Mole or cancer?

The deep learning algorithm diagnoses skin cancer no worse than a qualified dermatologist

marks, Geektimes

Deep learning is a promising method of teaching algorithms, which is involved in a large number of areas (information security, analysis of research results, image recognition). As for image recognition, here we are talking not only about the fact that a machine can distinguish a cat from a dog, as it was with the Google neural network. No, such technology can be useful in medicine, in particular, in oncology.

Stanford scientists have created a system that is able to make a diagnosis by analyzing a photo of a patient's skin (Taylor Kubota, Stanford News Service: Deep learning algorithm does as well as dermatologists in identifying skin cancer). Recent tests have shown impressive results: the algorithm made diagnoses as accurately as dermatologists with extensive experience and serious qualifications. To compare the capabilities of the technology, the authors of the project asked professional dermatologists to make a diagnosis based on the image of skin areas of various people (with verification of the diagnosis), and then the same images were shown to the machine. 

"We have created a very powerful deep learning algorithm that is able to learn using data," said Andre Esteva, one of the authors of the study. "Instead of hard-coding such a system, we let it make its own decisions." 

The algorithm was named "deep convolution neural network". Its capabilities are based on Google Brain, a project of Google Corporation, whose goal is to explore the possibilities of machine learning. The computing power of the Google Brain system enables third-party developers to create various machine learning projects. When scientists started working, the neural network could identify more than 1.28 million objects in the images, divided into several thousand different categories. But the researchers had a clear goal – they needed to train the neural network to correctly identify carcinoma and seborrheic keratosis, as well as teach the system to distinguish these two diseases from each other by images with areas of affected human skin. 

In addition, the computer needed to distinguish these elements from the usual age spots, rashes and other possible changes in the structure of the skin. A doctor with a lot of knowledge and experience is able to do this almost without mistakes. And scientists have set themselves the task of "educating" such a professional from the neural network. 

At the top – pictures of benign skin changes, at the bottom – malignant formations (VM)

The problem was also the fact that the specialists did not have a large enough sample of images, according to which it would be possible to train the system. Therefore, they had to create an image database on their own. "We collected photos from the Internet and asked doctors to help us sort the images," says one of the authors of the study. The authors took some pictures from foreign websites, so it was sometimes simply impossible to understand what was written in the description, since the accompanying texts were in Arabic, German, Latin and other languages. 

In order to study the condition of a patient's skin area, dermatologists often use a medical instrument called a dermatoscope. It gives a certain level of magnification so that the doctor can see the skin in detail. The device provides approximately the same "picture", so that a photo of a skin area taken with this tool is understandable for any dermatologist from any country in the world. Unfortunately for the study participants, not all photos from the Internet were taken using a dermatoscope. The shooting angle, the illumination, the degree of magnification – all this was different. 

As a result, scientists analyzed 130,000 images and identified about 2,000 different types of skin diseases. They created a dataset for the image library, and then "fed" it all to the neural network. Each image was represented by a separate block, a "pixel", with a brief description of the disease. Then the algorithm was "asked" to show the stages of development of the same disease, having previously identified the patterns of enlargement of the focus.

Various categories of images into which the algorithm divided the original database of photos

After everything was ready, the authors of the project compared the results of the diagnosis made with the system with the known results of the diagnosis of skin diseases of patients made by two dozen dermatologists from Stanford Medical School. To test the algorithm, scientists used only high-quality images made by professionals. The diagnostic accuracy was 91%, both for the algorithm and for doctors. 

The authors plan to gradually develop their development. In particular, the researchers want to create an application that will work directly with photos of skin areas with problem areas that are uploaded by patients themselves. This, according to the researchers, will simplify access to medical services for a large number of patients. And smartphones can be of invaluable help here. "My main Eureka moment was when I realized how ubiquitous smartphones would be," says one of the initiators of the project, describing the process of implementing the work from the idea to the working service. "Anyone now has a powerful computer with a large number of sensors, including a camera. What if you can use this to get photos of skin cancer or other types of diseases?". 

In any case, researchers need to conduct more tests before bringing their technology to the masses in order to finally configure the algorithm. In this case, it is extremely important to know how the machine classifies images. 

"The possibilities of computer classification of images are a great help for dermatologists who will be able to make more accurate diagnoses. But in the future, it is necessary to confirm the algorithm's operability, this must be done before implementing such a practice in hospitals," says Susan Svetter, professor of dermatology at Stanford.

Article by Esteva et al. Dermatologist-level classification of skin cancer with deep neural networks is published in the journal Nature – VM.

Portal "Eternal youth" http://vechnayamolodost.ru  27.01.2017