Artificial Intelligence in Skin Cancer: Man vs Machine Redux
The use of AI in the diagnosis of skin cancer has real potential, yet current algorithms lack standardization and regulatory approval.
The incidence of skin cancer is increasing in fair-skinned individuals, in the United States and worldwide, and is putting pressure on health care costs and specialist services. Estimates predict that as high as 20% of Americans will get skin cancer, with approximately 230,000 new cases of malignant melanoma worldwide and 50,000 resulting deaths.1,2 Although melanoma represents 5% of all skin cancers — it is the most deadly, with a high risk of metastasis — and is thought to be the cause of 75% of skin cancer deaths.
However, the most common forms of skin cancer are nonmelanocytic or nonpigmented, such as basal cell carcinoma and squamous cell carcinoma (80% and 20% of all skin cancers, respectively), and is more difficult to diagnose. The incidence of keratinocyte carcinomas (cutaneous squamous cell and basal cell carcinomas is 5.4 million cases.3
The diagnosis of skin cancer is primarily conducted through dermoscopic analysis followed by a biopsy and pathological examination. Early diagnosis of skin cancer is a cornerstone to improving outcomes and is correlated with 99% overall survival (OS). But once skin cancer is in advanced stages, OS falls to 5%.
In the last decade, the success of machine-learning systems has reshaped the role of dermatologists. Algorithms integrating disease taxonomy were developed to automate classification of skin cancer images.5 Convolutional neural networks (CNNs) are composed of several layers of algorithms or neurons performing convolution, a process of constructing meaningful representations of objects in an image from pixel data, with subsequent layers processing additional information to perfect the resulting image (using features such as edge, borders, and irregularity of a skin lesion).6
In their landmark article in Nature, Andre Esteva, the department of electrical engineering, Stanford University, Stanford, California, and colleagues, used pretrained CNN machines, validated with over 130,000 dermoscopic images and their corresponding 2000 clinical descriptions, proving a diagnostic ability comparable to that of 21 board-certified dermatologists in classifying melanoma, melanoma with dermoscopy, and keratinocyte carcinomas.1
CNN can be delivered through mobile networks, allowing patients remote access to dermatologist-level diagnoses. This technology has the potential to increase the reach of primary care beyond the clinic and expand the scope of clinical decision making.