pre-set thresholds with their products [41]. Thresholds can be altered in different settings depending on the use case, patient population characteristics and local service capacity for follow-on diagnostic testing [41]. CAD products appear to have at least equivalent accuracy to human readers for the screening of TB. Two systematic reviews from 2019 and 2022 found generally high accuracy for a range of CAD products, although meta-analysed estimates could not be produced due to heterogeneous study designs, methodological limitations, and concerns about the use of overlapping datasets of CXR images for training and testing [42, 43]. In three evaluations conducted on behalf of the WHO to inform their updated 2021 guideline, human radiologists and CAD products were compared directly using libraries of fresh images on which the CAD had not been trained. Both humans and CAD showed variable accuracy across settings and populations, but their sensitivity and specificity estimates overlapped. The WHO therefore concluded that there is little difference between the two, and recommended any CAD products that demonstrate noninferior performance on external validation to those evaluated in their guideline [9]. Subsequent analysis of five CAD products using a large fresh dataset found that all five products significantly outperformed human radiologists in three TB screening centres in Bangladesh [40]. Unlike traditional diagnostic tests, CAD products can be updated rapidly. A 2022 study comparing two of the CAD product versions evaluated by the WHO with their subsequent version updates on fresh image sets found that the newer versions significantly outperformed the older ones [44]. Although this presents obvious advantages as manufacturers can rapidly roll out updated and improved versions to their users globally, it also presents a challenge for programmes using the technology: a given threshold score will not always be associated with the same diagnostic accuracy in different versions of the same product, meaning that software updates may require programmes to adjust threshold scores to maintain the same performance and keep referrals for follow-on testing within local capacity [44]. In response to this challenge, FIGURE 2 Computer-aided detection of CXR abnormalities in a TB clinic. https://doi.org/10.1183/2312508X.10024322 83 IMAGING AND DIAGNOSIS |J. BIGIO ET AL.