FIND (the global alliance for diagnostics) has recently developed a validation platform that performs independent analysis of AI-based diagnostic products on diverse datasets. The platform consists of a third-party digital infrastructure for fast and standardised in silico assessment of CAD products (in addition to other digital diagnostics) on high-quality clinical, laboratory and digital imaging data collected by FIND in collaboration with partners in multiple countries. These assessments are designed to highlight where performance in specific populations may differ from those on which the products were trained [45]. The potential utility of these CAD products is large as a replacement for human readers in locations that lack such health workers and as an aid to human readers to reduce their workload and highlight abnormal images requiring prioritisation [35, 41], as well as being part of integrated disease screening efforts [46], which could help advance the universal health coverage agenda. However, they have some limitations. First, as neural networks process information in a manner beyond the understanding of users, and the algorithms underlying them are business secrets, a CAD product can be seen as a black box, reaching a decision through a process that is effectively impossible for a user to explain [35, 44]. This opacity means that each software update should be treated as an entirely new product for evaluation as it may have involved significant changes to the neural network that cannot be understood from the outside [36]. Unbiased software evaluations of the latest versions created using datasets from different parts of the world will therefore be vital for both maintaining trust in CAD products and for allowing users to understand the programmatic implications of the updated products [44]. The WHO is currently creating a set of requirements for CAD for TB as part of a prequalification technical specification that will include guidance on the validation method for CAD products [47], and the FIND validation platform mentioned earlier will support this work [45]. Second, CAD products have infrastructure, training and other requirements that may prove challenging. For example, an early evaluation of a CAD product for TB screening in five primary health centres in Peru found that limited internet connectivity, lack of access to radiograph viewers and heavy health worker workloads discouraged use of the product [48]. In addition, these products are currently only recommended by the WHO for the screening of TB, and not as products suitable for the general interpretation of CXR for any indication, a clear disadvantage when compared with human readers. General CXR interpretation is much more technically challenging than the diagnosis of an individual disease, and although CAD has seen expansion into non-TB use cases such as cancer and COVID-19, it does not yet cover all possible use cases [49–51]. Lastly, current CAD products are not validated for use in children, and this is a key evidence gap that must be addressed. Ultrasound Like CXR, ultrasound hardware has developed rapidly in recent years. Modern POCUS devices are inexpensive, safe and highly portable [52]. A 2023 study from the UK found that POCUS devices ranged in price from £2 500 to £6 000, compared with ∼£30 000 for a traditional hospital ultrasound device [53]. They can run on battery power and are increasingly used by nonradiologist healthcare staff for a range of indications in hospital and primary-care settings worldwide (figure 3) [54–57]. Given the high hardware cost of CXR and access challenges to CXR in many high-TB-burden settings, utilising POCUS for the screening of TB could substantially widen access to imaging for TB. However, evidence is currently limited and of low quality for the use of POCUS as a diagnostic aid for both PTB and EPTB. AI interpretation of ultrasound images has been attempted for many anatomical regions of the body [58, 59], but has yet to overcome critical issues for use in TB screening and diagnosis, in part due to a lack of appropriate databases of ultrasound images on which to train the AI [60]. 84 https://doi.org/10.1183/2312508X.10024322 ERS MONOGRAPH |THE CHALLENGE OF TB IN THE 21ST CENTURY