48, 8.6%).

Diagnostic performance of ShuffleNetV2

Table 2 and Figure 3A ShuffleNetV2 demonstrated exceptional discriminatory performance across all three validation cohorts, as illustrated in the comprehensive performance analysis. In the training cohort, the model achieved an area under the curve (AUC) of 0.926 (95% CI: 0.918–0.933), with balanced sensitivity of 0.847 (95% CI: 0.831–0.862) and specificity of 0.851 (95% CI: 0.837–0.864). The confusion matrix (Figure 3B) revealed robust classification accuracy, while calibration analysis (Figure 3C) demonstrated excellent agreement between predicted and observed probabilities across the full spectrum. External validation (Figures 3D) maintained strong performance with an AUC of 0.917 (95% CI: 0.902–0.932), sensitivity of 0.854 (95% CI: 0.826–0.879), and specificity of 0.818 (95% CI: 0.786–0.846), with the confusion matrix (Figure 3E) indicating consistent generalizability. The prospective cohort (Figures 3G) yielded an AUC of 0.896 (95% CI: 0.860–0.932) with sensitivity of 0.862 (95% CI: 0.799–0.907) and specificity of 0.830 (95% CI: 0.791–0.864), demonstrating 333 true negatives, 137 true positives, 68 false positives, and 22 false negatives in the confusion matrix (Figure 3H). The prospective calibration curve (Figure 3I) deviated from the ideal calibration line, particularly in the low-to-intermediate predicted-probability ranges, indicating that the estimated probabilities did not consistently correspond to the observed frequency of active TB. Thus, although discrimination remained favorable, the probability estimates were not optimally calibrated in the prospective cohort.

Table 2.

Performance of ShuffleNetV2 in identifying active tuberculosis from chest radiographs

Dataset	AUC (95%CI)	Accuracy (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)	PPV (95%CI)	NPV (95%CI)
Training set	0.926 (0.918–0.933)	0.849 (0.839–0.859)	0.847 (0.831–0.862)	0.851 (0.837–0.864)	0.822 (0.805–0.838)	0.873 (0.859–0.885)
External validation set	0.917 (0.902–0.932)	0.836 (0.816–0.856)	0.854 (0.826–0.879)	0.818 (0.786–0.846)	0.836 (0.807–0.862)	0.838 (0.807–0.865)
Prospective validation set	0.896 (0.860–0.932)	0.839 (0.807–0.867)	0.862 (0.799–0.907)	0.830 (0.791–0.864)	0.668 (0.601–0.729)	0.938 (0.908–0.959)

AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value.

Figure 3.

Diagnostic performance of ShuffleNetV2 in the training, retrospective external-validation, and prospective-validation cohorts

A–C show receiver operating characteristic curve, confusion matrix, and calibration curve in the training cohort (n = 4,613 patients; AUC 0.926, 95% CI 0.918–0.933). D–F show the corresponding results in the external-validation cohort (n = 1,318 patients; AUC 0.917, 95% CI 0.902–0.932). G–I show the prospective cohort (n = 560 patients; AUC 0.896, 95% CI 0.860–0.932). Here n denotes patients, each contributing one chest radiograph; CI denotes confidence interval, and AUC denotes area under the receiver operating characteristic curve.

Ablation and benchmarking analyses

Systematic ablation experiments evaluated the effects of input resolution, width multiplier, data augmentation, pre-training, and loss function across the training, retrospective external validation, and prospective validation cohorts (Table S1). The selected primary configuration provided the strongest overall prospective discrimination and a balanced sensitivity-specificity profile compared with the tested variants.

In particular, removing pre-training produced the largest reduction in performance, whereas changes in input resolution, width multiplier, augmentation, and loss function produced smaller but measurable changes across cohorts.

We also benchmarked ShuffleNetV2 against ResNet50, DenseNet121, EfficientNet-B0, and ViT-Base/16 under identical experimental conditions (Table S2). In the prospective cohort, ShuffleNetV2 achieved an AUC of 0.896, compared with 0.882 for ResNet50, 0.848 for DenseNet121, 0.687 for EfficientNet-B0, and 0.895 for ViT-Base/16. On retrospective external validation, the AUCs were 0.917, 0.930, 0.927, 0.817, and 0.927, respectively. ShuffleNetV2 required approximately 2.3 million parameters, compared with approximately 25.6 million for ResNet50, 8.0 million for DenseNet121, 5.3 million for EfficientNet-B0, and 86 million for ViT-Base/16 (Table S3).

Mean inference time was 0.9 s per case for ShuffleNetV2, compared with 2.4 s for ResNet50, 3.1 s for DenseNet121, 2.8 s for EfficientNet-B0, and 6.7 s for ViT-Base/16 under the same hardware and timing conditions.

Comparison between ShuffleNetV2 and radiologists’ diagnosis in the prospective cohort and subgroup analysis

Table 3 and Figure 4 present a comparative analysis of diagnostic performance between ShuffleNetV2 and radiologists in the prospective cohort, showing that AI assistance significantly improved the accuracy of active TB detection. Without AI support (Figure 4A), radiologists showed considerable variability in diagnostic performance, with AUC values ranging from 0.648 (95% CI: 0.607–0.689) for Junior_3 to 0.786 (95% CI: 0.746–0.824) for Senior_3, indicating significant experience-dependent disparities in tuberculosis identification capabilities. The ROC curves showed that all individual radiologists performed below the AI model’s benchmark AUC of 0.896 (95% CI: 0.860–0.932), with junior radiologists clustering in the lower performance range (AUC 0.648–0.700) and senior radiologists achieving moderate performance (AUC 0.786–0.811). Remarkably, AI-assisted interpretation (Figure 4B) dramatically enhanced diagnostic accuracy across all experience levels, with Senior_A, Intermediate_A, and Junior_A achieving AUC values of 0.937 (95% CI: 0.915–0.960), 0.902 (95% CI: 0.872–0.931), and 0.897 (95% CI: 0.869–0.923), respectively.

Table 3.

Performance of radiologists with and without ShuffleNetV2 assistance

Group	Reader	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)
Without AI assistance	senior 1	0.811 (0.774–0.848)	0.821 (0.788–0.851)	0.786 (0.716–0.843)	0.835 (0.796–0.869)	0.655 (0.585–0.718)	0.908 (0.874–0.933)
Without AI assistance	senior 2	0.798 (0.770–0.827)	0.730 (0.692–0.766)	0.956 (0.912–0.979)	0.641 (0.593–0.686)	0.514 (0.457–0.570)	0.974 (0.946–0.987)
Without AI assistance	senior 3	0.786 (0.746–0.824)	0.789 (0.754–0.821)	0.780 (0.709–0.837)	0.793 (0.751–0.830)	0.599 (0.531–0.663)	0.901 (0.865–0.928)
Without AI assistance	intermediate 1	0.722 (0.691–0.753)	0.629 (0.588–0.668)	0.937 (0.888–0.966)	0.506 (0.458–0.555)	0.429 (0.378–0.482)	0.953 (0.916–0.974)
Without AI assistance	intermediate 2	0.718 (0.677–0.759)	0.805 (0.771–0.836)	0.516 (0.439–0.592)	0.920 (0.890–0.943)	0.719 (0.631–0.794)	0.827 (0.790–0.860)
Without AI assistance	intermediate 3	0.718 (0.676–0.759)	0.729 (0.690–0.764)	0.692 (0.616–0.758)	0.743 (0.698–0.784)	0.516 (0.450–0.583)	0.859 (0.818–0.892)
Without AI assistance	junior 1	0.700 (0.665–0.734)	0.616 (0.575–0.655)	0.893 (0.836–0.932)	0.506 (0.458–0.555)	0.418 (0.366–0.471)	0.923 (0.880–0.951)
Without AI assistance	junior 2	0.679 (0.653–0.706)	0.552 (0.510–0.593)	0.975 (0.937–0.990)	0.384 (0.338–0.433)	0.386 (0.339–0.434)	0.975 (0.937–0.990)
Without AI assistance	junior 3	0.648 (0.607–0.689)	0.754 (0.716–0.788)	0.403 (0.330–0.480)	0.893 (0.859–0.919)	0.598 (0.503–0.686)	0.790 (0.750–0.825)
With AI assistance	senior A	0.937 (0.915–0.960)	0.938 (0.914–0.955)	0.937 (0.888–0.966)	0.938 (0.910–0.957)	0.856 (0.797–0.901)	0.974 (0.953–0.986)
With AI assistance	intermediate A	0.902 (0.872–0.931)	0.921 (0.896–0.941)	0.855 (0.792–0.902)	0.948 (0.921–0.966)	0.866 (0.804–0.911)	0.943 (0.916–0.962)
With AI assistance	junior A	0.897 (0.869–0.923)	0.902 (0.874–0.924)	0.887 (0.828–0.927)	0.908 (0.875–0.932)	0.792 (0.727–0.845)	0.953 (0.927–0.970)

AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value.

Figure 4.

Radiologist performance with and without ShuffleNetV2 assistance in the prospective cohort

Nine radiologists (three senior, three intermediate, and three junior) assessed the same 560 patient radiographs without model assistance (A); one radiologist from each experience level then reassessed the radiographs with assistance (B). AUCs and 95% confidence intervals are shown in the figure and Table 3. Improvements for the assisted senior, intermediate, and junior readers were evaluated using DeLong tests (all p < 0.001). Here n = 560 patients, with one chest radiograph per patient; AUC denotes area under the receiver operating characteristic curve.

Subgroup analyses showed that the ShuffleNetV2 exhibited robust performance across demographic and technical factors, including age, gender, occupation, residence, and scanner type (DeLong test, all p > 0.05) (Figure 5).

Figure 5.

Forest plot of subgroup performance in the retrospective external-validation cohort

Data are represented as area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (error bars). AUCs and 95% CIs are shown for age, gender, occupation, residence, and scanner-type subgroups. DeLong tests compared AUCs within each subgroup: age p = 0.191, gender p = 0.216, occupation p = 0.474, residence p = 0.792, and scanner type p = 0.064. The cohort included n = 1,318 patients, each contributing one chest radiograph; the numbers of active and non-active patients in each category are shown in the figure.

Interpretability analysis of ShuffleNetV2

ShuffleNetV2 demonstrates robust lesion localization capabilities within pulmonary regions, effectively discerning pathological changes while maintaining minimal responsiveness to extra-pulmonary structures such as bones and musculature despite the absence of manual annotations. Illustratively, Grad-CAM heatmaps in active TB cases (Figures 6A–6C) revealed peak model attention co-localized with the pathological findings, demonstrating strong spatial concordance between model-identified features and radiologically evident disease manifestations, whereas non-active TB cases (Figures 6D–6F) showed concentrated high-intensity signals in the right mid-to-upper lobe regions, indicating successful model recognition of characteristic post-tuberculous scarring patterns distinct from active disease manifestation. Misclassifications were observed in cases with potentially confounding thoracic findings. Figure 7A shows an active TB case that was incorrectly classified as non-active TB; overlap of the right clavicle with the apical pulmonary abnormality may have contributed to the error. Figure 7B shows a non-active TB case that was incorrectly classified as active TB; pleural calcification and chronic post-tuberculous changes may have been incorrectly interpreted by the model as features of active disease.

Figure 6.

Chest X-ray images with Grad-CAM visualizations in two representative cases

In A–C (active TB), a 54-year-old male exhibits a prominent consolidative opacity in the apical and subapical regions of the right upper lobe, appearing as a heterogeneous density with irregular margins. The Grad-CAM visualization quantitatively shows the attention mechanism of the model, showing maximal activation localized precisely within the radiographically confirmed consolidation. This shows strong spatial concordance between clinical observations and automated detection. In D–F (non-active TB), a 56-year-old male shows characteristic imaging findings of inactive pulmonary tuberculosis, including fibrotic striae in the right lobe. The Grad-CAM map reveals concentrated high-intensity activation in the mid to upper portion of the right lobe, corresponding to the region of fibrotic changes.

Figure 7.

Representative cases misclassified by ShuffleNetV2

(A) A reference-standard active TB case incorrectly predicted by the model as non-active TB; right clavicular overlap may have contributed to the misclassification.

(B) A reference-standard non-active TB case incorrectly predicted as active TB; pleural calcification and chronic post-tuberculous changes may have contributed to the misclassification.

Discussion

Accurate differentiation between active and non-active tuberculosis presents a crucial clinical challenge. Active TB, which is highly contagious through airborne transmission, necessitates prompt isolation and intensive multidrug therapy. Conversely, non-active TB carries no transmission risk and typically requires no therapeutic intervention. Rigorous monitoring of active TB cases is essential to evaluate treatment response and detect potential drug-resistant progression, while non-active cases demand only periodic surveillance to confirm sustained latency.^11,12,13

Our study developed and validated a DL model (ShuffleNetV2) for discriminating active from non-active TB using CXRs. While deeper architectures such as ResNet50 and ViT-Base/16 achieved marginally higher AUCs on retrospective external validation, ShuffleNetV2 demonstrated superior or equivalent performance on the prospective cohort—critical for real-world deployment—while requiring an order of magnitude fewer parameters, thereby validating its selection as a lightweight backbone for TB screening in resource-limited settings (Tables S2 and S3). The model achieved excellent performance in the retrospective external validation set (AUC: 0.917) and showed clinical applicability in prospective testing. Comparative analysis revealed ShuffleNetV2’s diagnostic superiority over radiologists across experience levels, significantly outperforming senior, intermediate, and junior practitioners. In addition, ShuffleNetV2 also showed profound augmentation potential, elevating diagnostic AUC to near-expert levels (0.897–0.937) regardless of reader seniority, thereby mitigating experience-related disparities. Regarding diagnostic efficiency, the model showed substantially faster interpretation times (0.9 s/case) versus human readers (seniors: 16.9 s; intermediates: 35.3 s; juniors: 87.1 s). The systematic ablation analyses further demonstrated the sensitivity of performance to key training and design choices (Table S1).

In the prospective cohort, in which the prevalence of active TB was 28.4%, the model achieved a positive predictive value of 0.668. Of the 205 patients predicted to have active TB, 68 were actually non-active cases, meaning that 33.2% of positive predictions were false positives. This finding has important clinical implications. If the model output were treated as a definitive diagnosis, false-positive results could lead to unnecessary patient isolation, additional diagnostic procedures, psychological distress, increased use of healthcare resources, and potentially inappropriate multidrug therapy. Therefore, ShuffleNetV2 should be used only as a triage or clinical decision-support tool. A positive prediction should prompt confirmatory evaluation using clinical findings and microbiological tests, including molecular testing and mycobacterial culture, rather than directly initiating anti-tuberculosis treatment. Furthermore, because positive predictive value depends strongly on disease prevalence, the model’s operating threshold should be evaluated and, where necessary, adjusted for the intended clinical population.

The prospective cohort demonstrated poorer calibration than the training and retrospective external validation cohorts. This finding illustrates that discrimination and calibration represent different aspects of model performance: a model may rank active and non-active cases effectively, as reflected by an AUC of 0.896, while still producing probability estimates that do not accurately represent absolute risk. One likely explanation is the marked prevalence shift between cohorts. Active TB accounted for 44.8% of the training cohort and 52.2% of the retrospective external validation cohort but only 28.4% of the prospective cohort. Consequently, a model trained on the higher-prevalence retrospective population may overestimate the probability of active TB in the prospective setting. Differences in recruitment period, institution, case mix, disease severity, and radiographic presentation may also have contributed. Before clinical implementation, the model should therefore undergo site-specific recalibration using representative local data, followed by prospective monitoring of calibration and operating thresholds.

Chest radiography remains a globally accessible imaging modality, particularly valuable in resource-limited settings due to its low cost, minimal patient discomfort, absence of contrast requirements, immediate image availability, and significantly lower radiation dose compared to computed tomography (CT). Given TB’s endemic status in developing regions, CXR-based DL models present a clinically pragmatic solution. In our previous work,¹⁴ we developed a CT-based model for active and non-active TB differentiation using 1160 active and 1131 inactive TB cases, employing a 3D Nested UNet for lesion delineation and comparing 3D VGG-16, EfficientNet, and ResNet-50 classifiers. External validation demonstrated the 3D ResNet-50’s superiority (AUCs: 0.961 and 0.946), outperforming radiologists with 10-fold faster interpretation. However, CT’s prohibitive cost and limited availability in endemic areas establish CXR-based models as essential initial triage tools for early detection.

Comparison with previous chest radiograph-based studies

Table 4 summarizes representative studies applying deep learning to chest radiograph-based tuberculosis assessment. Visuña et al.¹⁵ developed an explainable ensemble to identify tuberculosis and classify its stage as active or latent, reporting an accuracy of 96%. This represents a related activity-classification task; however, latent tuberculosis is not directly equivalent to the non-active pulmonary tuberculosis category used in our study, which was defined using microbiological, clinical, and radiographic criteria. Nafisah and Muhammad et al.¹⁶ reported an accuracy of 99.1% and an area under the receiver operating characteristic curve of 0.999 for tuberculosis-versus-non-tuberculosis classification using three public chest radiograph datasets. Acharya et al.¹⁷ reported an accuracy of 96.91% and an area under the receiver operating characteristic curve of 0.9938 for multiclass classification using a normalization-free network. Although these results are high, their endpoints primarily involved distinguishing tuberculosis from normal findings or other pulmonary diseases and therefore represent a different and generally less clinically specific task than differentiating active from non-active tuberculosis among patients with established tuberculosis.

Table 4.

Comparison with representative chest radiograph-based deep learning studies of tuberculosis

Study	Classification task	Data and validation	Model	Principal performance	Important distinction
Visuña et al., 2025	TB detection followed by active-versus-latent classification	CXR dataset with a small, imbalanced latent-TB class	CNN ensemble, random forest, and Grad-CAM	accuracy, 96%	closest endpoint, but latent TB is not identical to non-active pulmonary TB
Nafisah and Muhammad, 2022	TB versus non-TB	three publicly available CXR datasets	lung segmentation, CNNs, and explainable AI	accuracy, 99.1%; AUC, 0.999	no active-versus-non-active differentiation
Acharya et al., 2022	binary and multiclass TB detection	3,500 public CXRs	normalization-free network and Score-CAM	accuracy, 96.91%; AUC, 0.9938; sensitivity, 91.81%; specificity, 98.42%	no active-versus-non-active differentiation
Kazemzadeh et al., 2023	active TB versus no active TB	22,284 participants from 10 countries; multinational testing	deep learning system	test AUC, 0.89; sensitivity, 88%; specificity, 79%	active-TB screening rather than activity classification among patients with TB
Lee et al., 2021	active-TB screening	19,686 asymptomatic individuals; five participants had active TB	deep learning-based automated detection system	d all five active-TB CXRs; specificity, 0.959–0.997	very low-prevalence screening population
Present study	active versus non-active pulmonary TB	6,491 patients; multicenter retrospective and prospective validation	ShuffleNetV2	external AUC, 0.917; prospective AUC, 0.896	prospective validation and AI-assisted radiologist evaluation

Bold entries indicate the present study. Diagnostic performance should be compared cautiously because studies differed in target conditions, reference standards, disease prevalence, populations, datasets, and validation designs.

In clinically oriented studies, Kazemzadeh et al.¹⁸ reported an area under the receiver operating characteristic curve of 0.89 for active tuberculosis detection in a multinational test population, while Lee et al.¹⁹ demonstrated high sensitivity and specificity in systematic screening of asymptomatic individuals. In comparison, our model achieved areas under the receiver operating characteristic curve of 0.917 in the retrospective external validation cohort and 0.896 in the prospective validation cohort for active-versus-non-active tuberculosis classification. Our study additionally evaluated prospective performance, calibration, demographic and scanner-based subgroups, processing time, and the effect of AI assistance on radiologists with different experience levels. Nevertheless, numerical performance across studies should be interpreted cautiously because of differences in classification endpoints, disease prevalence, reference standards, dataset composition, and validation design.

Developing clinically relevant DL models deployable across diverse patient demographics and clinical settings could provide healthcare and clinicians with effective tools to mitigate disparities in global TB screening and triage. Radiologist scarcity remains critical in high-burden regions, particularly during community-led screening in remote settings, often preventing timely result communication prior to patient departure, which subsequently hinders adherence to follow-up care. While nucleic acid amplification tests (NAATs) offer high diagnostic specificity, their implementation costs restrict widespread use. Integrating NAATs with low-cost, moderately sensitive frontline screening via chest radiography represents a resource-conscious strategy to optimize population-level TB control through risk-stratified testing.

The consistent performance of the ShuffleNetV2 model across diverse demographic (age, gender, occupation, residence) and technical (CXR scanners) subgroups underscores its robustness as a diagnostic tool of active TB, suggesting minimal bias toward specific patient profiles or imaging platforms. This stability is clinically significant, indicating potential reliability in real-world settings where patient populations are heterogeneous and healthcare facilities utilize varied imaging equipment. Such generalizability reduces concerns about algorithmic fairness and technical dependency, key hurdles for clinical AI adoption, while supporting broader deployment across resource-variable environments.

This present study employed Grad-CAM for TB lesion visualization, though occasional misidentification of mediastinal structures as TB lesions was observed. These errors predominantly occurred in patients with mild TB symptoms, atypical radiographic presentations, or non-characteristic lesion distributions—findings more prevalent among immunocompetent individuals. Consequently, immune status evaluation merits consideration during TB classification to enhance diagnostic precision.

The model therefore shows potential as a rapid clinical decision-support tool for active-versus-non-active tuberculosis triage, while prospective impact on patient outcomes and health systems should be evaluated in future studies.

Limitations of the study

Our study has several limitations. First, the validation did not include international cohorts; thus, future work will assess diagnostic performance using larger, more diverse populations including international datasets. Second, our model did not incorporate significant clinical and laboratory parameters. We intend to integrate these metrics to enhance diagnostic accuracy. Third, classification errors may arise in patients with competent immune systems and atypical presentations. Therefore, clinical application necessitates careful consideration of individual patient characteristics during disease classification. Finally, we benchmarked ShuffleNetV2 against four representative alternative architectures under identical experimental conditions. Although ShuffleNetV2 demonstrated favorable prospective performance and computational efficiency, this comparison was not exhaustive and did not include all available architectures, model sizes, or optimization strategies. Future studies should evaluate additional architectures and independently reproduce these findings in larger and internationally diverse cohorts.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Mayidili Nijiati (mydl0911@163.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data: The clinical radiographs and associated records are not publicly available because they contain sensitive patient information. Researchers may contact the lead contact to request access to de-identified data, subject to review by the participating institutions and an appropriate data-use agreement.
• •Code: Public research scripts and a README are available at https://github.com/xujingxu/DL_APT.
• •Additional information: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (82360359); the Xinjiang Uighur Autonomous Health Commission, Special Scientific Research Project for Young Scientific and Technological Talents (WJWY-202304); the Autonomous Region Health Science and Technology Plan Project (2025001ZYWGHZSYJSTGXM653124849); the Key Research and Development Plan Project of Xinjiang Uygur Autonomous Region (2022B03032-1); the Special Fund Project for Central Government-Guided Local Science and Technology Development (2024B0202010005, XW); the Outstanding Youth Project of Xinjiang Uygur Autonomous Region Natural Science Foundation (2025D01E50); the second group of the Tianshan Talent Training Program: the Youth Support Talent Project (2023TSYCQNTJ0009); and the Xinjiang Autonomous Region “Tianshan Talents” Medical and Health Leading Talent Project (TSYC202401A011).

Author contributions

Conceptualization, M.N., J.M., A.T., and C.H.; project administration, M.N. and X.Z.; resources, A.T., S.A., J.M., C.H., and X.C.; data curation, A.T., S.A., J.M., and Y.L.; formal analysis and interpretation, C.H., J.X., and X.C.; writing – original draft, all authors; writing - review and editing, all authors. All authors approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used ChatGPT (OpenAI) for language-editing and document-formatting assistance. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
ShuffleNetV2	Ma et al., 2018; PyTorch torchvision	https://arxiv.org/abs/1807.11164; https://pytorch.org/vision/stable/models/shufflenetv2.html
PyTorch	Meta AI (formerly Facebook AI)	https://pytorch.org/; Version:1.8
Python	Python Software Foundation	https://www.python.org/; Version: 3.8
R	R Foundation for Statistical Computing	https://www.r-project.org/; Version: 3.5.2
Matplotlib	Matplotlib Development Team	https://matplotlib.org/; Version:1.7.2
Scikit-learn	scikit-learn developers	https://scikit-learn.org/; Version: 3.10.3
Public research scripts and README	This paper	https://github.com/xujingxu/DL_APT

Experimental model and study participant details

Study design and patient datasets

The overall study design is shown in Figure 1. In this multicenter cohort study, patients with microbiologically confirmed TB were included. All patients with available chest X-ray images. The exclusion criteria were as follows: incomplete clinical records, concurrent conditions that could impact chest X-ray findings (e.g., lung cancer, pacemaker or defibrillator placement, history of cardiac or pulmonary surgery), and extensive pleural effusion. Figure 2 displays the flowchart of patient inclusion.

Active and non-active TB cases were identified according to current expert consensus and diagnostic criteria.^20,21,22 Active TB was defined by: (i) positive microbiological confirmation of Mycobacterium tuberculosis, (ii) presence of symptoms including cough, weight loss, fever, night sweats, dyspnea, or chest pain, and (iii) imaging findings such as consolidations, nodules, cavitary lesions, ground-glass opacities, or tree-in-bud signs. Non-active TB cases were defined by negative pathogenetic confirmation of Mycobacterium tuberculosis and imaging features such as calcified lesions (single or multiple), well-defined cord-like lesions, sclerotic lesions, clean cavities, pleural thickening, adhesions, or calcifications.

Human participants

This multicenter study included patients with microbiologically confirmed pulmonary tuberculosis who had available chest radiographs. A total of 6,491 patients were included across the training, retrospective external validation, and prospective validation cohorts. Overall, the mean age was 60.73 ± 16.32 years, and there were 3,209 men and 3,282 women. The training cohort comprised 4,613 patients (2,067 active TB and 2,546 non-active TB; mean age, 60.36 ± 16.44 years; 2,260 men and 2,353 women). The retrospective external validation cohort comprised 1,318 patients (688 active TB and 630 non-active TB; mean age, 61.69 ± 15.36 years; 664 men and 654 women). The prospective validation cohort comprised 560 patients (159 active TB and 401 non-active TB; mean age, 61.51 ± 17.40 years; 285 men and 275 women). Sex was recorded as biological sex (male/female); gender, ancestry, race, and ethnicity were not separately collected or reported. All procedures involving human participants were approved by the Institutional Review Board of the First People’s Hospital of Kashi (Approval Number: 202243) and conducted in accordance with the Declaration of Helsinki and its later amendments.

Ethics statement

This study received ethical approval from the Institutional Review Board of the First People's Hospital of Kashi (Approval Number: 202243) in accordance with the Declaration of Helsinki and its later amendments. For the retrospective cohorts, the review board waived the requirement for informed consent. For the prospective cohort, all participants provided written informed consent prior to enrollment.

Method details

Data acquisition and preprocessing

We retrieved chest X-ray images in Digital Imaging and Communications in Medicine (DICOM) format from the medical imaging database of each participating hospital. The pixel values of DICOM images were normalized using window width and level adjustments to constrain values within the [0, 1] range. Images were then resized to the model's required input dimensions of 224×224 pixels to ensure compatibility.

Development and validation of a deep learning model

We developed ShuffleNetV2, a DL model using CXRs to distinguish active from non-active TB cases, based on the ShuffleNetV2 architecture. ShuffleNetV2 was selected because of its reported computational efficiency and favorable deployment characteristics, which may facilitate implementation in resource-limited clinical environments. The architecture enhances information flow through channel shuffle operations, which optimize feature propagation and significantly improve the model's sensitivity to discriminative features of active versus non-active TB, thereby boosting diagnostic accuracy. Furthermore, its group convolution strategy substantially reduces computational demands, enabling rapid inference on mobile platforms. This facilitates real-time active-versus-non-active pulmonary TB classification. Additionally, ShuffleNetV2 employs multi-scale feature fusion to integrate information across spatial dimensions, strengthening recognition and differentiation of TB status.

ShuffleNetV2 divided the network into four sequential stages, each comprising multiple ShuffleNet units. Throughout these stages, the spatial dimensions of feature maps progressively decreased while channel depth increased. Within each stage, ShuffleNet units were arranged sequentially, with outputs from one unit serving as inputs to the next. This hierarchical organization enabled progressive extraction and integration of image features. After the final stage, Global Average Pooling (GAP) and a Fully Connected Layer (FCL) were applied to generate differential diagnoses between active and non-active TB. Specifically, the GAP layer performed channel-wise averaging on the final feature maps, producing a fixed-length vector that encapsulated global image information. This feature vector was then passed through the fully connected layer, which mapped it to the model's output classes. The output layer employed a Softmax function to generate normalized probabilities for TB activity classification.

Retrospective data collected from Kashi Chest Hospital (KCH; n=2872) and Jiashi County Hospital (JCH; n=1741) were used for model training. Data pooled from Shufu County Hospital (SCH; n=1011) and Yecheng County Hospital (YCH; n=307) were used for retrospective validation, while data collected from Maigaiti County Hospital (MCH; n=560) were employed for prospective validation.

The ShuffleNetV2 architecture was implemented using PyTorch. Random horizontal and vertical flips were applied for data augmentation, which effectively quadrupled the training dataset from 4,613 to approximately 18,452 images (4,613 original images × 4 augmentation variants: original, horizontal flip, vertical flip, and both flips). The Adam optimizer was used to minimize cross-entropy loss with Softmax activation, together with a learning-rate scheduler (initial learning rate, 0.0001; weight decay, 0.001; momentum, 0.9). Training terminated upon model convergence or after reaching the maximum epoch count. Model development and inference were performed on a workstation equipped with an Intel Xeon E5-2697 v4 processor at 2.30 GHz, 505 GB of system memory, and two NVIDIA TITAN Xp graphics processing units, each with 12 GB of memory. The workstation ran Linux with Python 3.8, PyTorch 1.8.0, CUDA 11.1, and cuDNN 8.0.5. Training used a batch size of 64 and required approximately 57 hours. Inference speed was evaluated on a graphics processing unit using a batch size of 8. The mean processing time was 0.9 seconds per case, including image loading, DICOM preprocessing, and transfer to the graphics processing unit. The model’s generalization capability was validated in both retrospective and prospective cohorts.

Ablation and benchmarking experiments

All ablation and benchmarking experiments used the prespecified cohort splits: 4,613 training images, 1,318 retrospective external-validation images, and 560 prospective-validation images. Unless otherwise specified, experiments used the primary ShuffleNetV2 configuration: 224 × 224-pixel input images, a width multiplier of 1.0, ImageNet-pretrained initialization, random horizontal and vertical flips, the Adam optimizer, cross-entropy loss, an initial learning rate of 0.0001, weight decay of 0.001, a batch size of 64, and the same learning-rate scheduler and convergence criteria described above. Validation cohorts were not used for model fitting or parameter optimization.

For the ablation analysis, one design or training factor was changed at a time while all other settings were held constant. The evaluated variants were an input resolution of 384 × 384 pixels, width multipliers of 0.5 and 2.0, removal of data augmentation, removal of ImageNet pre-training, and replacement of cross-entropy loss with focal loss. Each variant was trained using the training cohort and evaluated independently in the training, retrospective external-validation, and prospective-validation cohorts. Accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated using the same procedures as for the primary model. The complete results are reported in Table S1.

For architecture benchmarking, ResNet50, DenseNet121, EfficientNet-B0, and ViT-Base/16 were trained using the same cohort splits, image preprocessing, input dimensions, augmentation pipeline, optimization settings, batch size, stopping criteria, and evaluation metrics as ShuffleNetV2. The models were evaluated without using the validation cohorts for training or model selection. The number of trainable parameters was recorded for each model. Inference time was measured on the same workstation using a batch size of 8 and the timing procedure described above, including image loading, DICOM preprocessing, and transfer to the graphics processing unit. Complete performance results are provided in Table S2, and the parameter-count and validation-AUC summary is provided in Table S3.

Performance of radiologists with and without ShuffleNetV2

We also evaluated the diagnostic performance of ShuffleNetV2, a DL model developed for active TB screening using chest X-rays, against nine radiologists stratified by experience: three senior radiologists (≥10 years), three intermediate radiologists (5-10 years), and three junior radiologists (<5 years). This comparative assessment examines standalone diagnostic accuracy across all cohorts using the identical, prospectively collected dataset (MCH cohort). Subsequently, a crossover study design was implemented where one randomly selected radiologist from each experience tier re-evaluates the same image set with ShuffleNetV2’s real-time assistance, enabling quantification of performance metrics before and after AI augmentation to determine diagnostic improvement.

Quantification and statistical analysis

Statistical analysis

Statistical analysis was performed using R software (version 3.5.2; R Foundation for Statistical Computing, Vienna, Austria). Quantitative variables were presented as mean ± standard deviation (SD), while categorical variables were expressed as frequencies or percentages. Continuous variables were analyzed using Student's t-test or Wilcoxon rank-sum test, and categorical variables with Chi-square or Fisher's exact test. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). All diagnostic performance estimates were reported with 95% confidence intervals (CIs). Confusion matrices were generated for detailed classification analysis. Model calibration was evaluated via calibration curves and Brier scores. Statistical comparisons between ShuffleNetV2 and radiologists employed the DeLong test for AUC differences and McNemar's test for binary performance metrics (such as accuracy, sensitivity, specificity). Gradient-weighted Class Activation Mapping (Grad-CAM) produced attention heatmaps at the radiograph level to enhance deep learning model interpretability. Statistical significance was defined as p <0.05.