Computers help oncologists

Machine learning algorithms help to recognize cancer cells

Medical news based on Google Blog materials: Assisting Pathologists in Detecting Cancer with Deep Learning

The verdict of the pathologist who examined the patient's tissue samples is extremely important in the treatment of many diseases. In particular, if we are talking about oncological diseases, the entire treatment regimen is based on the pathologist's diagnosis. Doctors have been training for years, honing their skills and gaining experience.

Even with constant training, diagnoses made to the same patient by different pathologists can vary significantly, which can eventually lead to incorrect therapy. When studying images of patients with some forms of breast and prostate cancer, opinions may coincide by only 48%. This is not surprising – the amount of information contained in the images is huge. Pathologists need to consider all the tissues in the picture without exception, and there may be several such pictures in one patient. If you digitize images at 40x magnification, they will "weigh" 10 gigapixels, and the doctor is responsible for each pixel. The time may be limited.

In order to resolve the difficulties with the difference of opinion and time constraints, Google specialists are studying how deep learning algorithms can be applied in the work of a pathologist, to supplement the "manual" study with an automatic recognition program for affected cells. To prepare the program, the researchers used images provided by the Medical Center of the University of Nijmegen (Radboud University). Programmers have trained artificial intelligence to find breast cancer cells with lymph node metastases in the images.

It was found that even standard algorithms – for example, Inception (aka GoogLeNet) – are able to effectively distinguish diseased cells from healthy ones, although the resulting heat maps (images showing which part of the tissue is probably affected) contained too much noise. However, after refinement, which also included training networks to scan images at different magnifications, the accuracy of the mathematical model was comparable or even exceeded the accuracy of the work of a pathologist, who is not limited in time when studying biomaterial.

In fact, after the modification of the algorithm, the heat maps prepared with its help were improved so much that their accuracy reached 89%. Experts compared their result with the work of a pathologist, whose time was not limited (in fact, he finished after 30 hours, having examined 130 images), and found that artificial intelligence is 16% more effective than a human (73%). The model performed well with another set of images provided by another hospital.

Despite the promising results, the authors of the algorithm emphasized that the model is inferior to a person in any case, if only because it is programmed to look for only certain pathologies, and a person will pay attention to signs of other diseases – autoimmune diseases, inflammatory processes, other types of cancer. The counting system itself also has its drawbacks – the number of false positive results when a pathologist mistakes healthy cells for sick ones increases the sensitivity of the algorithm during training. The best option, according to experts, is to combine both approaches, that is, to supplement the work of a human specialist with a deep learning algorithm program. 

Portal "Eternal youth" http://vechnayamolodost.ru  06.03.2017