Introduction

With an estimated 10.4 million new cases and 1.8 million fatalities1 from infectious diseases each year, tuberculosis (TB) is the main cause of infectious disease-related deaths globally, posing a constant challenge to public health2,3.

Mycobacterium tuberculosis (M.Tb) causes this disease, which can be transmitted via airborne means. Its impact significantly affects nearly 25% of the world's population, particularly in regions marked by socioeconomic deprivation2. The most profound impact of TB is observed in regions with lower to moderate incomes, where approximately two-thirds of all cases are concentrated within eight nations: India, Indonesia, China, Nigeria, Bangladesh, the Philippines, Pakistan, and South Africa2,3. Treatment for TB exists, with approximately 80% of infections being effectively managed through a six-month regimen of various antibiotics.

To effectively combat TB, early detection and identification of individuals with TB disease are essential. Unfortunately, a sizable portion, approximately 30% of all cases of TB disease remain unreported to the World Health Organization (WHO) every year4, underscoring the problem of underdiagnosis. The main TB screening method is based on chest X-ray (CXR) imaging because of its proven efficacy and affordability5. However, a notable challenge arises from the reliance on skilled human interpreters, such as radiologists, clinicians and/or technicians with training in radiology, to interpret CXR results, particularly due to their scarcity in the most affected regions6,7. Artificial intelligence-based solutions designed for resource-constrained settings have seen a noteworthy surge in attention due to the global scarcity of skilled CXR interpreters for TB screening8,9,10,11.

In March 2021, for the first time, the WHO recommended that computer-aided detection (CAD) software can replace human readers in interpreting digital CXRs for screening and triaging pulmonary TB disease12. The WHO recommends that CAD may be used to interpret antero-posterior or postero-anterior views of digital CXRs for pulmonary TB in individuals aged 15 years or older12. The ongoing use of digital radiography is more cost-effective than traditional methods, eliminating ongoing expenses associated with reagent use and radiologist services13.

The present TB workflow in resource-limited nations is marked by severe delays in finding and diagnosing TB patients14, resulting in a significant proportion of cases being diagnosed at the late stages of the disease15. To address this urgent need and in response to the WHO's endorsement of computer-assisted diagnosis of TB, our primary objective was to conduct TB screening on the most extensive cohort to date, comprising of 4495 participants and compare the accuracies of our novel AI-based CAD software named “DecXpert” with the gold standard molecular reference technique GeneXpert MTB/RIF16 in identification of active TB cases. The secondary objective of this study was to benchmark the diagnostic performance of DecXpert against that of 3 certified radiologists. DecXpert is a specialised deep convolutional neural network with self-attention mechanisms specifically designed for detection of active TB cases. This design holds relevance for real-world TB screening in areas lacking specialised personnel and facing resource limitations.

Results

Patient demographics

A total of 4495 participants were prospectively enrolled from January 2018 to November 2023 across 12 primary health care centers and one tertiary care center. From the initial cohort of 4495 participants, 132 individuals were excluded from the analysis for various reasons. Among these exclusions, 81 individuals reported an unproductive cough without an available bronchoalveolar lavage (BAL) sample, while 29 participants declined to participate in the study. Additionally, 4 individuals had a documented history of previous TB treatment, and 17 participants were pregnant. However, 1 individual's inclusion was deemed inconclusive due to uncertain results from either the CXR or GeneXpert MTB/RIF test (refer to Fig. 1). A total of 4363 individuals were ultimately included in the analysis. Among the 4363 individuals included in the study, 680 had an unproductive cough, but their BAL fluid was accessible and available for analysis. The median age of the participants was 43.1 years, and 2161 males (49.6%) and 2202 females (50.4%) were included. Predominantly, fever symptoms were evident in the majority of participants, accounting for 2565 individuals (58.8%) (refer to Table 1).

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Workflow for evaluating the performance of DecXpert CAD software. From the initially enrolled 4495 individuals, 4364 participants were included in the analysis after excluding those with no chest X-ray and/or inconclusive GeneXpert MTB/RIF results. The GeneXpert MTB/RIF test results categorized participants as TB(+) (2345) or TB(–) (2018). Radiologists interpreted the chest X-rays, identifying 1665 as TB(+) and 1695 as TB(–). The DecXpert CAD system classified 2064 cases as TB(+) and 1716 as TB(–). The performance metrics calculated include positive predictive value (PPV) and negative predictive value (NPV) for each method, using the GeneXpert MTB/RIF results as the reference standard.

Table 1 Demographic and clinical characteristics of the study population stratified by GeneXpert, radiology, and DecXpert test results.
Full size table

Hemoptysis and night sweats were reported by 437 (10%) and 306 (7%) participants, respectively. Among the individuals in the study cohort, 2345 individuals (53.7%) were confirmed to be TB positive, while 2018 individuals (46.3%) tested negative for TB (refer to Fig. 1). Notably, within the subgroup with positive TB results, there were 2161 (49.6%) males and 2202 (50.4%) females (Table 1). Gender, fever, cough, hemoptysis, and night sweats exhibited significant associations (P-value < 0.05) with the GeneXpert MTB/RIF test results (refer to Table 1). Moreover, gender, age, hemoptysis, night sweats and cough demonstrated significant associations (P-value < 0.01) with the radiological and DecXpert results (refer to Table 1).

Radiologists demonstrated a positive predictive value (PPV) of 71%, suggesting that out of the 2345 individuals identified as positive by GeneXpert MTB/RIF, 1665 individuals were confirmed positive by radiologists (refer to Table 2). In contrast, DecXpert exhibited a higher PPV of 88%, where out of the 2345 individuals identified as positive by GeneXpert MTB/RIF, 2064 were subsequently confirmed as positive by DecXpert. Regarding the negative predictive value (NPV), radiologists achieved an 83.9% NPV. This indicates that among the 2018 individuals classified as negative by GeneXpert MTB/RIF, 1695 were confirmed as negative by radiologists. Conversely, DecXpert exhibited a higher negative predictive value (NPV) of 85%, suggesting that of the 2018 individuals classified as negative for TB by GeneXpert MTB/RIF, 1716 were confirmed as negative by DecXpert (refer to Table 2).

Table 2 Performance evaluation of DecXpert against GeneXpert MTB/RIF and radiologists.
Full size table

Quantitative assessment

We evaluated the performance of different models for the identification of TB using the DecXpert score as a primary predictor (refer to Fig. 2). Initially, Model 1, which relies solely on DecXpert scores for TB detection, demonstrated an area under the receiver operating characteristic (ROC) curve (AUC) of 0.85 (95% CI 0.82–0.87), indicating a reasonably strong predictive ability (refer to Table 3 and Fig. 2). Model 2, which incorporated both the DecXpert score and symptom information, displayed an improved AUC of 0.88 (95% CI 0.83–0.92). When patient demographic information (specifically age and gender) was integrated with DecXpert scores in Model 3, the AUC further increased to 0.91 (95% CI 0.88–0.94) (refer to Table 3 and Fig. 2), indicating an enhanced predictive performance compared to that of earlier models. Furthermore, the development of a composite Model 4, which integrates the DecXpert score, symptom incidence, age, and gender, resulted in a greater AUC of 0.95 (95% CI 0.90–0.97) (refer to Table 3 and Fig. 2).

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Illustrates the diagnostic accuracy of DecXpert assessed through receiver operating characteristic (ROC) curves generated from the evaluated models in this study. The x-axis denotes different combinations of information used, while the y-axis indicates the corresponding AUC values. The bars represent scenarios including DecXpert scores alone, DecXpert scores combined with symptom information, DecXpert scores combined with age and gender information, and DecXpert scores combined with symptom information, age, and gender. Higher AUC values signify improved discrimination between individuals with and without active tuberculosis. The error bars depict the 95% confidence intervals for each AUC value.

Table 3 Summarizes the performance of the DecXpert Computer-Aided Detection system with added patient demographics and clinical information.
Full size table

Performance of the DecXpert algorithm against the gold standard molecular reference GeneXpert MTB/RIF

Next, we assessed the performance of the proposed DecXpert model that has proven to be highly effective in the identification of TB cases from CXR images. The GeneXpert MTB/RIF test reports served as the established ground truth for evaluation. The DecXpert model, operating solely on CXR imaging data without incorporating patient symptoms, achieved an AUC of 0.85 (95% CI = 0.82–0.87), indicated by the blue ROC curve in Fig. 3. However, upon inclusion of age and gender, the performance of the DecXpert model improved, yielding an AUC of 0.91 (95% CI = 0.88–0.94), indicated by the green ROC curve in Fig. 3. These results indicate a high level of accuracy ranging between 85 and 91% relative to GeneXpert MTB/RIF test reports, which are considered the gold standard (refer to Fig. 3), suggesting that DecXpert reports could potentially serve as a surrogate for GeneXpert MTB/RIF. Notably, when basic patient demographics, specifically age and gender, and patient symptoms such as cough, fever, hemoptysis and night sweats were omitted, the DecXpert model still demonstrated a strong 85% concordance with GeneXpert MTB/RIF testing.

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

Shows Receiver Operating Characteristic (ROC) curves for the DecXpert Computer-Aided Detection system. The blue curve represents DecXpert scores alone, yielding an Area Under the Curve (AUC) of 0.85 (95% CI 0.82–0.87), while the green curve represents DecXpert scores combined with age and gender information, resulting in an improved AUC of 0.91 (95% CI 0.88–0.94). The x-axis illustrates the false positive rate (1–specificity), while the y-axis depicts the true positive rate (sensitivity). Higher AUC values indicate better overall accuracy in discriminating between individuals with and without tuberculosis. Including age and gender information enhances DecXpert's diagnostic performance, as evidenced by the higher AUC value for the green curve.

Performance of the DecXpert algorithm against 3 board-certified radiologists

Analysing the overall cohort revealed that the DecXpert algorithm successfully identified 2064 TB patients (88%) out of the total 2345 GeneXpert MTB/RIF-confirmed positive TB patients, whereas the radiologists identified 1665 TB patients (71%). Thus, using the DecXpert algorithm increased the overall TB case detection rate by approximately 1.23 times compared to radiologists. Examining the performance of the three board-certified radiologists as illustrated in Fig. 4, the first radiologist achieved an AUC of 0.79 (95% CI 0.74–0.84), the second radiologist achieved an AUC of 0.72 (95% CI 0.67–0.76), and the third radiologist achieved an AUC of 0.75 (95% CI 0.71–0.78). Notably, each radiologist's sensitivity/specificity point fell outside the 95% CI space of the DecXpert model's ROC curve, indicating their performance was inferior to DecXpert (Fig. 4). Furthermore, within the unproductive cough and comorbidity subgroup, there was a considerable improvement in the TB case detection rate, according to DecXpert, which detected 303 patients (71.2%), whereas radiologists were able to detect only 244 patients (57.4%). This observation emphasises the enhanced performance of the DecXpert algorithm compared to that of radiologists, particularly within this subgroup, demonstrating its greater efficiency in identifying TB patients.

Fig. 4
Fig. 4The alternative text for this image may have been generated using AI.
Full size image

Diagnostic accuracy of DecXpert and three board certified radiologist experts compared with the reference GeneXpert MTB/RIF. The receiver operating characteristic (ROC) curve illustrates the performance of the DecXpert model (utilising age and gender) alongside the performance metrics of three board-certified radiologist experts, all plotted within the same ROC space. The area under the curve (AUC) quantifies the overall discriminative ability, while CI denotes the confidence interval surrounding the AUC estimation. The asterisk, bubble and square symbols represent the first, second and third radiologists respectively.

Qualitative assessment

DecXpert went through a validation process that focused on visualizing CXRs and highlighting specific areas in the image that are important for DecXpert to make decisions when classifying TB cases. The assessment shows DecXpert relies on clinically relevant lung regions from CXRs to make decisions. Figure 5 shows highlighted regions in CXRs of confirmed TB patients from the GeneXpert MTB/RIF test. The model relies on accurate visual information and does not consider misleading visual cues such as symbols, motion artifacts, embedded text or symbols, or imaging irregularities when making decisions. This finding demonstrated that DecXpert-related decision-making behaviour is primarily rooted in clinically relevant features.

Fig. 5
Fig. 5The alternative text for this image may have been generated using AI.
Full size image

Visualizing analyzed chest X-rays from sample TB patients. Sample chest X-ray images of TB patients are shown, with the highlighted areas demonstrating the most important aspects detected by the DecXpert system for identifying tuberculosis-related abnormalities.

Evaluation of DecXpert algorithm's suitability for deployment in remote isolated regions with offline and online functionality and minimal hardware needs

Subsequently, we evaluated the suitability of integrating DecXpert into the pre-existing CXR workflows within primary healthcare facilities and diagnostic centres, particularly for deployment in geographically isolated regions of the nation where computational hardware capabilities are limited. To this end, we examined both online and offline iterations of the DecXpert software and incorporated them into current TB CXR workflows for the purpose of TB screening and diagnosis at primary healthcare facilities and diagnostic centres. The investigation focused on the implementation of DecXpert across seven distinct providers of digital chest X-ray machines, including GE Healthcare™, Siemens™, Philips Healthcare™, FujiFilm Medical Systems™, Shimadzu Corporation™, Toshiba Medical Systems™, and Hitachi Healthcare™. This examination encompassed varying computational hardware setups, ranging from 500 MB to 16 GB of RAM, and spanning different versions of the Windows operating system (2000, 7, 8, and 10).

Additionally, the study investigated the compatibility of DecXpert with all five perspectives of CXR images—posteroanterior (PA), anteroposterior (AP), lateral, decubitus, and oblique views. This assessment was conducted at six remote and geographically dispersed locations within the northern region of India. Moreover, we aimed to ascertain the ease of use of DecXpert by the existing X-ray technicians at these sites within their current CXR workflow. DecXpert demonstrated seamless integration and compatibility with all CXR images from the seven vendors, supporting TB screening and diagnostic workflows. Notably, it functioned effectively with basic computational hardware, such as systems with 500 MB RAM and running Windows 2000. Furthermore, the on-site technicians at these healthcare facilities were easily trained on a simple four-step process for processing CXR images (refer to Fig. 6a–d). DecXpert was available in both offline and online configurations with the same functionality. Figure 6a–d shows the process wherein CXR images and patient demographics are uploaded to DecXpert, which then generates a diagnostic report in PDF format, available electronically or in print for physician assessment.

Fig. 6
Fig. 6The alternative text for this image may have been generated using AI.
Full size image

User interface and output displays of the DecXpert software. (a) The DecXpert software login screen. (b) The interface for uploading a chest X-ray image and entering patient details. (c) The output screen displaying the analyzed chest X-ray image along with a color-coded risk assessment for tuberculosis and other potential abnormalities. (d) A portable document format (PDF) report providing an overall risk score and quantitative assessment for TB and 18 other abnormalities, as well as a summary of clinical diagnosis.

Cost-effectiveness analysis of using DecXpert as a pre-screening tool for pulmonary TB in resource-limited settings

After obtaining encouraging results on the feasibility of integrating DecXpert into pre-existing CXR workflows, we subsequently conducted a cost-effectiveness analysis. This analysis aimed to evaluate the use of DecXpert as a pre-screening tool for patients presenting with symptoms suggestive of pulmonary TB at primary healthcare facilities in resource-limited settings, prior to molecular (GeneXpert MTB/RIF) testing. This strategy was compared to the current diagnostic algorithm for TB case finding and diagnosis, as outlined by the Central TB Division, Government of India17, which involves triaging by a general practitioner (GP) followed by microbiological testing with the GeneXpert MTB/RIF test.

The overall cost for a GeneXpert MTB/RIF test in the 145 countries eligible for subsidized GeneXpert MTB/RIF has been estimated at 14.90 USD18 which includes equipment, resources, maintenance and consumables. The cost analysis for the digital radiography system, including equipment and running costs such as labor, maintenance, consumables and depreciation has resulted in a cost of 1.49 USD19,20 per chest X-ray (CXR) screening.

Since CXR screening is significantly cheaper and faster than GeneXpert MTB/RIF testing and human reading, we performed a cost-effectiveness analysis of a scenario where DecXpert identifies a proportion of subjects eligible for subsequent GeneXpert MTB/RIF, and the remainder are discharged after the CXR screen (refer to Fig. 7).

Fig. 7
Fig. 7The alternative text for this image may have been generated using AI.
Full size image

Two-Stage Screening and Diagnostic Workflow for Tuberculosis Using DecXpert and GeneXpert MTB/RIF. The diagram illustrates the two-stage process for tuberculosis (TB) case finding and diagnosis. In the pre-screening stage, subjects are first subjected to chest X-ray (CXR) and evaluated using DecXpert. Results are obtained within one minute. Subjects with TB positive results on DecXpert are referred for further examination, while those found to be TB negative with DecXpert are not followed up. In the examination stage, subjects undergo GeneXpert MTB/RIF testing, with results available within two hours. Positive results lead to TB treatment, while negative results do not require further follow-up.

The DecXpert tool was compared to the current algorithm for TB case finding and diagnosis from the Central TB Division, Government of India, where patients presenting with symptoms suggestive of pulmonary TB at the primary healthcare facility are assessed by a general practitioner (GP). This comparison aimed to illustrate the effect on cost benefits.

The following metrics were calculated for both DecXpert and GP in a primary health setting:

  • Px: The proportion (0–1) of cases that will be sent for subsequent GeneXpert MTB/RIF testing. This includes all cases that would be marked as TB positive, including some false positives.

  • CostAVG: The average cost for a case arriving at the primary healthcare facility.

  • \({\text{Cost}}_{{{\text{AVG}}}} = {1}.{49} + \left( {{\text{Px}} \times {14}.{9}0} \right)\)

  • CostTB: The average cost per TB case detected. This calculation requires an estimate of TB prevalence, for which we used the incidence in the current study: CostTB = CostAVG/TB prevalence

Scenario 1: General Practitioner Screening.

A reasonable estimate for the accuracy of primary care physicians in India in identifying tuberculosis (TB) suspects is around 50%21. The study conducted by Satyanarayana et al.21 found that about 50% of presumptive TB cases in the community sought molecular testing, suggesting GPs correctly identify and refer around 50% of true TB suspects.

  • Px: 1–0.5 = 0.5 (Sensitivity of GP is 0.5 21)

  • CostAVG: 1.49 + (0.5 × 14.90) = 8.94 USD

  • CostTB: 8.94/(2345/4495) = 17.14 USD per TB case detected

Scenario 2: DecXpert Screening.

  • Px: 1–0.88 = 0.12 (Sensitivity of DecXpert is 0.88)

  • CostAVG: 1.49 + (0.12 × 14.90) = 3.278 USD

  • CostTB: 3.278/(2345/4495) = 6.29 USD per TB case detected

Cost reduction

Percentage decrease in cost per TB case detected = ((General practitioner cost–DecXpert cost)/General practitioner cost) × 100.

Percentage decrease in cost per TB case detected = ((17.14–6.29)/17.14) × 100 ≈ 63.32%

The results showed that when a general practitioner screens patients for further molecular testing, including GeneXpert MTB/RIF, the average cost per TB case detected was 17.14 USD. In contrast, when DecXpert screens patients for further molecular testing with GeneXpert MTB/RIF, the average cost per TB case detected was 6.29 USD. Therefore, the cost per TB case detected decreases by approximately 63.32% or 2.72 times when using DecXpert screening compared to a GP.

These findings suggest that the use of DecXpert as a pre-screening tool for pulmonary TB in resource-limited settings can significantly improve the cost-effectiveness of TB case detection compared to the current standard of care involving a general practitioner. This information can be valuable for decision-makers in resource-limited settings when considering the adoption of DecXpert for improving the efficiency and cost-effectiveness of TB case finding and diagnosis.

Discussion

Our investigation aimed to evaluate the efficacy of the AI-based DecXpert software in detecting TB in a resource-limited setting with a high disease burden but without HIV co-infection. Importantly, this represents the largest cohort study to date, comprising 4363 participants who underwent GeneXpert MTB/RIF testing for comparison with both DecXpert and radiologist interpretations. DecXpert effectively identified 88% (2064 out of 2345) of TB cases diagnosed by GeneXpert MTB/RIF, showcasing its potential to reduce the necessity for expensive molecular tests.

The current TB workflow in resource-limited nations suffers from severe diagnostic delays, with a significant proportion of cases detected at late disease stages. This delayed diagnosis poses a significant public health challenge, as it increases the risk of disease transmission and complications. The development of DecXpert, offering both online and offline deployment for automated CXR interpretation, represents a milestone in leveraging technological advancements for large-scale TB screening efforts. By enabling early and accurate identification of TB cases, DecXpert has the potential to revolutionize the existing diagnostic paradigm in resource-constrained settings. Furthermore, the use of DecXpert as a triage tool could enhance case identification and potentially reduce program expenses by optimizing the utilization of costly molecular testing resources such as Cartridge-Based Nucleic Acid Amplification Test (CBNAAT).

Our findings demonstrate that DecXpert achieves substantial diagnostic accuracy, with an area under the receiver operating characteristic curve (AUC) of 0.91 (95% confidence interval [CI] 0.88–0.94) when integrating patient age and gender (Model 3). The incorporation of demographic data facilitated the creation of personalized risk probabilities based on quantitative assessments of TB at the individual CXR level. This approach helps frontline healthcare workers prioritize testing effectively. In settings with limited GeneXpert resources, DecXpert optimizes testing by reducing cartridge use, benefiting high-burden facilities.

Furthermore, we explored the integration of DecXpert into existing CXR workflows within primary healthcare facilities and diagnostic centers, with a particular focus on deployment in geographically isolated regions with limited computational infrastructure. DecXpert demonstrated seamless integration and compatibility with CXR images from 7 different CXR vendors, supporting TB screening and diagnostic workflows across six remote locations in northern India, even on systems with basic computational hardware, such as those with 500 MB RAM and running Windows 2000. This highlights the versatility and adaptability of the software, making it suitable for deployment in resource-constrained settings with diverse hardware configurations.

Notably, on-site technicians at these healthcare facilities were easily trained on a simple four-step process for processing CXR images using DecXpert (Fig. 6). The software was made available in both offline and online configurations, offering the same functionality and thereby accommodating diverse operational environments. The streamlined process involved uploading CXR images and patient demographics, culminating in the generation of a probable diagnostic report in a portable document format (PDF) for physician assessment (Fig. 6). This user-friendly interface and workflow integration underscore the potential for widespread adoption of DecXpert as a valuable screening and diagnostic tool for TB, particularly in resource-limited and geographically isolated regions.

The cost-effectiveness analysis of DecXpert as a pre-screening tool for pulmonary TB in resource-limited settings revealed significant potential for reducing healthcare costs including TB screening programs. Our analysis compared the use of DecXpert with the current diagnostic algorithm involving general practitioners (GPs) and subsequent molecular testing (GeneXpert MTB/RIF). The findings demonstrated that using DecXpert for pre-screening reduced the average cost per TB case detected from 17.14 USD (with GP screening) to 6.29 USD. This represents a 63.32% reduction in cost per TB case detected, highlighting the economic advantages of integrating DecXpert into existing workflows.

These results underscore the potential of DecXpert to enhance the cost-effectiveness of TB case detection in resource-limited settings. By effectively triaging patients for subsequent molecular testing, DecXpert not only reduces the burden on expensive diagnostic resources but also facilitates earlier and more accurate identification of TB cases. This cost-saving potential is crucial for healthcare systems with limited budgets and high TB burden, suggesting that the adoption of DecXpert could lead to significant improvements in TB control efforts. This information can be valuable for decision-makers in resource-limited settings when considering the adoption of DecXpert for improving the efficiency and cost-effectiveness of TB case finding and diagnosis.

A critical aspect of our study was the validation process, which emphasized DecXpert's reliance on clinically relevant regions within the lung features extracted from CXRs for decision-making. By visualizing the specific areas within CXRs that were crucial for classifying TB cases, as illustrated in Fig. 5, the study highlights the algorithm's ability to identify and prioritize the relevant radiological features associated with TB disease. Notably, the study found that DecXpert's decision-making process relies on accurate visual information and does not consider misleading visual cues, such as symbols, motion artifacts, embedded text or symbols, or imaging irregularities. The emphasis on clinically relevant regions within the lung features extracted from chest X-rays aligns with the established understanding of TB manifestations and the radiological patterns associated with the disease. By focusing on these specific areas, DecXpert can leverage the diagnostic information contained within the images, potentially enhancing its ability to accurately identify TB cases. These findings contribute to the interpretability and trustworthiness of the system, potentially paving the way for its wider adoption and integration into clinical workflows for TB screening and diagnosis.

While the utilization of digital X-rays has expanded the potential for TB diagnosis and garnered community interest, operational considerations, health communication strategies, resource availability, and pathway development for individuals not diagnosed with TB remain crucial elements for the effective implementation of such programs. It is essential to address these factors to ensure that the benefits of DecXpert and similar technologies are fully realized and translated into improved patient outcomes.

Several limitations of this study should be acknowledged. Firstly, the use of GeneXpert MTB/RIF as the reference standard for TB diagnosis may have influenced the study's findings. While GeneXpert MTB/RIF is a highly sensitive and specific test, it is not a perfect diagnostic tool. False-negative results can occur, particularly in cases of extrapulmonary TB, drug-resistant TB, or in individuals with low bacterial loads. This could underestimate the true prevalence of TB in our study population and potentially overestimate the performance of DecXpert. Secondly, the study population may be subject to selection bias due to the inclusion criteria. Patients were recruited from primary healthcare facilities, which may not represent the entire spectrum of TB cases in the community. Individuals with severe symptoms or those residing in remote areas might be underrepresented in our sample. This could limit the generalizability of our findings to the broader population. Finally, the cost-effectiveness analysis relies on several key assumptions: the cost of GeneXpert MTB/RIF, the prevalence of TB, and the accuracy of General Practitioner diagnosis based on a single study21 and DecXpert. Changes in these factors could significantly impact the cost-effectiveness results. To enhance the robustness of our findings, additional economic evaluations using updated data are recommended.

Nonetheless, the introduction of DecXpert, a deep learning-powered tool for TB screening, demonstrated strong performance in detecting TB cases and explicable decision-making behaviour. This innovation supports frontline workers in high-risk regions combating TB, a major global health challenge. By using AI for clear and transparent predictions, DecXpert could revolutionize TB screening and diagnosis, especially in resource-limited settings.

In summary, our study underscores the potential of DecXpert as a valuable screening tool for TB, particularly in resource-limited settings. Its integration with demographic data shows promise for personalized risk assessment, offering insights that could enhance testing strategies and clinical workflows in the ongoing fight against TB. The robust diagnostic performance, adaptability to diverse healthcare settings, and explicable decision-making process contribute to the overall utility and trustworthiness of DecXpert. As we continue to grapple with the global burden of TB, innovative technologies like DecXpert represent a significant stride towards improving early detection, optimizing resource utilization, and ultimately saving lives.

Methods

Study design

This study included an extensive passive case-finding investigation of prospectively enrolled patients. This multicenter study was carried out across 12 primary health care (PHC) centres and one tertiary care (TC) centre in the northern Indian region. Patient enrollment occurred from January 2018 to January 2022 at the 12 PHC centres and from April 2022 to November 2023 at the one tertiary care centre. All patients who presented to the concerned centre with a history of fever, cough, expectoration, or constitutional symptoms for more than two weeks were included in the study. Patients with nonproductive cough and the listed symptoms underwent bronchoalveolar lavage (BAL) for diagnosis if available; otherwise, they were excluded. The study excluded pregnant females, patients who declined to participate, individuals already undergoing TB treatment, those presenting with a nonproductive cough, and participants for whom BAL fluid samples were unavailable (refer to Fig. 1).

Patients with TB were identified as individuals aged 15 years and older who had both a CXR and a GeneXpert MTB/RIF test confirming pulmonary TB. Patients without TB were identified as individuals aged 15 years and older who underwent a CXR, with a negative GeneXpert MTB/RIF16 test ruling out TB diagnosis.

The diagnostic capabilities were evaluated by utilising the sensitivity (Se), specificity (Sp), and the count of false positives, false negatives, true positives, and true negatives (FP, FN, TP, and TN, respectively) in relation to the results of the GeneXpert MTB/RIF test.

Scoring chest X-rays using DecXpert

By utilising outputs from its detection systems, the software treats them as descriptive features extracted from images. These features are then used to train a K-nearest neighbors classifier22, allowing for the computation of a cumulative abnormality/severity score for each CXR, which ranges from 0 to 100. A higher score indicates a more pronounced abnormality, potentially indicating TB. Individuals with DecXpert scores of less than or greater than 50 were sent for gold standard molecular GeneXpert MTB/RIF test (refer to Fig. 8).

Fig. 8
Fig. 8The alternative text for this image may have been generated using AI.
Full size image

Evaluation of DecXpert CAD performance against the gold standard molecular reference test GeneXpert MTB/RIF. The figure illustrates the workflow followed for comparing DecXpert and GeneXpert MTB/RIF test results, and determining false positives (FP), false negatives (FN), true positives (TP), and true negatives (TN) for DecXpert, using GeneXpert MTB/RIF as the reference standard.

Dataset and patients enrolled

We prospectively included all patients who met the inclusion criteria across 12 primary health care centres, along with individuals who sought care at the Pulmonary Medicine Department of a tertiary care centre in northern India. All participants underwent CXR and GeneXpert MTB/RIF testing as part of the screening process.

Among the initial 4495 participants screened, 4364 individuals were recruited into our study. However, the study included 4363 individuals, and one of the recruited individuals had an inconclusive GeneXpert MTB/RIF test result. Within this cohort, 2345 CXR images were obtained from patients diagnosed with TB, while 2018 CXR images were acquired from individuals without TB, as ascertained by their GeneXpert MTB/RIF test results, regarded as the reference gold standard.

Deep learning algorithm

We used a previously developed deep learning algorithm called DecXpert (version 1.4, which is available for download at https://github.com/Nikhil-techgit/decxpert/releases) was trained on 9876 chest X-ray images. The training data included 4932 X-ray images from TB patients and 4944 from non-TB patients. The patients for the training, validation, and test datasets were selected randomly from the chest X-ray images dataset, constituting 80%, 10%, and 10% of the dataset, respectively. All CXRs that were input into the algorithm were resized to the dimensions of 224 × 224 pixels. The following guidelines or limitations were established for developing DecXpert algorithm, outlined through an indicator function: (a) ensuring a sensitivity exceeding 80%, (b) achieving a specificity of more than 75%, and (c) constraining the number of parameters to 2 million.

To specify the general framework and specific attributes of a deep neural network customised for TB detection, we employed generative synthesis23. Using an optimal generator G, these designs were automatically found. Let A represent the generated architecture of the neural network based on a given set of initial configurations C, aimed at optimising a universal performance metric P24. This optimisation process is subject to specific constraints determined by an indicator function ϕ.

$$A=\text{max}P\left(A\left(c\right)\right) \, \text{such that} \, {1}_{\phi }\left(A\left(c\right)\right)=1;\forall c\in C$$
(1)

The aim was to find architectures that maximise performance while meeting predetermined criteria. First, the significant diversity of both macro and microarchitectures within the overall network architecture is a clear feature. This variation results from applying a meticulously tailored machine learning approach for TB case detection using CXR images, aiming to strike the best possible balance between efficiency and accuracy.

Second, the network architecture mainly comprises of convolutions applied on a depth and pointwise basis. This use of lightweight design patterns illustrates how the machine-driven design exploration strategy can modify the deep neural network microarchitecture in response to the imposed architectural complexity constraints. For DecXpert to potentially be widely adopted, its effective and efficient architecture is especially important. This is because DecXpert frequently needs to be deployed on low-end, low-cost computing devices in resource-constrained regions affected by poverty and economic constraints25.

Third, A notable aspect of the DecXpert network architecture includes visual attention condensers26, recently introduced, these mechanisms are a type of efficient attention condenser. These visual attention condensers enhance representational ability, requiring less computational and architectural complexity. Finally, this architecture predicts a patient's tuberculosis status as positive or negative.

Model training

Through the utilisation of stochastic gradient descent optimisation, the proposed neural network structure underwent training27. The training process involved a learning rate of 0.001, a momentum value of 0.6, and a batch size of 12, spanning across 400 epochs. Furthermore, data augmentation techniques were applied, including horizontal flipping, random cropping (15% margin), random contrast shifting (10% margin), and random intensity shifting (10% margin), were applied during the training phase28. After that, each image underwent resizing to a dimension of 224 × 224 pixels. Our experiments have shown that this specific size provides the highest level of performance, allowing for the conservation of important textural characteristics needed to differentiate individuals who are TB-positive from those who are TB-negative. No additional performance improvements were observed when using higher resolutions. The DecXpert deep neural network architecture underwent all construction, training, and evaluation procedures using the TensorFlow deep learning framework29.

Statistical analysis

For the statistical analysis, the results were scrutinised using DecXpert, GeneXpert MTB/RIF, and the evaluations of three certified radiologists. Regarding the demographic data, continuous variables were presented as the mean and standard deviation, while categorical variables were described in terms of numbers and percentages. A comparative analysis was conducted between the other methods and GeneXpert, which was considered as the gold standard. Additionally, a comparison was made between DecXpert and the radiologists' assessments. The sensitivity (Sn), specificity (Sp), positive predictive value (PPV), and negative predictive value (NPV) were calculated based on the positive and negative results obtained. The performance of these methods, including 95% confidence intervals (CI), was compared in differentiating between positive and negative patients for TB. The Sn, Sp, PPV, NPV, and their corresponding 95% CI were computed and reported. Receiver operating characteristic (ROC) curve analyses and area under curve (AUC) values, along with their respective 95% CI, were presented for the outcomes of positive patients versus negative. All statistical analyses were performed using R programming (version 4.1)30, and a statistical significance was determined with a two-tailed p-value below 0.05. Statistical significance was assessed with *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001, indicating progressively higher significance levels.

Positive Predictive Value (PPV) The positive predictive value is the proportion of true positive cases among all the cases that tested positive. In other words, it measures the probability that a positive result indicates the presence of the condition.

$$ {\text{PPV}} = \left( {\text{true positives}} \right)/\left( {{\text{true positives}} + {\text{false positives}}} \right) $$

Negative Predictive Value (NPV) The negative predictive value is the proportion of true negative cases among all the cases that tested negative. It measures the probability that a negative result indicates the absence of the condition.

$$ {\text{NPV}} = \left( {\text{true negatives}} \right)/\left( {{\text{true negatives}} + {\text{false negatives}}} \right) $$

Molecular validation using GeneXpert MTB/RIF

GeneXpert MTB/RIF16 was employed to verify the findings derived from DecXpert, encompassing both TB positive and negative cases, thereby enhancing confidence in the results.

Radiological validation

Three board-certified radiologists analysed and reported the patient’s chest X-rays. The first radiologist had more than 3 years of experience, the second had more than 5 years of experience, and the third had more than 6 years of experience. A majority vote (75%) among the three radiologists was required to classify a CXR as TB or normal. The radiological interpretations of the CXRs were subsequently compared to those of the DecXpert and GeneXpert MTB/RIF results.

Ethical compliance

The research study obtained ethical clearance from the Institutional Ethics Committee (IEC) at Sanjay Gandhi Post Graduate Institute of Medical Sciences, with the approval code 2022-59-IMP-EXP-46, ensuring adherence to ethical standards. The IEC waived the need for written informed consent for this study. All data related to the study participants was de-identified and anonymized. Additionally, every participant verbally consented to be part of the study after being informed about its details, ensuring compliance with ethical norms set by the ethics committee. The study methodologies adhered to applicable guidelines and regulations. The radiologists and researchers involved in the study only had access to non-identifiable patient data for analysis purposes. In contrast, all identifiable patient screening and diagnostic information remained securely stored on an internally protected server accessible only through credential-based authentication.