PNG Medical Blog

Background

Tuberculosis (TB), although preventable and curable, remains a substantial global health challenge, with the World Health Organization (WHO)[1] estimating a quarter of the world’s population has been infected with the bacteria, and 5-10% of infected persons eventually developing TB disease.[2] In 2022, TB was the world’s second leading infectious killer, ranking above HIV/AIDS, with only COVID-19 having claimed more lives.[3]

Accurate and timely diagnosis of TB is critical for effective disease management and control.

Chest X-ray (CXR) is a mainstay in diagnosing chest/lung pathologies because it is cost-effective and easily deployable in low-resource settings, making it especially useful in countries where TB poses a high disease burden. The limiting factor of using CXR is the availability of trained human readers to interpret the X-rays. To address this gap, AI tools have been developed to assist in interpreting CXRs, enhancing diagnostic capacity by enabling both diagnosis and screening before symptoms develop.

Paediatric TB diagnosis remains challenging due to the variable presentation of the disease and the limited specificity of clinical symptoms and traditional diagnostic methods. In Papua New Guinea (PNG), the diagnosis of TB in the paediatric population is a clinical diagnosis confirmed by CXR findings and laboratory tests.

Since the early 2000s, there is increasing evidence showing the value of CAD systems in imaging diagnosis, culminating with WHO stating in 2020 that “CAD may be used as an alternative to human reader interpretation of plain digital CXR for screening and triage for TB.”[4] This development, however, was limited in application to persons aged 15 and above. Pediatric TB diagnosis is particularly challenging due to differences in disease presentation, lower bacterial load, and difficulty in obtaining sputum samples, making it harder to develop reliable CAD systems. Since most advancements in the field have been made with adult populations, the application of these technologies in paediatric TB diagnosis still requires comprehensive evaluation.

Artificial Intelligence

Healthcare is a key facet in the development of any civilization, and new technological advancements inevitably find their way into healthcare to ensure our survival as a species.

In the last decade or so, developments in machine learning and deep learning, cloud computing, big data and data management, telemedicine, and computational power have allowed computers to access large datasets for analysis and problem-solving, aiding humans in an unprecedented manner in solving our most complex problems.

Artificial Intelligence Governance

In 2018, the International Telecommunication Union (ITU) and the World Health Organization established the Focus Group on “Artificial Intelligence for Health” (FG-AI4H) to create a standard framework and guidelines for AI development in healthcare. The Focus Group interfaces various disciplines not only in medicine and healthcare but also in fields that converge where AI in healthcare is concerned, such as statistics, public health, machine learning, and ethics. The platform enables experts in various fields to collaborate to achieve their common goals.

Participants in the FG-AI4H fall into two categories: topic groups and working groups. Topic groups focus on specific health use cases (e.g., tumor tissue discrimination, tuberculosis, diabetes). They gather experts, compile data, and propose procedures to evaluate AI models for these health tasks. They finalize their work through consultation with the World Health Organization (WHO). Working groups address themes that affect all topic groups, such as ethical considerations or regulatory aspects. They define best practices, establish processes, and create reference documents for AI applications in health.

The goal is to improve AI tools, share evidence and case studies, and define best practices for AI applications in health.

Artificial Intelligence for Health

Deep learning is a branch of machine learning that uses artificial neural networks. One of its defining features is the use of neural network parameters based on complex automated processes, which are tuned to run multiple layers of iterative training and can involve billions of parameters.

The training, validation, testing, and performance comparisons of CAD systems require many chest radiographs. Since creating a large, annotated medical image dataset is not easy, most researchers rely on publicly available CXR datasets.

Although AI models need a large dataset of annotated X-ray images to be trained, these models can also be capable of “end-to-end training,” which requires no domain-specific knowledge on the part of the operator to train the AI except images with the relevant labels. This was demonstrated by Esteva et al. in their 2019 landmark paper.

AI models trained on a specific dataset will need validation over a wider population and geographical setting where the variance from the dataset will be more pronounced. For example, AI trained on recognizing skin lesions in one ethnic group in one location may give varying results when applied to people in another country. AI models cannot extrapolate data but can only learn patterns present in the data used to train them. As a result, the training data is just as important as the algorithms, and the data must not only be of high quality and large quantity but also cover the different possibilities and combinations of cases.

The rationale for integrating artificial intelligence (AI) with chest X-ray (CXR) imaging for tuberculosis (TB) diagnosis is founded on the ability of AI systems to enhance the accuracy, efficiency, and scalability of radiographic analysis. CXR is a readily available, cost-effective screening tool for detecting various pulmonary diseases, including TB, which presents specific patterns recognizable by trained algorithms. AI, particularly deep learning models, can systematically analyze these images, identifying subtle patterns that may be missed by human eyes. This capability not only increases diagnostic accuracy but also enables rapid screening in high-volume, resource-limited settings where trained radiologists may be scarce.

AI models, trained on large, diverse datasets of annotated images, have demonstrated high accuracy and reliability in identifying TB, making them a valuable adjunct to clinical diagnosis and public health strategies (Wenxing Zhou et al., 2021)[5]. Such integration promises to streamline workflows in healthcare settings and improve patient outcomes by facilitating early and accurate detection.

Current AI Tools and Models

In recent years, the tuberculosis (TB) field has welcomed several computer-aided detection (CAD) products that provide automated and standardized interpretation of digital chest X-rays based on artificial intelligence. The AI4HLTH resource center from the Stop TB Partnership and FIND aims to list all CAD products that may be used for TB detection, irrespective of their stage of development, regulatory or policy approval to provide implementors and researchers with relevant information on CAD products.

The AI4HLTH database categorizes various AI tools for TB detection, including those certified for clinical use such as RAPID and qure.ai qXR, and those still undergoing certification like TXNet and Genki (Stop TB Partnership & FIND, n.d.). The database is updated regularly to ensure that the information provided remains current and relevant. The following are the current AI tools and models available, classified into three categories:

Certified and Market-Ready

• CAD4TB: Certified and CE Class II. Developed by Delft Imaging, intended for age group 4+ years.
• RAPID: CE Certified. Developed by Fujifilm (Japan) and PAI Health. Intended age group: 12+ years.
• InferRead CXR Chest: CE Certified. Developed by InferVision. Intended age group: 15+ years.
• JLK (JVIEWER): CE Certified Class II. Developed by JLK. Intended age group: 18+ years.
• Lunit INSIGHT CXR: CE Certified, Class II. Developed by Lunit. Intended age group: 4+ years.
• OXIPT: CE Certified, Class II. Developed by OXIPT. Intended age group: 12+ years.
• qure.ai qXR: CE Certified, Class II. Developed by qure.ai. Intended age group: 4+ years.
• Radiscan AIDR: CE Certified, Class II. Developed by Radiscan. Intended age group: 15+ years.
• VUNO MedChest X-ray Pro: Developed by VUNO. Certified for Korea MFDS Class IIa, intended age group: 10+ years.

Pending Certification

• TXnet (Artelus): Pending certification. Development stage: Validation. Intended age group: 18+ years.
• Genki (DEEPTEK): CE pending, FDA pending. On the market. Intended age group: 14+ years.
• Dr CADx: Pending certification. Development stage: Validation. Intended age group: 16+ years.
• XrayAME (EPCON): Pending certification. On the market. Intended age group: 18+ years.
• JF CXR-2: China NMPA tier-3 pending. On the market. Intended age group: 15+ years.
• TiSepX TB (MEDICAL IP): Korea MFDS pending. On the market. Intended age group: 20+ years.
• Imagiflex (Yantrakashi): Pending certification. Development stage: Validation. Intended age group: 8+ years.

Under Development

• OpenTB (provisional): Certification not available. Development stage: Under development. Intended age group: 18+ years.

Limitations and Challenges of AI in Healthcare

AI adoption in healthcare, especially for diagnosing pulmonary TB (PTB), faces several challenges. Many AI models, particularly deep learning ones, function as black boxes, making it difficult for clinicians to interpret their decision-making process. Trust in AI systems can be limited when the logic behind predictions is not transparent. Additionally, AI models require high-quality, well-annotated data to function effectively. Inconsistent or low-quality imaging data can reduce diagnostic accuracy, while differences in equipment, patient demographics, or imaging settings can create robustness issues, affecting generalization across healthcare settings. While AI models can perform well in controlled studies, their real-world effectiveness might differ. There are concerns about safety if AI systems replace rather than augment clinical judgment. Regular calibration and validation are necessary to maintain model performance across populations and over time.

Papua New Guinea (PNG) Context

AI holds promise in enhancing healthcare delivery in PNG, but the local context introduces specific challenges and opportunities. PNG has one of the highest TB burdens in the Western Pacific, and limited access to healthcare facilities, diagnostic labs, and skilled radiologists makes timely TB diagnosis difficult, increasing transmission risks. AI systems can help bridge the diagnostic gap by providing rapid analysis of chest X-rays in remote areas. AI-powered portable devices could offer immediate results, reducing delays in treatment initiation. Collaborations between PNG healthcare providers, international organizations, and AI developers could create customized tools for the country. Pilot studies using AI for TB screening could validate the effectiveness of these models and guide larger-scale implementation.

Conclusion

AI tools for TB diagnosis have demonstrated significant potential, especially in resource-limited settings. Certified models like qure.ai qXR and RAPID have already achieved market readiness, while others are in the pipeline. Continued research is needed to address the limitations of AI systems, including data quality and explainability challenges. Models must be validated across diverse populations to ensure equitable healthcare delivery. National healthcare policies should consider integrating AI into diagnostic workflows, with clear regulatory frameworks to ensure safety and efficacy. Training programs can help clinicians understand and use AI tools effectively.

Additional Topics

Future research should compare AI-based tools with traditional diagnostic methods, focusing on clinical outcomes, cost-effectiveness, and patient satisfaction. The use of AI in healthcare also raises ethical questions about data privacy, bias, and accountability. Robust regulatory oversight is essential to ensure patient safety and uphold ethical standards. Additionally, studies should assess how AI tools impact patient outcomes, including earlier diagnosis, faster treatment initiation, and improved recovery rates. Real-world evidence will be critical to validating the long-term benefits of AI systems.

Glossary

• AI: Artificial Intelligence
• CAD: Computer-Aided Detection
• CXR: Chest X-ray
• DL: Deep Learning
• FG-AI4H: Focus Group on Artificial Intelligence for Health
• ITU: International Telecommunication Union
• ML: Machine Learning
• MDR TB: Multi-Drug Resistant Tuberculosis
• PNG: Papua New Guinea
• PTB: Pulmonary Tuberculosis
• TB: Tuberculosis
• WHO: World Health Organization

References

1. World Health Organization. (2021). WHO consolidated guidelines on tuberculosis: Module 2 – Screening: systematic screening for tuberculosis disease. Geneva, World Health Organization. Retrieved from https://www.who.int/publications/i/item/9789240022676
2. World Health Organization. (2022). Global tuberculosis report 2022. Geneva, World Health Organization. Retrieved from https://www.who.int/teams/global-tuberculosis-programme/tb-reports/global-tuberculosis-report-2022
3. World Health Organization. (2022). Global Tuberculosis Report 2022 Factsheet. Geneva, World Health Organization. Retrieved from https://www.who.int/publications/m/item/global-tuberculosis-report-2022-factsheet
4. World Health Organization. (2021). Determining the local calibration of computer-assisted detection (CAD) thresholds and other parameters: A toolkit to support the effective use of CAD for TB screening. Geneva, World Health Organization. Retrieved from https://iris.who.int/bitstream/handle/10665/345925/9789240028616-eng.pdf
5. Wenxing Zhou, Guanxun Cheng, Ziqi Zhang, Litong Zhu, Stefan Jaeger, Fleming Y. M. Lure, & Lin Guo. (2022). Deep learning-based pulmonary tuberculosis automated detection on chest radiography: large-scale independent testing. Quantitative Imaging in Medicine and Surgery, 12(4), 1624-1636.
6. Lakhani P, Sundaram B. Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks. Radiology. 2017 Aug;284(2):574-582. doi: 10.1148/radiol.2017162326. Epub 2017 Apr 24. PMID: 28436741.
7. Wiegand, T., Lee, N., Pujari, S., Sing, M., Xu, S., Kuglitsch, M., Lecoutre, M., Riviere-Cinnamon, A., Weicken, E., Wenzel, M., Werneck Leite, A., Campos, S., & Quast, B. (n.d.). Whitepaper for the ITU/WHO Focus Group on Artificial Intelligence for Health. International Telecommunication Union (ITU) & World Health Organization (WHO). Retrieved from https://www.itu.int/en/ITU-T/focusgroups/ai4h/Documents/FG-AI4H_Whitepaper.pdf
8. Stop TB Partnership & FIND. (n.d.). AI4HLTH: Resource center on computer-aided detection products for the diagnosis of tuberculosis. Retrieved October 12, 2024, from https://ai4hlth.wixsite.com/website-1