Advances in Melanoma-Nevus Classification Using Artificially Generated Image Data Sets

Abstract

The challenging task of skin cancer classification via artificial intelligence (AI) is a combined effort of dermatologists, patients, and the machine-learning community. 1,2 A classifier in this case is a software based on AI that assigns skin photographs to categories (ie, benign or malignant).Artificial intelligence is already part of care as an assistant system in Europe, commercially available to dermatologists for integration in their skin cancer screenings. However, a lack of interpretability and reliability has been criticized as putting the patients at risk. 3 The reliability of a given algorithm largely depends on the quality of the images it is trained with, since these images determine the rules for assigning a class (ie, nevus or melanoma) to a given test image.An important limitation of AI-assisted skin cancer classification remains the scarcity and quality of publicly available data sets both for training and testing image classifiers, especially for research units without direct connection to patient care. The heterogeneity of the data are also highly relevant with regard to both technical aspects (eg, capturing devices, resolutions, angles, preprocessing and postprocessing, contrasts) and biological characteristics (eg, skin type, ethnicity, sex, age). Furthermore, ensuring that dermatologic clinical images accurately represent the diagnosis ideally requires histopathologic verification or other molecular testing, which establishes the "ground truth" of a data set. Only data sets that fulfill most of these criteria may lead to robust classifiers that sustain artificial and clinical benchmarks in a given external environment. 4,5 Privacy concerns lead to a challenging process of collecting high-quality and heterogenous data from university hospitals or specialized medical centers. There are techniques in which algorithms rather than data are transferred for training, which may resolve these issues; however, to date, there is a lack of clinical data supporting the use of these techniques in the field of dermatology.In this issue of JAMA Dermatology, Cho et al 6 propose a novel, broadly accessible way to compile data sets that can address both data scarcity and possible privacy concerns in the development of AI for skin cancer detection. By leveraging the power of generative AI, a type of AI that can create a wide variety of data such as images, videos, audio, text, and 3-dimensional models, coupled with publicly available skin lesion photographs on the internet, they demonstrate that it is possible to acquire sufficient training data for a melanomanevus classifier that is performant in experimental settings.