‘Man With Machine’ Approach to Diagnosing Skin Cancer Lesions
Posted: Monday, September 23, 2019
In the field of skin cancer detection, the advancement of artificial intelligence in diagnosing skin lesions has raised a compelling question: Is man or machine the better diagnostician? According to a study published in the European Journal of Cancer, a combination of human and computerized diagnostics may be more effective than either method alone.
“To the best of our knowledge, this is the first study in the field of digital skin diagnostics that has combined the decisions of dermatologists and artificial intelligence,” stated Titus J. Brinker, MD, of the German Cancer Research Center and University Hospital Heidelberg in Heidelberg, Germany, and colleagues.
After the researchers divided 11,444 dermoscopic images into 5 diagnostic categories, novel deep learning techniques were used to train a single convolutional neural network. Then, 112 dermatologists from 13 German university hospitals and the trained convolutional neural network independently classified a set of 300 biopsy-verified skin lesions into those 5 classes. The decisions of each system were merged to form an overall classifier, taking the degree of certainty of both systems into account.
The combination of both systems did significantly increase the sensitivity from 86% with the convolutional neural network alone to 89% by incorporating the human decision (P < .05). However, although the combination of man and machine increased overall accuracy to 82.95% compared with 81.59% achieved by the convolutional neural network and 42.94% by the physicians, it was not statistically significant. “Although our results are not significant, they indicate superiority of the ‘man with machine’ approach on 300 test images obtained from a heterogenous data set with high external validity,” concluded Dr. Brinker and colleagues.
Disclosure: The study authors reported no conflicts of interest.