Skin cancer better diagnosed by deep learning than doctors
A convolutional neural network (CNN) has beaten a team of 11 pathologists at diagnosing melanoma.
Details are presented in a study running in the September edition of the European Journal of Cancer.
For the study, dermatologist and cancer researcher Titus Brinker, MD, and colleagues at several universities in Germany trained a deep-learning system on 595 images.
The images were taken from pathology slides that were randomly cropped and previously classified by an expert histopathologist.
The expert simultaneously classified another 100 slide images to use as a test set.
Using three statistical tests to gauge the neural network’s performance vs. that of 11 histopathology physicians, the researchers found their CNN had better sensitivity and accuracy than the physicians (76% and 68% vs. 51.8% and 59.2%, respectively).
The physicians only topped the CNN at specificity, which rules out cancer, 66.5% to 60%.
Allowing that the work is preliminary and needs to be confirmed by prospective studies using whole slides rather than cropped images, Brinker et al. state that theirs is the first study to directly pit a deep-learning algorithm against practicing pathologists in diagnosing melanoma.
In their discussion, the authors suggest digital pathology is ripe for AI augmentation of pathologists’ eyes.
The clear win by their CNN over the physicians “may be explained by the ability of artificial intelligence to mine ‘sub-visual’ image features that may not be visually discernible by a pathologist,” they write. “Consequently, computer vision is able to gather more information with diagnostic relevance from an image section than a pathologist.”
The study is available in full for free.
Dave P. has worked in journalism, marketing and public relations for more than 30 years, frequently concentrating on hospitals, healthcare technology and Catholic communications. He has also specialized in fundraising communications, ghostwriting for CEOs of local, national and global charities, nonprofits and foundations.