This is a good point and was an issue for skin cancer AI in the early days but has largely been solved now - newer models use ensembles of vision transformers and much greater number of training images and have been proven accurate in clinical trials. See for example https://skin-analytics.com/ (I am not linked in any way)