Recommendations Developed for Dermatology Artificial Intelligence
Checklist for evaluation of image-based AI reports grouped into topics of data, technique, technical assessment, application
By Dermsquared Editorial Team | December 01, 2021
In CLEAR Derm consensus guidelines issued by the International Skin Imaging Collaboration Artificial Intelligence Working Group and published online Dec. 1 in JAMA Dermatology , recommendations are presented for imaging-based artificial intelligence (AI) algorithm development and assessment in dermatology.
Roxana Daneshjou, M.D., Ph.D., from the Stanford School of Medicine in Redwood City, California, and colleagues guide developers and reviewers of dermatology AI by consolidating limited existing literature with expert opinion. Key recommendations were formulated and grouped into topics of data, technique, technical assessment, and application.
The authors developed a checklist of items that outlined the best practices of image-based AI development and assessment. The checklist relating to data included image types, image artifacts, technical acquisition details, preprocessing procedures, synthetic images made public, adequate referencing of public images, patient-level metadata, skin tone information, potential biases, data set partitions, sample sizes, external test data, multivendor images, class distribution and balance, and images that are out of distribution. The checklist for techniques included labeling method, references to common/accepted diagnostic labels, histopathologic review for malignant neoplasms, and detailed description of the development of algorithms. Technical assessment included publicly evaluating algorithms; performance measures; benchmarking, technical comparison, and novelty; and assessment of bias. Finally, the application checklist included the use of cases and target conditions and potential impacts on the health care team and patients.
"While we propose guidelines for clinical and peer-review evaluation of AI, these recommendations are also relevant for a regulatory framework and should be considered for any automated dermatology algorithm that may affect the wider community," the authors write.
Several authors disclosed financial ties to the biopharmaceutical, technology, and health care industries; one author holds several related patents.