195 GENERAL DISCUSSION photographs taken by GPs and patients beforehand and to supply real-time feedback and guidance to them if the taken images are of insufficient quality. It is relevant to provide additional training for GPs in the use of photography equipment and how to obtain adequate (dermoscopic) photographs but GPs have to keep track of knowledge in several other medical disciplines, have limited time, and perform several non-medical tasks. AI solutions could be suitable to bridge this tension field. In fact, Dutch GPs perceived that AI applications have an educational purpose because these applications can explain why a lesion is classified as suspicious or benign, thereby possibly enlarging GPs’ dermatology knowledge [56]. Dermatologists in this study also acknowledged the educational potential of AI for GPs, referring to AI as an instrument to train GPs and allowing them to spend less time on training their photography skills. Therefore, future studies could explore whether these kinds of AI solutions may eventually reduce the training requirements for GPs performing digital dermatology consultation. Studies in the literature showed that image-based AI may improve GPs’ diagnostic accuracy of skin disorders and results in fewer missed skin cancer diagnoses and fewer unnecessary excisions or referrals for patients with benign skin lesions [53,56,57]. Furthermore, GPs indicated that the improved diagnostic accuracy with AI subsequently boosted their confidence in the management of patients with suspected skin lesions [56]. Future research should examine whether AI photo recognition algorithms in the digital dermatology services can indeed be used to interpret clinical or dermoscopic skin images and to assist GPs in triaging or diagnosing various skin conditions. However, methodological limitations hamper the diagnostic accuracy of the latest available AI image recognition applications that aim to determine the risk of skin cancer in suspicious skin lesions [58]. For example, not all previous developed AI algorithms in dermatology are trained on clinical images taken with equipment in real practice or are prepared to manage lower-quality images [58-60]. Another methodological limitation is that not all AI models are trained on difficult to diagnose skin conditions or on an inclusive training set (for example, only lesions of white-skinned patients or chronic skin conditions) [58,59,61]. Ibrahim et al. consider this problem as “health data poverty: the inability for individuals, groups, or populations to benefit from a discovery or innovation due to a scarcity of data that are adequately representative” [62]. To avoid this health data poverty, AI algorithms should be validated on different external image datasets so that the results of the algorithm are generalizable to other populations independent of the images used for training the algorithm. As an example, researchers could combine the training dataset with clinical and dermoscopic images from shared international archived image galleries [63,64] or create a multicenter training imaging set based on multiple dermatology datasets from different countries [65]. Before these diagnostic AI algorithms are used in real-life conditions in general dermatology practice, appropriate evaluations are required to ensure that these AI algorithms are fair, reliable, and safe. Therefore, the International Skin Imaging Collaboration AI working group formulated recommendations 8
RkJQdWJsaXNoZXIy MTk4NDMw