TBScreen.AI: Study Protocol for an Inclusive AI-Based Chest X-Ray Screening System for Tuberculosis in Remote Areas of Indonesia

TBScreen.AI: Study Protocol for an Inclusive AI-Based Chest X-Ray Screening System for Tuberculosis in Remote Areas of Indonesia | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Study protocol TBScreen.AI: Study Protocol for an Inclusive AI-Based Chest X-Ray Screening System for Tuberculosis in Remote Areas of Indonesia Antonia Morita Iswari Saktiawati, Wahyono Wahyono, Ari Probandari, and 12 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8937526/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Indonesia ranks second in tuberculosis (TB) burden globally and continues to face a significant diagnosis gap, particularly in remote areas with limited access to radiologists. Approximately 14% of new TB cases are left undetected by TB services. Utilizing artificial intelligence-based computer-aided detection (AI-CAD) applied to chest X-ray (CXR) imaging offers a potential solution to improve TB screening. However, AI systems may exacerbate existing inequities if they are developed and implemented without consideration of gender, age, disability, ethnicity, and socioeconomic barriers to care. This research aims to develop and train an AI-CAD model that considers the balance of gender, age, ethnicity, and disability status. The study will validate an AI-CAD model for TB screening among different socio-cultural and disability status groups. Additionally, it will assess barriers to TB services and inclusive strategies for AI-CAD implementation in remote areas of Indonesia. Method This mixed-methods study comprises three integrated components. Part I focuses on the development of TBScreen.AI, an AI-CAD model trained on gender-, age-, ethnicity-, and disability-balanced CXR data from hospitals in Java and Papua. Part II uses a prospective cross-sectional design to evaluate the diagnostic accuracy of TBScreen.AI among individuals with presumptive TB attending four health facilities in Indonesia. AI-generated interpretations are compared with radiologist readings and the final patient diagnosis established by site physicians, based on comprehensive examinations, including anamnesis, physical examination, laboratory investigations, and follow-up assessments. Part III integrates quantitative and qualitative methods to assess access to TB services and to identify inclusive implementation strategies that support equitable deployment of AI-assisted screening. Discussion This study investigates the development of AI-CAD tools for TB case detection using digital chest radiography, while also considering gender equality, disability, and social inclusion. The implementation of an accurate and inclusive AI-CAD tool as a second interpreter of CXR images for TB case detection will prevent treatment delays due to radiologist unavailability and reduce waiting time for CXR image interpretation in remote areas with limited health resources. artificial intelligence computer-aided detection tuberculosis screening chest X-ray tuberculosis gender disability remote areas inclusivity Figures Figure 1 Figure 2 Background Tuberculosis (TB) remains one of the world’s leading causes of death from a single infectious disease ( 1 ). Globally, Indonesia ranks second in TB burden, with an estimated 1.08 million cases and 126,100 deaths in 2024 ( 1 ). Despite substantial national efforts to control TB, a significant diagnosis gap has persisted in recent years ( 1 ). Approximately 14% of new TB cases are not detected by TB services, potentially leading to sustained community transmission and increased mortality among undetected and untreated individuals ( 2 ). To address this gap, the World Health Organization (WHO) recommends systematic screening for early detection to reduce transmission and improve treatment outcomes ( 3 , 4 ). TB screening with a chest X-ray (CXR) can correctly identify 85% of patients who actually have TB and rule out 96% of patients who do not, demonstrating greater accuracy compared to screening based on symptoms only ( 3 ). However, the use of CXR as a screening tool might be challenging due to the shortage and unequal distribution of radiologists and healthcare facilities, especially in remote areas of high-burden countries such as Indonesia. Nationwide, Indonesia has 2,303 professional radiologists or only 1.2 radiologists per 100,000 population ( 5 , 6 ). This estimation is one of the lowest rates in the world, and the actual number may be lower ( 7 – 9 ). According to the 2024 Indonesian Health Profile, a significant disparity exists between the number of radiologists and hospitals in Indonesia. Of 56,769 specialist doctors, only 2,303 (4.06%) are radiologists. In the same year, Indonesia had 3,155 hospitals (2,636 general and 519 specialty hospitals, predominantly located in the western region. Radiologists are heavily concentrated in Java and Bali (66.2%), with 18% in Sumatra, and only 15.8% distributed across Nusa Tenggara, Kalimantan, Sulawesi, Maluku, and Papua ( 6 ). This uneven distribution limits equitable access to radiology services, particularly in remote areas. The WHO advocates using computer-aided detection (CAD) software to support the interpretation of digital chest X-rays for TB screening, particularly in remote areas. CAD can provide initial screening in settings with limited specialist resources, with abnormal findings suggestive of TB referred to a radiologist for further evaluation ( 3 , 4 ).Despite global CAD development for TB screening, Indonesia lacks its own CAD software. Additionally, artificial intelligence (AI) may inherit and amplify biases, including those related to gender, disability status, age, and race ( 10 , 11 ). Meanwhile, Indonesia, with more than 600 ethnic groups, is one of the most diverse countries in the world ( 12 ), and people with disabilities, women, children, and the elderly are often underrepresented in access to health services. For example, although an estimated 16% of TB cases occur in children under 15 years of age, only 12% of TB patients presenting to health centers are in this age group ( 1 ). Approximately 3.9 million Indonesians (1.4% of the population) live with disabilities, and women with disabilities have been shown to access health services less frequently than men with disabilities, reflecting compounding inequities ( 13 ). These disparities underscore the importance of developing CAD systems that explicitly account for diversity in gender, disability status, age, and ethnic context. This research aims to: 1) develop and train an AI-CAD model using datasets balanced by gender, age, ethnicity, and disability status, 2) validate the diagnostic performance of the AI-CAD model for TB screening across diverse socio-cultural groups and by disability status, and 3) assess barriers to TB services and identify inclusive strategies for AI-CAD implementation in remote areas in Indonesia. Methods Research Setting The study will be conducted in four sites in three provinces in Indonesia: Balai Kesehatan Masyarakat Klaten (Balkesmas Klaten) in Central Java Province, Respira Lung Hospital and Sardjito General Hospital in Yogyakarta Province, and Mimika Regional Hospital in Central Papua Province. These four study sites represent different level of health facilities in Indonesia’s health system and were selected based on their diverse locations, high TB burden, and the availability of X-ray machines. Balkesmas Klaten is a secondary public health center located in the Klaten district that also covers six surrounding districts in Central Java. Balkesmas was initially dedicated to pulmonary care and now provides promotive, preventive, and curative services for the broader community. Respira Lung Hospital is a secondary-level hospital that primarily provides specialized care for respiratory conditions. Sardjito General Hospital is a tertiary-level hospital located in Yogyakarta, receiving patients from across Yogyakarta Province and the southern part of Central Java. Mimika Regional Hospital is a secondary-level public hospital in Mimika District, Central Papua. Central Papua ranked fifth as a province with the fewest number of radiologists ( 6 ). Study Design The study consists of the following three integrated components: Part 1: TBScreen.AI development phase In this phase, we develop TBScreen.AI to recognize chest X-ray patterns suggestive of TB and non-TB. The model generates a probability score indicating the likelihood of TB. As shown in Fig. 1 , the TBScreen.AI development is divided into four main stages: ( 1 ) Data preprocessing, ( 2 ) AI modeling and training, ( 3 ) Model optimization, and ( 4 ) User interface (UI) design and implementation. During the data preprocessing stage, CXR photos are collected from Sardjito General Hospital, Yogyakarta and Mimika Regional Hospital, Mimika. All CXR images are labeled according to the final diagnosis as pulmonary TB (PTB; clinically-diagnosed or bacteriologically-confirmed), extrapulmonary TB, or non-TB. We include 1,580 gender-balanced CXR images. The sample size is calculated based on articles published by Hitzl et al . and Balki et al. ( 14 , 15 ). The CXR data collection considers gender, age, and ethnicity balance to anticipate the biological impacts on lung physiology and immunity ( 16 – 18 ). CXR images of patients with scoliosis are included in the development phase to train the model to distinguish pulmonary features in individuals with and without scoliosis ( 19 , 20 ). Raw CXR images (.jpg, .jpeg, .png) are annotated by the radiologist team for TB-related abnormalities, including infiltrates, consolidation, cavities, effusion, fibrosis, and calcification. These annotations serve as the ground truth for classifying TB and non-TB cases, including normal and abnormal CXR. Finally, as the image quality may vary, Contrast-Limited Adaptive Histogram Equalization (CLAHE) is applied to enhance image quality and improve consistency. This step is essential to eliminate noise and ensure better feature extraction during AI modelling. In the second stage (AI modelling), core machine learning tasks are performed. The optimal TBScreen.AI model is developed to segment lung abnormalities and generate TB predictions. In this phase, the preprocessed data is used to train a deep learning model. The process starts with feature extraction from CXR images, including the position, area, perimeter, and other geometric information of TB-related abnormalities identified through segmentation using a U-Net deep learning model. A hybrid classification strategy is then applied, incorporating a weighted multiplier within a support vector machine framework. Based on these extracted features, the model generates decision rules and produces diagnosis recommendations that can serve as a preliminary tool for TB screening. Furthermore, hyperparameter tuning is conducted to optimize the model’s performance. By adjusting parameters such as learning rate, batch size, and number of layers, the model can achieve better accuracy and generalization. The performance of the trained lung segmentation model is analyzed using the mean Intersection over Union (IoU), by comparing predicted segmentation masks with ground truth annotations. The sensitivity and specificity of the TBScreen.AI model is assessed using Support Vector Machine and Random Forest classifiers, based on labeled TB features. The final output is categorized into two classes: TB or non-TB. The third stage (model optimization) aims to identify the best-performing model using the available dataset. Performance evaluation is done using various metrics to ensure the model’s reliability across different scenarios. Once the most accurate and stable model is selected, the model will be converted into an Application Programming Interface (API). This will enable seamless integration with other systems and facilitate real-time predictions or classifications in a production environment. The fourth stage (UI development) applies the best-performing model to a prototype web-based application. The interface includes a login page, an upload page for CXR image submission and analysis, a results page displaying the TBScreen.AI classification and detailed TB-related findings, and a list page documenting user identity and input history on the device. This initial design will be reviewed by potential users during the TBScreen.AI validation phase. Part 2: Validation phase of the TBScreen.AI This phase employs a cross-sectional design to validate the TBScreen.AI model among patients with signs and symptoms of TB. Prospective data collection for validation takes place in Balkesmas Klaten, Respira Lung Hospital, Sardjito General Hospital, and Mimika Regional Hospital. The data collection is planned over a 9-month period, from the end of August 2025 to the end of May 2026. The study population comprises all presumptive TB patients attending the health facilities. We calculated the sample size using the formula for estimating a population proportion, assuming a 99% confidence level (Z = 2.57), 5% margin of error, and design effect of 1. Based on reported TB bacteriological positivity rates in Indonesia (11.56%-12.82%) and Mimika (35.8%) ( 21 , 22 ), we applied hypothesized TB positivity proportions among presumptive TB patients (N) in Yogyakarta province ( 14% of 3,759,500), Klaten Municipality ( 14% of 1,302,648), and Mimika Municipality (35% of 313,016) ( 23 ). Allowing for 10% non-response, we will recruit 800 participants from Papua and 584 participants from study sites in Yogyakarta and Klaten. Mimika is intentionally oversampled as the primary remote implementation setting and to ensure an adequate number of TB-positive cases for validation. In total, the TBScreen.AI model will be evaluated on at least 1,384 CXR images from newly enrolled presumptive TB patients. In addition, the study will assess barriers to care and out-of-pocket expenditures among patients receiving TB treatment. The sample size for this descriptive component was calculated based on an assumed TB prevalence of 35%, with a 95% confidence level (Z = 1.96), 5% margin of error, and design effect of 1. Based on standard formulas for descriptive studies, we plan to recruit 350 patients receiving TB treatment. Overall, the total planned sample size is 1,734 participants. The inclusion criteria are ( 1 ) presumptive TB patients presenting with signs and symptoms suggestive of TB ( e.g. , persistent cough, unintentional ≥ 5% weight loss, night sweats) and ( 2 ) age ≥ 15 years. The exclusion criteria are ( 1 ) incomplete clinical data and ( 2 ) receipt of fixed-dose combination of anti-TB therapy for more than 7 days. We will collect participants' demographic and clinical data, including signs, symptoms, CXR images, and bacteriological examination results. Eligible individuals with presumptive TB will be consecutively enrolled. Indications for CXR imaging and bacteriological testing will follow routine procedures at each health facility. CXR examinations will be performed in accordance with WHO guidelines ( 4 ). Sputum samples will be collected for bacteriological testing (GeneXpert MTB/RIF, acid-fast bacilli/AFB smear microscopy, or culture, as available). This phase will apply a quadruple-blinding, in which the participants, radiographers, radiologists, and treating physicians will be blinded to the TBScreen.AI results. For the qualitative component, the research team will recruit 2–3 specialists (radiologists, internists, or pulmonologists) and 13–14 presumptive TB patients at each regional site for in-depth interviews. In addition, approximately 6–8 health workers (general practitioners, nurses, and paramedics) at each facility will participate in a focus group discussion. These data collection strategies are designed to achieve thematic saturation regarding barriers to TB services and opportunities for future implementation of TBScreen.AI technology. AI analysis will be performed after completion of data collection. CXR images obtained from each study site will be uploaded to the TBScreen.AI application and automatically analyzed to detect patterns of lung abnormalities. As a second reader, TBScreen.AI outputs will be considered predictive results rather than definitive diagnoses. The AI interpretations will be compared with radiologist readings and the patients’ final diagnoses established by site physicians, based on comprehensive examinations, including anamnesis, physical examination, laboratory investigations, and follow-up assessments, which will serve as the reference standards. Part 3: Assessment of access barriers and inclusive implementation strategies This phase complements the development and validation of TBScreen.AI by examining health system, sociocultural, and individual-level barriers to TB service access, and by identifying equity-oriented strategies for inclusive AI-assisted CXR screening in diverse Indonesian settings. Specifically, this study component will ( 1 ) identify barriers and facilitators to TB screening and diagnostic services across gender, age, disability, and geographic groups; ( 2 ) explore health providers’ perceptions, acceptability, trust, and concerns regarding AI-based TB screening tools, and ( 3 ) generate evidence-informed, context-sensitive recommendations for the inclusive integration of TBScreen.AI into routine TB services, particularly in remote and resource-constrained settings. This phase will employ a mixed-methods design, combining patient exit surveys with qualitative interviews and focus group discussions (FGDs). The integration of quantitative and qualitative data will allow triangulation of findings related to access barriers, user experiences, and implementation feasibility. Patient Exit Survey All participants will be invited to complete a structured exit survey following their TB screening and diagnostic procedures. Trained data collectors will administer the survey in Bahasa Indonesia or the relevant local language, depending on participant preference and literacy level. For participants with communication, cognitive, or psychosocial impairments, reasonable accommodations will be provided, including caregiver assistance. Where direct participation is not feasible, caregivers may respond as proxies, and this will be clearly documented. The exit survey will collect data across four domains: Sociodemographic and clinical characteristics, including age, sex, ethnicity, occupation, smoking status, comorbidities, and disability status (using self-reported functional limitations guided by the Washington Group Short Set on Functioning (WG-SS) ( 24 ); TB awareness and care-seeking behavior, including symptom recognition, perceived severity, and timing of healthcare utilization; Experiences of accessing TB services, including availability of services, perceived quality of care, stigma or discrimination, and cultural or gender-related constraints, and Economic barriers, including direct medical costs, non-medical costs (transport, food, accommodation), and income loss related to TB-related healthcare visits. The questionnaire is adapted from existing validated tools ( 25 , 26 ) and piloted prior to data collection. The case-report form and interview guide can be accessed in the S1 Appendix and S2 Appendix. Qualitative Interviews 1. Semi-structured interviews with patients Semi-structured interviews will be conducted with a purposive subsample of presumptive TB patients to capture diverse experiences across gender, age, disability status, and geographic location. Interviews will explore: (a) pathways to TB care and decision-making processes; (b) experiences of stigma, discrimination, or exclusion, and; (c) perceived accessibility and acceptability of TB. Participants with disabilities may be accompanied by caregivers. When communication barriers persist despite accommodations, caregiver proxy interviews will be conducted, with careful attention to ethical considerations. 2. Key informant interviews with health specialists Key informant interviews will be conducted with specialists involved in TB service delivery, including radiologists, pulmonologists, and internists. These interviews will examine: (a) current diagnostic workflows and bottlenecks, (b) perceived opportunities and risks of AI integration into TB service, (c) concerns related to diagnostic responsibility, accountability, and trust in AI systems, and ( 4 ) system-level requirements for equitable and sustainable AI deployment. Brief demographic information (age, professional role, years of experience) will be collected to contextualize findings. 3. Focus group discussions (FGD) with frontline health workers One FGD will be conducted at each study site with frontline healthcare workers (general practitioners, nurses, other paramedics involved in TB service) who have at least one year of experience delivering TB services. FGDs (60–90 minutes) will explore: (a) practical workflow implications of AI-assisted TB screening, (b) usability and acceptability of TBScreen.AI in routine practice, (c) perceived benefits and risks for different patient groups, and (d) recommendations for inclusive and context-appropriate implementation. Qualitative interview and FGD guides are provided in S3 Appendix. AI System The model and web-based application are co-developed by teams from the Department of Computer Science and Electronics, Faculty of Mathematics and Natural Sciences, Universitas Gadjah Mada, and the Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, with additional clinical and radiological input from RSUD Mimika. TBScreen.AI model To develop the AI model, the Python programming language was utilized along with the OpenCV library, which provides a wide range of functionalities for image processing and computer vision tasks. In the preprocessing stage, several image enhancement methods were explored and compared, including Histogram Equalization, Gamma Correction, Retinex, and CLAHE. Additionally, segmentation techniques were examined, where both Single U-Net and Double U-Net architectures were tested to determine their effectiveness. Following the preprocessing and segmentation stages, the modeling of artificial intelligence was carried out using two different machine learning algorithms, namely Support Vector Machine and Random Forest. These models were compared to evaluate their performance in handling the classification task. Based on the preliminary experiments, the best performance was achieved using a combination of CLAHE for image enhancement, Double U-Net for segmentation, and Support Vector Machine for classification. This combination showed the most promising results in the initial testing phase. Therefore, this model will be applied to the web application. TBScreen.AI application To make the system accessible from various platforms, including laptops, desktop computers, and mobile devices, it was decided to develop the application as a web-based system. A web-based application is considered more flexible and can be easily accessed through internet browsers without the need to install anything on the user’s device. The development of this web application was done using a combination of several programming languages, including Python, PHP, HTML, JavaScript, CSS, Flask, and SQLite. Each language has its own function in building the system, such as backend processing, user interface design, and database communication. The application is designed to support at least two main types of users. The first one is a guest user who can access the system without the need to log in. This type of user can perform self-screening independently and immediately receive the results. The second type is a healthcare worker, who is required to log in to the system. After logging in, healthcare professionals can perform screenings for multiple patients. The results of these screenings will then be stored in the system's database for further review and record-keeping. Figure 2 shows an example of the interface when the screening analysis results are displayed on the screen. The developed application can be accessed publicly through the following link: https://tbscreen.ai/ Analysis Plan Diagnostic accuracy In the TBScreen.AI development and training phase, receiver operating characteristic (ROC) analysis will be performed, and the area-under-curve (AUC) will be calculated to assess the diagnostic performance against a composite reference standard (CRS). The CRS consists of clinical assessment, radiologist readings, laboratory examinations, and follow-up findings (if available). During the AI-CAD validation phase, diagnostic performance will be evaluated using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with GeneXpert results and final clinical diagnoses serving as reference standards. Descriptive analyses will be used to summarize participants’ sociodemographic and clinical characteristics. Statistical analysis will be performed using STATA 18. Quantitative survey data The three primary outputs and their corresponding analytical approaches are outlined below: Delay in TB diagnosis and treatment, operationalized using patient-reported timelines ( e.g. , symptom onset to first care-seeking; first care-seeking to diagnostic evaluation; and, where applicable, diagnosis to treatment initiation). Delays will be analyzed as continuous measures (days) and categorized using pre-specified cut-offs appropriate to the Indonesian context. Barriers to accessing TB services, including health system barriers (e.g., availability of diagnostics/medicines, waiting time, referral complexity), geographic and transport constraints, sociocultural barriers (including stigma and gender-related constraints), and disability-related accessibility barriers. Barrier items will be summarized as proportions (overall and by category) and, where multi-response options are used, as multi-response frequencies. Patient-incurred costs and financial burden, aligned with the WHO TB Patient Cost Survey framework. Costs will be estimated at the visit level and aggregated to the TB episode level, disaggregated into direct medical costs, direct non-medical costs, and indirect costs (income loss and related productivity costs). Net out-of-pocket (OOP) expenditure will be calculated as total patient-incurred costs minus any reported reimbursements. Catastrophic TB-related costs will be defined as total TB-related patient-incurred costs exceeding 20% of annual household income, with sensitivity analyses using alternative thresholds ( e.g. , 10% and 30%) Stratified and equity-focused analysis All primary outputs will be analyzed overall and stratified by relevant equity dimensions and implementation-relevant characteristics, including (as sample size permits): gender, age group, disability status (including severity proxy), ethnicity, rural/urban residence, study site/province, and insurance coverage. For financial outcomes, analyses will also stratify by diagnostic vs treatment pathway and by whether a companion was required. Descriptive results will be presented using frequencies (%) for categorical variables and medians (IQR) for skewed continuous variables ( e.g ., costs, delays). Group differences will be assessed using chi-square tests for categorical outcomes and non-parametric tests (Wilcoxon rank-sum/Kruskal–Wallis) for continuous outcomes. Effect sizes (risk differences/odds ratios, median differences) will be reported with 95% confidence intervals. Multivariable modelling Multivariable models will be used to identify factors associated with: (i) longer delays; (ii) reporting key barriers; and (iii) higher OOP expenditure and catastrophic costs. Delay outcomes will be modelled using appropriate approaches based on distributions (e.g., log-transformed linear models or time-to-event style models if feasible). OOP expenditure will be modelled using generalized linear models (log link, gamma distribution), while catastrophic costs will be modelled using logistic regression. Models will include key covariates reflecting equity dimensions (gender, disability status, rurality, ethnicity), socioeconomic indicators, insurance status, and study site, with robust standard errors to account for clustering by facility/site, will be analyzed descriptively to characterize access barriers and out-of-pocket expenditures across participant subgroups. Where appropriate, subgroup analyses by gender, age group, disability status, and study site will be conducted. Qualitative data analysis Audio records are transcribed manually by the research team and translated into English. The quality of data will be checked before the coding and analysis. Transcripts of the interview will be thoroughly read to identify patterns and assess the fitness and relevance of the information. Thematic analysis will be conducted with Dedoose. The codebook will be developed and applied to a sample transcript, then cross-checked among researchers. Any discrepancies will be discussed and resolved, and the codes refined accordingly. This process is repeated until the codebook adequately captures the interview content. The final codebook is applied to all transcripts. Researchers will analyze emerging trends and patterns in the data and iteratively refine thematic categories. Data saturation will be considered achieved when no new themes emerge and similar findings are consistently observed across interviews. Ethical considerations The study will be conducted in accordance with the Declaration of Helsinki ( 27 ). Ethical approvals were obtained from the Health Research Ethics Committee, National Research and Innovation Agency, Indonesia (001/KE.03/SK/01/2025) and the Medical Health Research Ethics Committee, Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia (KE/FK/1439/EC/2025). All eligible patients recruited for the validation are given a detailed ‘participant information sheet’ and will sign an informed consent before participating in this study. During the validation phase, nurses at the study sites will provide brief information and offer patients the opportunity to meet with a member of the research team or staff to receive research details. The research team or staff will explain the research verbally, based on the information provided in the informed consent form, including the potential risks and the risk mitigation program. The research team will ask participants to sign the consent form, while the legal guardian or caregiver will sign on behalf of participants with intellectual or psychosocial disabilities. Participants can withdraw from the research during any research process or health procedures. The research teams can also decide to withdraw the participants from the research for urgent medical reasons. There will be no consequences after the withdrawal process, and patients will be diagnosed and treated according to routine healthcare practice. Discussion This is the first research protocol to develop and evaluate an AI-CAD tool for TB detection that incorporates gender equality, disability, and social inclusion considerations. The development of TBScreen.AI will examine group-based differences that may influence CXR interpretation within an AI framework. Furthermore, to promote an inclusive implementation of digital health technologies in TB services, the study will also investigate sociocultural determinants of access, including gender norms, stigma, accessibility barriers for persons with disabilities, and out-of-pocket expenditures. The findings are expected to inform equity-oriented TB screening strategies and promote more inclusive access to screening services. Wider implementation of TBScreen.AI could contribute to national TB control efforts by serving as a second reader for CXR interpretation. This may help reduce diagnostic delays due to radiologist shortages and prolonged waiting times, particularly in remote and resource-limited settings. Abbreviations ACF Active case finding AFB Acid-fast bacilli AI Artificial intelligence API Application Programming Interface AUC Area-under-curve CAD Computer-aided detection CLAHE Contrast Limited Adaptive Histogram Equalization CRS Composite Reference Standard CXR Chest radiography FGD Focus group discussion IoU Intersection over Union IQR Interquartile range NPV Negative predictive value OOP Out-of-pocket PPV Positive predictive value PTB Pulmonary tuberculosis RIF Rifampicin TB Tuberculosis WG-SS Washington Group Short Set on Functioning WHO World Health Organization Declarations Ethics approval and consent to participate Ethical clearance has been obtained from the Health Research Ethics Committee, the National Research and Innovation Agency, Indonesia (001/KE.03/SK/01/2025) and the Medical Health Research Ethics Committee, Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia (KE/FK/1439/EC/2025). Study permits were obtained from all study sites prior to data collection. Written consent and assent, either electronic or paper-based, are obtained from each participant before undergoing any research procedures. Participants can withdraw from the research during any research process or health procedures. Participant data protection and confidentiality are guaranteed. Consent for publication Informed consents are obtained from each participant for any data to be published in this manuscript. Availability of data and materials All data generated or analysed during this study are included in this published article and its supplementary information files. Competing interests The authors declare that they have no competing interests. Funding This study is supported by grants from the Department of Foreign Affairs and Trade (DFAT) Australia through KONEKSI (Collaboration for Knowledge, Innovation, and Technology Australia and Indonesia) under the supervision of Cowater International. The views expressed in this publication are the authors’ alone and are not necessarily the views of the Australian Government. Authors’ contributions AMS, W, TM, AP, HS, LA designed the study. AMS, TM, AP, RT, DU, NS, LA, NRA, JRP, AS, CAP contributed to the development of the interview guides, survey questionnaire, and study protocol. W, AMS, NHA, NRA, JRP, LC contributed to the development and validation of the AI-CAD model. AMS, TM, HS, AP developed analyses plan and performed sample size calculation. All authors are part of the core research team that meets regularly to discuss the design and implementation of the study. All authors have contributed intellectually by providing input to the manuscript, and all have read and approved the submitted version. Acknowledgments The authors gratefully acknowledge Chatarina Sari (Director of the YAKKUM Rehabilitation Center), Ayatullah (SAPDA), and their team in SAPDA and YAKKUM Rehabilitation Center for their support in strengthening the inclusivity approach of this study. We also extend our sincere appreciation to Cathy Vaughan and Min-hui Law from Nossal Institute for Global Health for their valuable methodological insights. We gratefully acknowledge Faisal Dharma Adhinata and Adi Suheryadi from the Department of Computer Science and Electronics at Universitas Gadjah Mada for their assistance with the AI system development. We acknowledge the critical inputs of the Tuberculosis Task Force and the General Directorate of Health Services for Vulnerable Populations, Ministry of Health, Republic of Indonesia, the Mimika District Health Office, also the Expert Committee on Tuberculosis (Komli TB) and Tuberculosis Research Network (JetSet TB) in Indonesia. Our gratitude also due to Silvester Sikora, the research assistant and field coordinator at the Mimika study site. We thank Erwan Hartadi and Rosalia Ratna from the Center for Tropical Medicine, Faculty of Medicine, Public Health and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia, for their assistance in administering the ethics clearance and research permits process. We are grateful to the healthcare workers and staff at RSUP Dr. Sardjito, RS Paru Respira, Balkesmas Klaten, and RSUD Kabupaten Mimika for their input in the development of this protocol. References WHO. Global tuberculosis report [Internet]. Geneva: World Health Organization. Licence: CC BY-NC-SA 3.0 IGO. 2025. Available from: https://iris.who.int/server/api/core/bitstreams/e97dd6f4-b567-4396-8680-717bac6869a9/content World Health Organization. The second national TB inventory study in Indonesia. Global Tuberculosis Report 2024. 2024. World Health Organization. WHO consolidated guidelines on tuberculosis. Module 2: screening – systematic screening for tuberculosis disease. Licence: C. Geneva: World Health Organization; 2021. World Health Organization. Chest Radiography in Tuberculosis Detection - Summary of current WHO recommendations and guidance on programmatic approaches. WHO Libr Cat Data. 2016. Yunus RE. Radiology Loading and Coverage Hours in Indonesia. Korean J Radiol. 2024. Kementerian Kesehatan Indonesia. Profil Kesehatan Indonesia (Health Profile of Indonesia) 2024 [Internet]. Minist. Heal. Indones. 2025. Available from: https://drive.google.com/file/d/1-lNRA3k9o9jM5vGacbnKY4OZorUQ-_Sc/view Pedrazzoli D, Lalli M, Boccia D, Houben R, Kranzer K. Can tuberculosis patients in resource-constrained settings afford chest radiography? Eur Respir J. 2017. van Cleeff MRA, Kivihya-Ndugga LE, Meme H, Odhiambo JA, Klatser PR. The role and performance of chest X-ray for the diagnosis of tuberculosis: A cost-effective analysis in Nairobi, Kenya. BMC Infect Dis. 2005. 10.1186/1471-2334-5-111 . Cited: in:: PMID: 16343340. Frija G, Blažić I, Frush DP, Hierath M, Kawooya M, Donoso-Bach L, Brkljačić B. How to improve access to medical imaging in low- and middle-income countries ? eClinicalMedicine. 2021. Zou J, Schiebinger L. AI can be sexist and racist — it’s time to make it fair. Nature. 2018. 10.1038/d41586-018-05707-8 . Qin ZZ, Van der Walt M, Moyo S, Ismail F, Maribe P, Denkinger CM, Zaidi S, Barrett R, Mvusi L, Mkhondo N, et al. Computer-aided detection of tuberculosis from chest radiographs in a tuberculosis prevalence survey in South Africa: external validation and modelled impacts of commercially available artificial intelligence software. Lancet Digit Heal. 2024. 10.1016/S2589-7500(24)00118-3 . Cited: in:: PMID: 39033067. BPS Statistics Indonesia. Profile of Ethnic Groups and Regional Language Diversity: 2020 Population Census Long Form Results. BPS-Statistics Indones; 2025. Putri NK, Syahansyah RJ. Disability-based disparities under universal health coverage among chronically ill adults during the COVID-19 pandemic in Indonesia: an interrupted time series analysis. Glob Health Action. 2025; doi: 10.1080/16549716.2025.2581946. Cited: in:: PMID: 41201369. Hitzl W, Reitsamer HA, Hornykewycz K, Mistlberger A, Grabner G. Application of discriminant, classification tree and neural network analysis to differentiate between potential glaucoma suspects with and without visual field defects. J Theor Med. 2003. 10.1080/10273360410001728011 . Balki I, Amirabadi A, Levman J, Martel AL, Emersic Z, Meden B, Garcia-Pedrero A, Ramirez SC, Kong D, Moody AR et al. Sample-Size Determination Methodologies for Machine Learning in Medical Imaging Research: A Systematic Review. Can. Assoc. Radiol. J. 2019. Lomauro A, Aliverti A. Sex and gender in respiratory physiology. Eur Respir Rev. 2021; doi: 10.1183/16000617.0038-2021. Cited: in:: PMID: 34750114. Schneider JL, Rowe JH, Garcia-de-Alba C, Kim CF, Sharpe AH, Haigis MC. The aging lung: Physiology, disease, and immunity. Cell. 2021. Ekström M, Mannino D. Research race-specific reference values and lung function impairment, breathlessness and prognosis: Analysis of NHANES 2007–2012. Respir Res. 2022; 10.1186/s12931-022-02194-4 . Cited: in:: PMID: 36182912. Johari J, Sharifudin MA, Rahman AA, Omar AS, Abdullah AT, Nor S, Lam WC, Yusof MI. Relationship between pulmonary function and degree of spinal deformity, location of apical vertebrae and age among adolescent idiopathic scoliosis patients. Singapore Med J. 2016; 10.11622/smedj.2016009 . Cited: in:: PMID: 26831315. Ragborg LC, Dragsted C, Ohrt-Nissen S, Mortensen J, Gehrchen M, Dahl B. Pulmonary function in patients with idiopathic scoliosis 40 years after diagnosis. Spine J. 2024. 10.1016/j.spinee.2024.07.006 . Cited: in:: PMID: 39097102. Ministry of Health Republic of Indonesia. Tuberculosis control program report 2023. 2024. Lestari T, Kamaludin, Lowbridge C, Kenangalem E, Poespoprodjo JR, Graham SM, Ralph AP. Impacts of tuberculosis services strengthening and the COVID-19 pandemic on case detection and treatment outcomes in Mimika District, Papua, Indonesia: 2014–2021. PLOS Glob Public Heal [Internet]. 2022;2:1–19. 10.1371/journal.pgph.0001114 . Ministry of Home Affairs of Indonesia. Population of the district/municipality [Internet]. 2020. Available from: https://pelita.kemendagri.go.id/kemendagri/dataset/257/tabel-data Washington Group on Disability Statistics. The Washington Group Short Set on Functioning (WG-SS) [Internet]. 2020. Available from: https://www.washingtongroup-disability.com/fileadmin/uploads/wg/Documents/Questions/Washington_Group_Questionnaire__1_-_WG_Short_Set_on_Functioning.pdf WHO. Tuberculosis patient cost surveys: a handbook. World Health Organization; 2017. Organization WH. Consolidated guidance on tuberculosis data generation and use. Module 4. Surveys of costs faced by households affected by tuberculosis. Geneva: World Health Organization; 2025. Practice IEG. clinical. Scientific guideline European Medicines Agency (EMA) [Internet]. 2002. Available from: https://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline Additional Declarations No competing interests reported. Supplementary Files S1AppendixExitsurvey.pdf S2AppendixPatientinterviewguide.pdf S3AppendixInterviewandFGDguide.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8937526","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Study protocol","associatedPublications":[],"authors":[{"id":601025248,"identity":"5fef23d4-7500-463d-a322-696a44e08b52","order_by":0,"name":"Antonia Morita Iswari Saktiawati","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzElEQVRIiWNgGAWjYNACAwYGfhibjSGBKC0GDJINpGkBWmNwAM4hoIW/gTvxcUXBH3njG+lXNzDU2DHwsRPQInGAd7PhGQMDw203cspuMBxLZmDjeUDASQd4t0k2GBgwArWk3WBgO8DAJkHAFnmoFvvNM0Ba/hGhxQCqJXGDRPqxG4xtRGgxPAz0S4OBcfKMM2/YbiT2JfMQ9Ivc8d6NDxv+yNn2t6c/u/Hhm52cfDsBWxiY4SweA1CM8BBQjwLYCThoFIyCUTAKRiwAABCkQFFn8YbDAAAAAElFTkSuQmCC","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":true,"prefix":"","firstName":"Antonia","middleName":"Morita Iswari","lastName":"Saktiawati","suffix":""},{"id":601025249,"identity":"6fa9f07a-803c-42c4-8fcc-c24eeacc555b","order_by":1,"name":"Wahyono Wahyono","email":"","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":false,"prefix":"","firstName":"Wahyono","middleName":"","lastName":"Wahyono","suffix":""},{"id":601025250,"identity":"8353737e-f1ae-46d0-94ce-eac9ed00db03","order_by":2,"name":"Ari Probandari","email":"","orcid":"","institution":"Universitas Sebelas Maret","correspondingAuthor":false,"prefix":"","firstName":"Ari","middleName":"","lastName":"Probandari","suffix":""},{"id":601025251,"identity":"cf037278-a7df-4b03-a741-deb754b16649","order_by":3,"name":"Naela Himayati Afifah","email":"","orcid":"","institution":"Gadjah Mada University/Dr. Sardjito Hospital","correspondingAuthor":false,"prefix":"","firstName":"Naela","middleName":"Himayati","lastName":"Afifah","suffix":""},{"id":601025252,"identity":"06176983-3458-4dc9-b95a-08f9ef6b585f","order_by":4,"name":"Nur Rahmi Ananda","email":"","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":false,"prefix":"","firstName":"Nur","middleName":"Rahmi","lastName":"Ananda","suffix":""},{"id":601025253,"identity":"29665055-ef70-4fec-b97f-3ee735578890","order_by":5,"name":"Luthfi Azizatunnisa","email":"","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":false,"prefix":"","firstName":"Luthfi","middleName":"","lastName":"Azizatunnisa","suffix":""},{"id":601025254,"identity":"cb6c81b0-cf7e-4131-baf6-6317e4b19d36","order_by":6,"name":"Henry Surendra","email":"","orcid":"","institution":"Monash University Indonesia","correspondingAuthor":false,"prefix":"","firstName":"Henry","middleName":"","lastName":"Surendra","suffix":""},{"id":601025255,"identity":"b2f16e0f-b50e-44b9-a184-f1ebefc400c5","order_by":7,"name":"Jeanne Rini Poespoprodjo","email":"","orcid":"","institution":"Papuan Health and Community Development Foundation","correspondingAuthor":false,"prefix":"","firstName":"Jeanne","middleName":"Rini","lastName":"Poespoprodjo","suffix":""},{"id":601025256,"identity":"19cf508a-654a-4390-8931-e082ccdb44c6","order_by":8,"name":"Rita Triharyani","email":"","orcid":"","institution":"YAKKUM Rehabilitation Center","correspondingAuthor":false,"prefix":"","firstName":"Rita","middleName":"","lastName":"Triharyani","suffix":""},{"id":601025257,"identity":"ce46ca3c-70aa-4362-a800-6383d68c6146","order_by":9,"name":"Dewi Utari","email":"","orcid":"","institution":"YAKKUM Rehabilitation Center","correspondingAuthor":false,"prefix":"","firstName":"Dewi","middleName":"","lastName":"Utari","suffix":""},{"id":601025258,"identity":"a7e67353-4335-44fc-b814-9c64ea9510da","order_by":10,"name":"Nurul Sa’adah","email":"","orcid":"","institution":"Advocacy Center for Women, Persons with Disabilities, and Children (SAPDA)","correspondingAuthor":false,"prefix":"","firstName":"Nurul","middleName":"","lastName":"Sa’adah","suffix":""},{"id":601025259,"identity":"2597895d-f86b-45b0-8f39-c1701d2a6cf7","order_by":11,"name":"Chairunnisa Athena Pelawi","email":"","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":false,"prefix":"","firstName":"Chairunnisa","middleName":"Athena","lastName":"Pelawi","suffix":""},{"id":601025260,"identity":"d9c9c099-e184-4419-a459-ae0c92a7a384","order_by":12,"name":"Aulia Shafira","email":"","orcid":"","institution":"Universitas Gadjah Mada","correspondingAuthor":false,"prefix":"","firstName":"Aulia","middleName":"","lastName":"Shafira","suffix":""},{"id":601025262,"identity":"5da69d0d-f808-4ad3-a5aa-00c992aca990","order_by":13,"name":"Lina Choridah","email":"","orcid":"","institution":"Gadjah Mada University/Dr. Sardjito Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lina","middleName":"","lastName":"Choridah","suffix":""},{"id":601025263,"identity":"6d1461df-c4dc-4601-aca5-d2d1bfd63027","order_by":14,"name":"Tiara Marthias","email":"","orcid":"","institution":"The University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Tiara","middleName":"","lastName":"Marthias","suffix":""}],"badges":[],"createdAt":"2026-02-22 07:23:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8937526/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8937526/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105764159,"identity":"5e949580-e3db-41eb-b29b-1c02fb9c8ef7","added_by":"auto","created_at":"2026-03-30 19:24:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":148189,"visible":true,"origin":"","legend":"\u003cp\u003ePipeline of TBScreen.AI Development\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/c7c0c8e5c54c6eaa690e4952.png"},{"id":105764132,"identity":"d58dbbca-3b60-4c29-884a-ae59bdbad0e0","added_by":"auto","created_at":"2026-03-30 19:24:41","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":292317,"visible":true,"origin":"","legend":"\u003cp\u003eIllustration of TB screening results using our application in https://tbscreen.ai\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/27b328fa47fb4f2ab41e6c8c.png"},{"id":105764166,"identity":"bdbe4d1e-ed38-469f-b927-925011a2d30e","added_by":"auto","created_at":"2026-03-30 19:24:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1131701,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/25c11283-d7b8-4f56-97c4-2e7935ec6c8c.pdf"},{"id":105764158,"identity":"acc73783-f1d5-46fa-a273-dd62b84528d0","added_by":"auto","created_at":"2026-03-30 19:24:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":604775,"visible":true,"origin":"","legend":"","description":"","filename":"S1AppendixExitsurvey.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/04b89a6d02b4ea3e91d68ca8.pdf"},{"id":105764135,"identity":"d93127b8-8e9b-4411-82ad-540fd1fd24fe","added_by":"auto","created_at":"2026-03-30 19:24:42","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":261920,"visible":true,"origin":"","legend":"","description":"","filename":"S2AppendixPatientinterviewguide.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/55903dc7f83cc62e60ef29c1.pdf"},{"id":105764160,"identity":"d1e631e2-46c8-409d-9947-987c82c75307","added_by":"auto","created_at":"2026-03-30 19:24:49","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":255353,"visible":true,"origin":"","legend":"","description":"","filename":"S3AppendixInterviewandFGDguide.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8937526/v1/8816c5ef785f2b2142b4a7c6.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"TBScreen.AI: Study Protocol for an Inclusive AI-Based Chest X-Ray Screening System for Tuberculosis in Remote Areas of Indonesia","fulltext":[{"header":"Background","content":"\u003cp\u003eTuberculosis (TB) remains one of the world\u0026rsquo;s leading causes of death from a single infectious disease (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Globally, Indonesia ranks second in TB burden, with an estimated 1.08\u0026nbsp;million cases and 126,100 deaths in 2024 (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Despite substantial national efforts to control TB, a significant diagnosis gap has persisted in recent years (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Approximately 14% of new TB cases are not detected by TB services, potentially leading to sustained community transmission and increased mortality among undetected and untreated individuals (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). To address this gap, the World Health Organization (WHO) recommends systematic screening for early detection to reduce transmission and improve treatment outcomes (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTB screening with a chest X-ray (CXR) can correctly identify 85% of patients who actually have TB and rule out 96% of patients who do not, demonstrating greater accuracy compared to screening based on symptoms only (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). However, the use of CXR as a screening tool might be challenging due to the shortage and unequal distribution of radiologists and healthcare facilities, especially in remote areas of high-burden countries such as Indonesia.\u003c/p\u003e \u003cp\u003eNationwide, Indonesia has 2,303 professional radiologists or only 1.2 radiologists per 100,000 population (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). This estimation is one of the lowest rates in the world, and the actual number may be lower (\u003cspan additionalcitationids=\"CR8\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). According to the 2024 Indonesian Health Profile, a significant disparity exists between the number of radiologists and hospitals in Indonesia. Of 56,769 specialist doctors, only 2,303 (4.06%) are radiologists. In the same year, Indonesia had 3,155 hospitals (2,636 general and 519 specialty hospitals, predominantly located in the western region. Radiologists are heavily concentrated in Java and Bali (66.2%), with 18% in Sumatra, and only 15.8% distributed across Nusa Tenggara, Kalimantan, Sulawesi, Maluku, and Papua (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). This uneven distribution limits equitable access to radiology services, particularly in remote areas.\u003c/p\u003e \u003cp\u003eThe WHO advocates using computer-aided detection (CAD) software to support the interpretation of digital chest X-rays for TB screening, particularly in remote areas. CAD can provide initial screening in settings with limited specialist resources, with abnormal findings suggestive of TB referred to a radiologist for further evaluation (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).Despite global CAD development for TB screening, Indonesia lacks its own CAD software.\u003c/p\u003e \u003cp\u003eAdditionally, artificial intelligence (AI) may inherit and amplify biases, including those related to gender, disability status, age, and race (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Meanwhile, Indonesia, with more than 600 ethnic groups, is one of the most diverse countries in the world (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e), and people with disabilities, women, children, and the elderly are often underrepresented in access to health services. For example, although an estimated 16% of TB cases occur in children under 15 years of age, only 12% of TB patients presenting to health centers are in this age group (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Approximately 3.9\u0026nbsp;million Indonesians (1.4% of the population) live with disabilities, and women with disabilities have been shown to access health services less frequently than men with disabilities, reflecting compounding inequities (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). These disparities underscore the importance of developing CAD systems that explicitly account for diversity in gender, disability status, age, and ethnic context.\u003c/p\u003e \u003cp\u003eThis research aims to: 1) develop and train an AI-CAD model using datasets balanced by gender, age, ethnicity, and disability status, 2) validate the diagnostic performance of the AI-CAD model for TB screening across diverse socio-cultural groups and by disability status, and 3) assess barriers to TB services and identify inclusive strategies for AI-CAD implementation in remote areas in Indonesia.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eResearch Setting\u003c/h2\u003e \u003cp\u003eThe study will be conducted in four sites in three provinces in Indonesia: \u003cem\u003eBalai Kesehatan Masyarakat Klaten\u003c/em\u003e (Balkesmas Klaten) in Central Java Province, Respira Lung Hospital and Sardjito General Hospital in Yogyakarta Province, and Mimika Regional Hospital in Central Papua Province. These four study sites represent different level of health facilities in Indonesia\u0026rsquo;s health system and were selected based on their diverse locations, high TB burden, and the availability of X-ray machines. Balkesmas Klaten is a secondary public health center located in the Klaten district that also covers six surrounding districts in Central Java. Balkesmas was initially dedicated to pulmonary care and now provides promotive, preventive, and curative services for the broader community. Respira Lung Hospital is a secondary-level hospital that primarily provides specialized care for respiratory conditions. Sardjito General Hospital is a tertiary-level hospital located in Yogyakarta, receiving patients from across Yogyakarta Province and the southern part of Central Java. Mimika Regional Hospital is a secondary-level public hospital in Mimika District, Central Papua. Central Papua ranked fifth as a province with the fewest number of radiologists (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy Design\u003c/h3\u003e\n\u003cp\u003eThe study consists of the following three integrated components:\u003c/p\u003e\n\u003ch3\u003ePart 1: TBScreen.AI development phase\u003c/h3\u003e\n\u003cp\u003eIn this phase, we develop TBScreen.AI to recognize chest X-ray patterns suggestive of TB and non-TB. The model generates a probability score indicating the likelihood of TB. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the TBScreen.AI development is divided into four main stages: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Data preprocessing, (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) AI modeling and training, (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Model optimization, and (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) User interface (UI) design and implementation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDuring the data preprocessing stage, CXR photos are collected from Sardjito General Hospital, Yogyakarta and Mimika Regional Hospital, Mimika. All CXR images are labeled according to the final diagnosis as pulmonary TB (PTB; clinically-diagnosed or bacteriologically-confirmed), extrapulmonary TB, or non-TB. We include 1,580 gender-balanced CXR images. The sample size is calculated based on articles published by Hitzl \u003cem\u003eet al\u003c/em\u003e. and Balki \u003cem\u003eet al.\u003c/em\u003e (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). The CXR data collection considers gender, age, and ethnicity balance to anticipate the biological impacts on lung physiology and immunity (\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). CXR images of patients with scoliosis are included in the development phase to train the model to distinguish pulmonary features in individuals with and without scoliosis (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRaw CXR images (.jpg, .jpeg, .png) are annotated by the radiologist team for TB-related abnormalities, including infiltrates, consolidation, cavities, effusion, fibrosis, and calcification. These annotations serve as the \u003cem\u003eground truth\u003c/em\u003e for classifying TB and non-TB cases, including normal and abnormal CXR. Finally, as the image quality may vary, Contrast-Limited Adaptive Histogram Equalization (CLAHE) is applied to enhance image quality and improve consistency. This step is essential to eliminate noise and ensure better feature extraction during AI modelling.\u003c/p\u003e \u003cp\u003eIn the second stage (AI modelling), core machine learning tasks are performed. The optimal TBScreen.AI model is developed to segment lung abnormalities and generate TB predictions. In this phase, the preprocessed data is used to train a deep learning model. The process starts with feature extraction from CXR images, including the position, area, perimeter, and other geometric information of TB-related abnormalities identified through segmentation using a U-Net deep learning model. A hybrid classification strategy is then applied, incorporating a weighted multiplier within a support vector machine framework. Based on these extracted features, the model generates decision rules and produces diagnosis recommendations that can serve as a preliminary tool for TB screening. Furthermore, hyperparameter tuning is conducted to optimize the model\u0026rsquo;s performance. By adjusting parameters such as learning rate, batch size, and number of layers, the model can achieve better accuracy and generalization. The performance of the trained lung segmentation model is analyzed using the mean Intersection over Union (IoU), by comparing predicted segmentation masks with ground truth annotations. The sensitivity and specificity of the TBScreen.AI model is assessed using Support Vector Machine and Random Forest classifiers, based on labeled TB features. The final output is categorized into two classes: TB or non-TB.\u003c/p\u003e \u003cp\u003eThe third stage (model optimization) aims to identify the best-performing model using the available dataset. Performance evaluation is done using various metrics to ensure the model\u0026rsquo;s reliability across different scenarios. Once the most accurate and stable model is selected, the model will be converted into an Application Programming Interface (API). This will enable seamless integration with other systems and facilitate real-time predictions or classifications in a production environment.\u003c/p\u003e \u003cp\u003eThe fourth stage (UI development) applies the best-performing model to a prototype web-based application. The interface includes a login page, an upload page for CXR image submission and analysis, a results page displaying the TBScreen.AI classification and detailed TB-related findings, and a list page documenting user identity and input history on the device. This initial design will be reviewed by potential users during the TBScreen.AI validation phase.\u003c/p\u003e\n\u003ch3\u003ePart 2: Validation phase of the TBScreen.AI\u003c/h3\u003e\n\u003cp\u003eThis phase employs a cross-sectional design to validate the TBScreen.AI model among patients with signs and symptoms of TB. Prospective data collection for validation takes place in Balkesmas Klaten, Respira Lung Hospital, Sardjito General Hospital, and Mimika Regional Hospital. The data collection is planned over a 9-month period, from the end of August 2025 to the end of May 2026. The study population comprises all presumptive TB patients attending the health facilities.\u003c/p\u003e \u003cp\u003eWe calculated the sample size using the formula for estimating a population proportion, assuming a 99% confidence level (Z\u0026thinsp;=\u0026thinsp;2.57), 5% margin of error, and design effect of 1. Based on reported TB bacteriological positivity rates in Indonesia (11.56%-12.82%) and Mimika (35.8%) (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e), we applied hypothesized TB positivity proportions among presumptive TB patients \u003cem\u003e(N)\u003c/em\u003e in Yogyakarta province \u003cem\u003e(\u003c/em\u003e14% of 3,759,500), Klaten Municipality \u003cem\u003e(\u003c/em\u003e14% of 1,302,648), and Mimika Municipality (35% of 313,016) (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Allowing for 10% non-response, we will recruit 800 participants from Papua and 584 participants from study sites in Yogyakarta and Klaten. Mimika is intentionally oversampled as the primary remote implementation setting and to ensure an adequate number of TB-positive cases for validation. In total, the TBScreen.AI model will be evaluated on at least 1,384 CXR images from newly enrolled presumptive TB patients. In addition, the study will assess barriers to care and out-of-pocket expenditures among patients receiving TB treatment. The sample size for this descriptive component was calculated based on an assumed TB prevalence of 35%, with a 95% confidence level (Z\u0026thinsp;=\u0026thinsp;1.96), 5% margin of error, and design effect of 1. Based on standard formulas for descriptive studies, we plan to recruit 350 patients receiving TB treatment. Overall, the total planned sample size is 1,734 participants.\u003c/p\u003e \u003cp\u003eThe inclusion criteria are (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) presumptive TB patients presenting with signs and symptoms suggestive of TB (\u003cem\u003ee.g.\u003c/em\u003e, persistent cough, unintentional\u0026thinsp;\u0026ge;\u0026thinsp;5% weight loss, night sweats) and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) age\u0026thinsp;\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e\u0026ge;\u003c/span\u003e\u0026thinsp;15 years. The exclusion criteria are (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) incomplete clinical data and (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) receipt of fixed-dose combination of anti-TB therapy for more than 7 days. We will collect participants' demographic and clinical data, including signs, symptoms, CXR images, and bacteriological examination results.\u003c/p\u003e \u003cp\u003eEligible individuals with presumptive TB will be consecutively enrolled. Indications for CXR imaging and bacteriological testing will follow routine procedures at each health facility. CXR examinations will be performed in accordance with WHO guidelines (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). Sputum samples will be collected for bacteriological testing (GeneXpert MTB/RIF, acid-fast bacilli/AFB smear microscopy, or culture, as available). This phase will apply a quadruple-blinding, in which the participants, radiographers, radiologists, and treating physicians will be blinded to the TBScreen.AI results.\u003c/p\u003e \u003cp\u003eFor the qualitative component, the research team will recruit 2\u0026ndash;3 specialists (radiologists, internists, or pulmonologists) and 13\u0026ndash;14 presumptive TB patients at each regional site for in-depth interviews. In addition, approximately 6\u0026ndash;8 health workers (general practitioners, nurses, and paramedics) at each facility will participate in a focus group discussion. These data collection strategies are designed to achieve thematic saturation regarding barriers to TB services and opportunities for future implementation of TBScreen.AI technology.\u003c/p\u003e \u003cp\u003eAI analysis will be performed after completion of data collection. CXR images obtained from each study site will be uploaded to the TBScreen.AI application and automatically analyzed to detect patterns of lung abnormalities. As a second reader, TBScreen.AI outputs will be considered predictive results rather than definitive diagnoses. The AI interpretations will be compared with radiologist readings and the patients\u0026rsquo; final diagnoses established by site physicians, based on comprehensive examinations, including anamnesis, physical examination, laboratory investigations, and follow-up assessments, which will serve as the reference standards.\u003c/p\u003e\n\u003ch3\u003ePart 3: Assessment of access barriers and inclusive implementation strategies\u003c/h3\u003e\n\u003cp\u003eThis phase complements the development and validation of TBScreen.AI by examining health system, sociocultural, and individual-level barriers to TB service access, and by identifying equity-oriented strategies for inclusive AI-assisted CXR screening in diverse Indonesian settings. Specifically, this study component will (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) identify barriers and facilitators to TB screening and diagnostic services across gender, age, disability, and geographic groups; (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) explore health providers\u0026rsquo; perceptions, acceptability, trust, and concerns regarding AI-based TB screening tools, and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) generate evidence-informed, context-sensitive recommendations for the inclusive integration of TBScreen.AI into routine TB services, particularly in remote and resource-constrained settings.\u003c/p\u003e \u003cp\u003eThis phase will employ a mixed-methods design, combining patient exit surveys with qualitative interviews and focus group discussions (FGDs). The integration of quantitative and qualitative data will allow triangulation of findings related to access barriers, user experiences, and implementation feasibility.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePatient Exit Survey\u003c/h2\u003e \u003cp\u003eAll participants will be invited to complete a structured exit survey following their TB screening and diagnostic procedures. Trained data collectors will administer the survey in Bahasa Indonesia or the relevant local language, depending on participant preference and literacy level. For participants with communication, cognitive, or psychosocial impairments, reasonable accommodations will be provided, including caregiver assistance. Where direct participation is not feasible, caregivers may respond as proxies, and this will be clearly documented.\u003c/p\u003e \u003cp\u003eThe exit survey will collect data across four domains:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSociodemographic and clinical characteristics, including age, sex, ethnicity, occupation, smoking status, comorbidities, and disability status (using self-reported functional limitations guided by the Washington Group Short Set on Functioning (WG-SS) (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e);\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eTB awareness and care-seeking behavior, including symptom recognition, perceived severity, and timing of healthcare utilization;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eExperiences of accessing TB services, including availability of services, perceived quality of care, stigma or discrimination, and cultural or gender-related constraints, and\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eEconomic barriers, including direct medical costs, non-medical costs (transport, food, accommodation), and income loss related to TB-related healthcare visits.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe questionnaire is adapted from existing validated tools (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e) and piloted prior to data collection. The case-report form and interview guide can be accessed in the S1 Appendix and S2 Appendix.\u003c/p\u003e \u003cp\u003e \u003cb\u003eQualitative Interviews\u003c/b\u003e \u003c/p\u003e \u003cp\u003e1. Semi-structured interviews with patients\u003c/p\u003e\u003cp\u003eSemi-structured interviews will be conducted with a purposive subsample of presumptive TB patients to capture diverse experiences across gender, age, disability status, and geographic location. Interviews will explore: (a) pathways to TB care and decision-making processes; (b) experiences of stigma, discrimination, or exclusion, and; (c) perceived accessibility and acceptability of TB.\u003c/p\u003e \u003cp\u003eParticipants with disabilities may be accompanied by caregivers. When communication barriers persist despite accommodations, caregiver proxy interviews will be conducted, with careful attention to ethical considerations.\u003c/p\u003e \u003cp\u003e2. Key informant interviews with health specialists\u003c/p\u003e \u003cp\u003eKey informant interviews will be conducted with specialists involved in TB service delivery, including radiologists, pulmonologists, and internists. These interviews will examine: (a) current diagnostic workflows and bottlenecks, (b) perceived opportunities and risks of AI integration into TB service, (c) concerns related to diagnostic responsibility, accountability, and trust in AI systems, and (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) system-level requirements for equitable and sustainable AI deployment.\u003c/p\u003e \u003cp\u003eBrief demographic information (age, professional role, years of experience) will be collected to contextualize findings.\u003c/p\u003e \u003cp\u003e3. Focus group discussions (FGD) with frontline health workers\u003c/p\u003e \u003cp\u003eOne FGD will be conducted at each study site with frontline healthcare workers (general practitioners, nurses, other paramedics involved in TB service) who have at least one year of experience delivering TB services. FGDs (60\u0026ndash;90 minutes) will explore: (a) practical workflow implications of AI-assisted TB screening, (b) usability and acceptability of TBScreen.AI in routine practice, (c) perceived benefits and risks for different patient groups, and (d) recommendations for inclusive and context-appropriate implementation. Qualitative interview and FGD guides are provided in S3 Appendix.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAI System\u003c/h3\u003e\n\u003cp\u003eThe model and web-based application are co-developed by teams from the Department of Computer Science and Electronics, Faculty of Mathematics and Natural Sciences, Universitas Gadjah Mada, and the Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, with additional clinical and radiological input from RSUD Mimika.\u003c/p\u003e\n\u003ch3\u003eTBScreen.AI model\u003c/h3\u003e\n\u003cp\u003eTo develop the AI model, the Python programming language was utilized along with the OpenCV library, which provides a wide range of functionalities for image processing and computer vision tasks. In the preprocessing stage, several image enhancement methods were explored and compared, including Histogram Equalization, Gamma Correction, Retinex, and CLAHE. Additionally, segmentation techniques were examined, where both Single U-Net and Double U-Net architectures were tested to determine their effectiveness.\u003c/p\u003e \u003cp\u003eFollowing the preprocessing and segmentation stages, the modeling of artificial intelligence was carried out using two different machine learning algorithms, namely Support Vector Machine and Random Forest. These models were compared to evaluate their performance in handling the classification task. Based on the preliminary experiments, the best performance was achieved using a combination of CLAHE for image enhancement, Double U-Net for segmentation, and Support Vector Machine for classification. This combination showed the most promising results in the initial testing phase. Therefore, this model will be applied to the web application.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eTBScreen.AI application\u003c/h2\u003e \u003cp\u003eTo make the system accessible from various platforms, including laptops, desktop computers, and mobile devices, it was decided to develop the application as a web-based system. A web-based application is considered more flexible and can be easily accessed through internet browsers without the need to install anything on the user\u0026rsquo;s device. The development of this web application was done using a combination of several programming languages, including Python, PHP, HTML, JavaScript, CSS, Flask, and SQLite. Each language has its own function in building the system, such as backend processing, user interface design, and database communication.\u003c/p\u003e \u003cp\u003eThe application is designed to support at least two main types of users. The first one is a guest user who can access the system without the need to log in. This type of user can perform self-screening independently and immediately receive the results. The second type is a healthcare worker, who is required to log in to the system. After logging in, healthcare professionals can perform screenings for multiple patients. The results of these screenings will then be stored in the system's database for further review and record-keeping. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows an example of the interface when the screening analysis results are displayed on the screen. The developed application can be accessed publicly through the following link: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://tbscreen.ai/\u003c/span\u003e\u003cspan address=\"https://tbscreen.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eAnalysis Plan\u003c/h2\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003eDiagnostic accuracy\u003c/h2\u003e \u003cp\u003eIn the TBScreen.AI development and training phase, receiver operating characteristic (ROC) analysis will be performed, and the area-under-curve (AUC) will be calculated to assess the diagnostic performance against a composite reference standard (CRS). The CRS consists of clinical assessment, radiologist readings, laboratory examinations, and follow-up findings (if available). During the AI-CAD validation phase, diagnostic performance will be evaluated using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with GeneXpert results and final clinical diagnoses serving as reference standards. Descriptive analyses will be used to summarize participants\u0026rsquo; sociodemographic and clinical characteristics. Statistical analysis will be performed using STATA 18.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eQuantitative survey data\u003c/h2\u003e \u003cp\u003eThe three primary outputs and their corresponding analytical approaches are outlined below:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDelay in TB diagnosis and treatment, operationalized using patient-reported timelines (\u003cem\u003ee.g.\u003c/em\u003e, symptom onset to first care-seeking; first care-seeking to diagnostic evaluation; and, where applicable, diagnosis to treatment initiation). Delays will be analyzed as continuous measures (days) and categorized using pre-specified cut-offs appropriate to the Indonesian context.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBarriers to accessing TB services, including health system barriers (e.g., availability of diagnostics/medicines, waiting time, referral complexity), geographic and transport constraints, sociocultural barriers (including stigma and gender-related constraints), and disability-related accessibility barriers. Barrier items will be summarized as proportions (overall and by category) and, where multi-response options are used, as multi-response frequencies.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePatient-incurred costs and financial burden, aligned with the WHO TB Patient Cost Survey framework. Costs will be estimated at the visit level and aggregated to the TB episode level, disaggregated into direct medical costs, direct non-medical costs, and indirect costs (income loss and related productivity costs).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eNet out-of-pocket (OOP) expenditure will be calculated as total patient-incurred costs minus any reported reimbursements. Catastrophic TB-related costs will be defined as total TB-related patient-incurred costs exceeding 20% of annual household income, with sensitivity analyses using alternative thresholds (\u003cem\u003ee.g.\u003c/em\u003e, 10% and 30%)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eStratified and equity-focused analysis\u003c/h2\u003e \u003cp\u003eAll primary outputs will be analyzed overall and stratified by relevant equity dimensions and implementation-relevant characteristics, including (as sample size permits): gender, age group, disability status (including severity proxy), ethnicity, rural/urban residence, study site/province, and insurance coverage. For financial outcomes, analyses will also stratify by diagnostic vs treatment pathway and by whether a companion was required.\u003c/p\u003e \u003cp\u003eDescriptive results will be presented using frequencies (%) for categorical variables and medians (IQR) for skewed continuous variables (\u003cem\u003ee.g\u003c/em\u003e., costs, delays). Group differences will be assessed using chi-square tests for categorical outcomes and non-parametric tests (Wilcoxon rank-sum/Kruskal\u0026ndash;Wallis) for continuous outcomes. Effect sizes (risk differences/odds ratios, median differences) will be reported with 95% confidence intervals.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eMultivariable modelling\u003c/h2\u003e \u003cp\u003eMultivariable models will be used to identify factors associated with: (i) longer delays; (ii) reporting key barriers; and (iii) higher OOP expenditure and catastrophic costs. Delay outcomes will be modelled using appropriate approaches based on distributions (e.g., log-transformed linear models or time-to-event style models if feasible). OOP expenditure will be modelled using generalized linear models (log link, gamma distribution), while catastrophic costs will be modelled using logistic regression. Models will include key covariates reflecting equity dimensions (gender, disability status, rurality, ethnicity), socioeconomic indicators, insurance status, and study site, with robust standard errors to account for clustering by facility/site, will be analyzed descriptively to characterize access barriers and out-of-pocket expenditures across participant subgroups. Where appropriate, subgroup analyses by gender, age group, disability status, and study site will be conducted.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eQualitative data analysis\u003c/h2\u003e \u003cp\u003eAudio records are transcribed manually by the research team and translated into English. The quality of data will be checked before the coding and analysis. Transcripts of the interview will be thoroughly read to identify patterns and assess the fitness and relevance of the information. Thematic analysis will be conducted with Dedoose. The codebook will be developed and applied to a sample transcript, then cross-checked among researchers. Any discrepancies will be discussed and resolved, and the codes refined accordingly. This process is repeated until the codebook adequately captures the interview content. The final codebook is applied to all transcripts. Researchers will analyze emerging trends and patterns in the data and iteratively refine thematic categories. Data saturation will be considered achieved when no new themes emerge and similar findings are consistently observed across interviews.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eEthical considerations\u003c/h2\u003e \u003cp\u003eThe study will be conducted in accordance with the Declaration of Helsinki (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e). Ethical approvals were obtained from the Health Research Ethics Committee, National Research and Innovation Agency, Indonesia (001/KE.03/SK/01/2025) and the Medical Health Research Ethics Committee, Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia (KE/FK/1439/EC/2025). All eligible patients recruited for the validation are given a detailed \u0026lsquo;participant information sheet\u0026rsquo; and will sign an informed consent before participating in this study.\u003c/p\u003e \u003cp\u003eDuring the validation phase, nurses at the study sites will provide brief information and offer patients the opportunity to meet with a member of the research team or staff to receive research details. The research team or staff will explain the research verbally, based on the information provided in the informed consent form, including the potential risks and the risk mitigation program. The research team will ask participants to sign the consent form, while the legal guardian or caregiver will sign on behalf of participants with intellectual or psychosocial disabilities. Participants can withdraw from the research during any research process or health procedures. The research teams can also decide to withdraw the participants from the research for urgent medical reasons. There will be no consequences after the withdrawal process, and patients will be diagnosed and treated according to routine healthcare practice.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis is the first research protocol to develop and evaluate an AI-CAD tool for TB detection that incorporates gender equality, disability, and social inclusion considerations. The development of TBScreen.AI will examine group-based differences that may influence CXR interpretation within an AI framework. Furthermore, to promote an inclusive implementation of digital health technologies in TB services, the study will also investigate sociocultural determinants of access, including gender norms, stigma, accessibility barriers for persons with disabilities, and out-of-pocket expenditures. The findings are expected to inform equity-oriented TB screening strategies and promote more inclusive access to screening services. Wider implementation of TBScreen.AI could contribute to national TB control efforts by serving as a second reader for CXR interpretation. This may help reduce diagnostic delays due to radiologist shortages and prolonged waiting times, particularly in remote and resource-limited settings.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eACF Active case finding\u003c/p\u003e \u003cp\u003eAFB Acid-fast bacilli\u003c/p\u003e \u003cp\u003eAI Artificial intelligence\u003c/p\u003e \u003cp\u003eAPI Application Programming Interface\u003c/p\u003e \u003cp\u003eAUC Area-under-curve\u003c/p\u003e \u003cp\u003eCAD Computer-aided detection\u003c/p\u003e \u003cp\u003eCLAHE Contrast Limited Adaptive Histogram Equalization\u003c/p\u003e \u003cp\u003eCRS Composite Reference Standard\u003c/p\u003e \u003cp\u003eCXR Chest radiography\u003c/p\u003e \u003cp\u003eFGD Focus group discussion\u003c/p\u003e \u003cp\u003eIoU Intersection over Union\u003c/p\u003e \u003cp\u003eIQR Interquartile range\u003c/p\u003e \u003cp\u003eNPV Negative predictive value\u003c/p\u003e \u003cp\u003eOOP Out-of-pocket\u003c/p\u003e \u003cp\u003ePPV Positive predictive value\u003c/p\u003e \u003cp\u003ePTB Pulmonary tuberculosis\u003c/p\u003e \u003cp\u003eRIF Rifampicin\u003c/p\u003e \u003cp\u003eTB Tuberculosis\u003c/p\u003e \u003cp\u003eWG-SS Washington Group Short Set on Functioning\u003c/p\u003e \u003cp\u003eWHO World Health Organization\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eEthical clearance has been obtained from the Health Research Ethics Committee, the National Research and Innovation Agency, Indonesia (001/KE.03/SK/01/2025) and the Medical Health Research Ethics Committee, Faculty of Medicine, Public Health, and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia (KE/FK/1439/EC/2025). Study permits were obtained from all study sites prior to data collection. Written consent and assent, either electronic or paper-based, are obtained from each participant before undergoing any research procedures. Participants can withdraw from the research during any research process or health procedures. Participant data protection and confidentiality are guaranteed.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eInformed consents are obtained from each participant for any data to be published in this manuscript.\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials\u003c/p\u003e\n\u003cp\u003eAll data generated or analysed during this study are included in this published article and its supplementary information files.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003eFunding\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis study is supported by grants from\u0026nbsp;the Department of Foreign Affairs and Trade (DFAT) Australia through KONEKSI (Collaboration for Knowledge, Innovation, and Technology Australia and Indonesia) under the supervision of Cowater International. The views expressed in this publication are the authors\u0026rsquo; alone and are not necessarily the views of the Australian Government.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026rsquo; contributions\u003c/p\u003e\n\u003cp\u003eAMS, W, TM, AP, HS, LA designed the study. AMS, TM, AP, RT, DU, NS, LA, NRA, JRP, AS, CAP contributed to the development of the interview guides, survey questionnaire, and study protocol. W, AMS, NHA, NRA, JRP, LC contributed to the development and validation of the AI-CAD model. AMS, TM, HS, AP developed analyses plan and performed sample size calculation. All authors are part of the core research team that meets regularly to discuss the design and implementation of the study. All authors have contributed intellectually by providing input to the manuscript, and all have read and approved the submitted version.\u003c/p\u003e\n\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eThe authors gratefully acknowledge Chatarina Sari (Director of the YAKKUM Rehabilitation Center), Ayatullah (SAPDA), and their team in SAPDA and YAKKUM Rehabilitation Center for their support in strengthening the inclusivity approach of this study. We also extend our sincere appreciation to Cathy Vaughan and Min-hui Law from Nossal Institute for Global Health for their valuable methodological insights. We gratefully acknowledge Faisal Dharma Adhinata and Adi Suheryadi from the Department of Computer Science and Electronics at Universitas Gadjah Mada for their assistance with the AI system development. We acknowledge the critical inputs of the Tuberculosis Task Force and the General Directorate of Health Services for Vulnerable Populations, Ministry of Health, Republic of Indonesia, the Mimika District Health Office, also the Expert Committee on Tuberculosis (Komli TB) and Tuberculosis Research Network (JetSet TB) in Indonesia. Our gratitude also due to Silvester Sikora, the research assistant and field coordinator at the Mimika study site. We thank Erwan Hartadi and Rosalia Ratna from the Center for Tropical Medicine, Faculty of Medicine, Public Health and Nursing, Universitas Gadjah Mada, Yogyakarta, Indonesia, for their assistance in administering the ethics clearance and research permits process. We are grateful to the healthcare workers and staff at RSUP Dr. Sardjito, RS Paru Respira, Balkesmas Klaten, and RSUD Kabupaten Mimika for their input in the development of this protocol.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWHO. Global tuberculosis report [Internet]. Geneva: World Health Organization. Licence: CC BY-NC-SA 3.0 IGO. 2025. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://iris.who.int/server/api/core/bitstreams/e97dd6f4-b567-4396-8680-717bac6869a9/content\u003c/span\u003e\u003cspan address=\"https://iris.who.int/server/api/core/bitstreams/e97dd6f4-b567-4396-8680-717bac6869a9/content\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. The second national TB inventory study in Indonesia. Global Tuberculosis Report 2024. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. WHO consolidated guidelines on tuberculosis. Module 2: screening \u0026ndash; systematic screening for tuberculosis disease. Licence: C. Geneva: World Health Organization; 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. Chest Radiography in Tuberculosis Detection - Summary of current WHO recommendations and guidance on programmatic approaches. WHO Libr Cat Data. 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYunus RE. Radiology Loading and Coverage Hours in Indonesia. Korean J Radiol. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKementerian Kesehatan Indonesia. Profil Kesehatan Indonesia (Health Profile of Indonesia) 2024 [Internet]. Minist. Heal. Indones. 2025. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://drive.google.com/file/d/1-lNRA3k9o9jM5vGacbnKY4OZorUQ-_Sc/view\u003c/span\u003e\u003cspan address=\"https://drive.google.com/file/d/1-lNRA3k9o9jM5vGacbnKY4OZorUQ-_Sc/view\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedrazzoli D, Lalli M, Boccia D, Houben R, Kranzer K. Can tuberculosis patients in resource-constrained settings afford chest radiography? Eur Respir J. 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Cleeff MRA, Kivihya-Ndugga LE, Meme H, Odhiambo JA, Klatser PR. The role and performance of chest X-ray for the diagnosis of tuberculosis: A cost-effective analysis in Nairobi, Kenya. BMC Infect Dis. 2005. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1471-2334-5-111\u003c/span\u003e\u003cspan address=\"10.1186/1471-2334-5-111\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Cited: in:: PMID: 16343340.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrija G, Blažić I, Frush DP, Hierath M, Kawooya M, Donoso-Bach L, Brkljačić B. How to improve access to medical imaging in low- and middle-income countries ? eClinicalMedicine. 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZou J, Schiebinger L. AI can be sexist and racist \u0026mdash; it\u0026rsquo;s time to make it fair. Nature. 2018. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/d41586-018-05707-8\u003c/span\u003e\u003cspan address=\"10.1038/d41586-018-05707-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQin ZZ, Van der Walt M, Moyo S, Ismail F, Maribe P, Denkinger CM, Zaidi S, Barrett R, Mvusi L, Mkhondo N, et al. Computer-aided detection of tuberculosis from chest radiographs in a tuberculosis prevalence survey in South Africa: external validation and modelled impacts of commercially available artificial intelligence software. Lancet Digit Heal. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S2589-7500(24)00118-3\u003c/span\u003e\u003cspan address=\"10.1016/S2589-7500(24)00118-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Cited: in:: PMID: 39033067.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBPS Statistics Indonesia. Profile of Ethnic Groups and Regional Language Diversity: 2020 Population Census Long Form Results. BPS-Statistics Indones; 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePutri NK, Syahansyah RJ. Disability-based disparities under universal health coverage among chronically ill adults during the COVID-19 pandemic in Indonesia: an interrupted time series analysis. Glob Health Action. 2025; doi: 10.1080/16549716.2025.2581946. Cited: in:: PMID: 41201369.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHitzl W, Reitsamer HA, Hornykewycz K, Mistlberger A, Grabner G. Application of discriminant, classification tree and neural network analysis to differentiate between potential glaucoma suspects with and without visual field defects. J Theor Med. 2003. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/10273360410001728011\u003c/span\u003e\u003cspan address=\"10.1080/10273360410001728011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBalki I, Amirabadi A, Levman J, Martel AL, Emersic Z, Meden B, Garcia-Pedrero A, Ramirez SC, Kong D, Moody AR et al. Sample-Size Determination Methodologies for Machine Learning in Medical Imaging Research: A Systematic Review. Can. Assoc. Radiol. J. 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLomauro A, Aliverti A. Sex and gender in respiratory physiology. Eur Respir Rev. 2021; doi: 10.1183/16000617.0038-2021. Cited: in:: PMID: 34750114.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchneider JL, Rowe JH, Garcia-de-Alba C, Kim CF, Sharpe AH, Haigis MC. The aging lung: Physiology, disease, and immunity. Cell. 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEkstr\u0026ouml;m M, Mannino D. Research race-specific reference values and lung function impairment, breathlessness and prognosis: Analysis of NHANES 2007\u0026ndash;2012. Respir Res. 2022; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12931-022-02194-4\u003c/span\u003e\u003cspan address=\"10.1186/s12931-022-02194-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Cited: in:: PMID: 36182912.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohari J, Sharifudin MA, Rahman AA, Omar AS, Abdullah AT, Nor S, Lam WC, Yusof MI. Relationship between pulmonary function and degree of spinal deformity, location of apical vertebrae and age among adolescent idiopathic scoliosis patients. Singapore Med J. 2016; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.11622/smedj.2016009\u003c/span\u003e\u003cspan address=\"10.11622/smedj.2016009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Cited: in:: PMID: 26831315.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRagborg LC, Dragsted C, Ohrt-Nissen S, Mortensen J, Gehrchen M, Dahl B. Pulmonary function in patients with idiopathic scoliosis 40 years after diagnosis. Spine J. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.spinee.2024.07.006\u003c/span\u003e\u003cspan address=\"10.1016/j.spinee.2024.07.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Cited: in:: PMID: 39097102.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMinistry of Health Republic of Indonesia. Tuberculosis control program report 2023. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLestari T, Kamaludin, Lowbridge C, Kenangalem E, Poespoprodjo JR, Graham SM, Ralph AP. Impacts of tuberculosis services strengthening and the COVID-19 pandemic on case detection and treatment outcomes in Mimika District, Papua, Indonesia: 2014\u0026ndash;2021. PLOS Glob Public Heal [Internet]. 2022;2:1\u0026ndash;19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pgph.0001114\u003c/span\u003e\u003cspan address=\"10.1371/journal.pgph.0001114\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMinistry of Home Affairs of Indonesia. Population of the district/municipality [Internet]. 2020. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pelita.kemendagri.go.id/kemendagri/dataset/257/tabel-data\u003c/span\u003e\u003cspan address=\"https://pelita.kemendagri.go.id/kemendagri/dataset/257/tabel-data\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWashington Group on Disability Statistics. The Washington Group Short Set on Functioning (WG-SS) [Internet]. 2020. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.washingtongroup-disability.com/fileadmin/uploads/wg/Documents/Questions/Washington_Group_Questionnaire__1_-_WG_Short_Set_on_Functioning.pdf\u003c/span\u003e\u003cspan address=\"https://www.washingtongroup-disability.com/fileadmin/uploads/wg/Documents/Questions/Washington_Group_Questionnaire__1_-_WG_Short_Set_on_Functioning.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWHO. Tuberculosis patient cost surveys: a handbook. World Health Organization; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrganization WH. Consolidated guidance on tuberculosis data generation and use. Module 4. Surveys of costs faced by households affected by tuberculosis. Geneva: World Health Organization; 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePractice IEG. clinical. Scientific guideline European Medicines Agency (EMA) [Internet]. 2002. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline\u003c/span\u003e\u003cspan address=\"https://www.ema.europa.eu/en/ich-e6-good-clinical-practice-scientific-guideline\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, computer-aided detection, tuberculosis screening, chest X-ray, tuberculosis, gender, disability, remote areas, inclusivity","lastPublishedDoi":"10.21203/rs.3.rs-8937526/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8937526/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eIndonesia ranks second in tuberculosis (TB) burden globally and continues to face a significant diagnosis gap, particularly in remote areas with limited access to radiologists. Approximately 14% of new TB cases are left undetected by TB services. Utilizing artificial intelligence-based computer-aided detection (AI-CAD) applied to chest X-ray (CXR) imaging offers a potential solution to improve TB screening. However, AI systems may exacerbate existing inequities if they are developed and implemented without consideration of gender, age, disability, ethnicity, and socioeconomic barriers to care. This research aims to develop and train an AI-CAD model that considers the balance of gender, age, ethnicity, and disability status. The study will validate an AI-CAD model for TB screening among different socio-cultural and disability status groups. Additionally, it will assess barriers to TB services and inclusive strategies for AI-CAD implementation in remote areas of Indonesia.\u003c/p\u003e\u003ch2\u003eMethod\u003c/h2\u003e \u003cp\u003eThis mixed-methods study comprises three integrated components. Part I focuses on the development of TBScreen.AI, an AI-CAD model trained on gender-, age-, ethnicity-, and disability-balanced CXR data from hospitals in Java and Papua. Part II uses a prospective cross-sectional design to evaluate the diagnostic accuracy of TBScreen.AI among individuals with presumptive TB attending four health facilities in Indonesia. AI-generated interpretations are compared with radiologist readings and the final patient diagnosis established by site physicians, based on comprehensive examinations, including anamnesis, physical examination, laboratory investigations, and follow-up assessments. Part III integrates quantitative and qualitative methods to assess access to TB services and to identify inclusive implementation strategies that support equitable deployment of AI-assisted screening.\u003c/p\u003e\u003ch2\u003eDiscussion\u003c/h2\u003e \u003cp\u003eThis study investigates the development of AI-CAD tools for TB case detection using digital chest radiography, while also considering gender equality, disability, and social inclusion. The implementation of an accurate and inclusive AI-CAD tool as a second interpreter of CXR images for TB case detection will prevent treatment delays due to radiologist unavailability and reduce waiting time for CXR image interpretation in remote areas with limited health resources.\u003c/p\u003e","manuscriptTitle":"TBScreen.AI: Study Protocol for an Inclusive AI-Based Chest X-Ray Screening System for Tuberculosis in Remote Areas of Indonesia","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-30 19:23:34","doi":"10.21203/rs.3.rs-8937526/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8304659f-bf71-453a-9731-5b4edc3585f3","owner":[],"postedDate":"March 30th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-30T19:23:34+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-30 19:23:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8937526","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8937526","identity":"rs-8937526","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}