Artificial Intelligence in Dermatology: A Primer

Abstract

Artificial intelligence is becoming increasingly important in dermatology, with studies reporting accuracy matching or exceeding dermatologists for the diagnosis of skin lesions from clinical and dermoscopic images. However, real-world clinical validation is currently lacking. We review dermatological applications of deep learning, the leading artificial intelligence technology for image analysis, and discuss its current capabilities, potential failure modes, and challenges surrounding performance assessment and interpretability. We address the following three primary applications: (i) teledermatology, including triage for referral to dermatologists; (ii) augmenting clinical assessment during face-to-face visits; and (iii) dermatopathology. We discuss equity and ethical issues related to future clinical adoption and recommend specific standardization of metrics for reporting model performance.

Stress testing reveals gaps in clinic readiness of image-based diagnostic artificial intelligence models

“…We evaluate our models in a clinically meaningful way by comparing them to dermatologists’ management decisions. It is currently difficult to compare CNN models across studies due to proprietary models and, heterogeneous and proprietary test datasets; additional standardized benchmarks are needed to compare model performance 10 .…”

“…Users should be aware that a model that has not been developed specifically to handle the out-of-distribution problem will do its best to blindly predict according to the disease classes it was trained on when it encounters a new disease it was not trained on, with high-confidence predictions potentially leading to false reassurance. An ideal model would first screen images to assess adequacy for decision-making (e.g., based on focus, lighting, presence of artifacts, etc., or similarity to images seen during training) and direct users to retake an image or defer to human experts when appropriate 10 .…”

“…Such incorrect predictions may arise from images that are minimally altered 9 , raising concern for model robustness to images taken in variable conditions. Several other critical concerns for real-world use, such as discrimination and calibration, remain unaddressed 10 , and proof of practice for published models has not been demonstrated.…”

The role of technology in melanoma screening and diagnosis

“…Recent advances in deep learning, an artificial intelligence method that excels in visual analysis, and the collection of large datasets have led to impressive gains in the automated diagnosis of melanoma (Young, Xiong, Pfau, Keiser, & Wei, 2020). In a landmark study, Esteva et al compiled 129,450 digital and dermoscopic images representing 2,032 different skin diseases to train a deep learning convolutional neural network that achieved an area under the receiver operating characteristic curve (AUC) of 0.96 for melanoma diagnosis, with an overall accuracy on par with dermatologists (Esteva et al., 2017).…”

Comparing the self‐perceived effects of a facial anti‐aging product to those automatically detected from selfie images of Chinese women of different ages and cities

“…Such objectives lead many skin care cosmeticians to investigate on the possible input of connected or nomadic devices such as telemedicine, different connected devices or analysis of selfie images taken by sophisticated smartphones. [1][2][3][4][5][6][7][8][9] All these approaches were obviously made possible by the tremendous progresses in imaging techniques coupled to continually more powerful hardware.…”