A deep-learning algorithm has been proven to recognise cancerous cells in skin lesions as accurately as trained dermatologists. The algorithm was developed by researchers at the Stanford Artificial Intelligence Laboratory and has been shown to achieve results just as well and in some cases even better than 21 board-certified dermatologists at diagnosing skin cancer.
These promising results were reported in this month’s issue of Nature where lead researcher Dr. Sebastian Thrun outlined his hope that the deep-learning algorithm may help doctors to identify which skin lesions to biopsy.
The algorithm – called the deep convolutional neural network (CNN) was initially dreamt up by Andre Esteva who recognised how universal smartphones have become, and how their increasingly sophisticated technology can be utilised for medical purposes. “Everyone will have a supercomputer in their pockets with a number of sensors in it, including a camera,” said Andre. “What if we could use it to visually screen for skin cancer? Or other ailments?”
The technology adapted by the Stanford researchers had in fact initially been developed by Google to differentiate cats from dogs by identifying 1.28 million images from 1,000 object categories. The researchers, along with dermatologists at Stanford Medicine and professor of microbiology and immunology Helen M. Blau, then ingeniously modified the algorithm to differentiate between benign and malignant lesions, saving themselves the complexity of creating a deep-learning algorithm from scratch.
The algorithm was later assessed on almost 2,000 digitally interpreted images of skin with biopsy-proven diagnoses of cancer. The algorithm correctly identified the majority of the cancerous specimens with a rate of accuracy very similar to expert clinical dermatologists.
This is just one example of the many ways in which deep-learning is transforming the medical industry and is reducing the likelihood of serious conditions being overlooked or misdiagnosed. The algorithm in question is still some way off being implemented on actual patients but after further testing we can certainly expect to see this, and other similar technologies appearing in hospitals, assisting in the process of both diagnosis and treatment.