A 2021 landscape analysis found eight digital stationary CXR products suitable for TB programmes and 21 portable and ultra-portable products [28], evidence of a growing global market in this product area with multiple competing manufacturers. However, published evidence of their use is limited. A portable system is currently being evaluated by the national TB programme of Peru in a Global Fund-supported project [31]. An early evaluation of an ultra-portable system conducted in Vietnam showed promising results, with no significant difference found in mean AI abnormality scores between radiographs produced using the ultra-portable system and those produced using a stationary system on the same patients, in both a national hospital and a community screening setting [32]. The Stop TB Partnership is currently undertaking a project deploying these systems in “hard-to-reach populations that currently face barriers to accessing services” in seven countries [33], early findings from which are highlighting the need for guidance on radiation protection for staff suitable for these ultra-portable machines to ensure more widespread uptake outside health facilities (Z.Z. Qin, Stop TB Partnership, Geneva, Switzerland personal communication). Two of the ultra-portable systems are available at negotiated prices through the Stop TB Partnership’s Global Drug Facility [34], and the Partnership offers a practical guide to their use for TB screening [35]. The Fujifilm FDR Xair System full kit is available for US$47 000, and the Delft Light full kit is available for US$66 750 via the Global Drug Facility. Thus, even ultra-portable X-ray systems are expensive, and this poses a major concern for the wider use of these technologies in low-income countries. CAD of CXR Historically, interobserver variation during radiograph interpretation and challenges in finding trained radiologists in low-resource settings have always been major barriers. In 2021, the WHO recommended CAD for the first time as an alternative to human interpretation of digital CXRs for the screening of TB in individuals aged ⩾15 years [9]. Product development in this space has been extremely rapid. In 2017, only one CAD product evaluating CXR findings suggestive of TB was commercially available [6]. A continually updated landscape analysis currently shows 13 TB-specific CAD products on the market [36], most with certifications including CE marking from the European Union [37], and several more in the development or validation stages [38]. These CAD products use AI to detect CXR abnormalities that are associated with TB (figure 2) [39]. Specifically, they use deep neural networks, a machine-learning technique, to train using datasets comprising large numbers of CXR images from patients with and without TB [35]. These training datasets are generally labelled based on other sources of information (e.g. labelled as “TB” or “not TB” based on linked results from molecular, culture or other tests performed on the same patients, as well as radiologist readings), allowing the neural network to compare its performance with the so-called ground truth of the labelling in a process known as supervised learning [35]. Producing accurate CAD models requires millions of iterations of this process using CXR image datasets that are as large as possible [35]. Most CAD products can be used with any computed or DR system that produces digital images in the correct DICOM (Digital Imaging and Communications in Medicine) format [28]. Some portable and ultra-portable X-ray systems can be also purchased in bundles that include CAD products [28]. CAD products typically give their CXR interpretation in the form of an abnormality score (e.g. between 0 and 1 or 0 and 100), with increasing numbers indicating a higher estimated likelihood of TB [40]. Abnormality scores can be given a threshold by the user to indicate whether the patient requires referral for further TB testing, although many companies provide 82 https://doi.org/10.1183/2312508X.10024322 ERS MONOGRAPH |THE CHALLENGE OF TB IN THE 21ST CENTURY