No worse than an ophthalmologist
Lily Peng and her colleagues from the Google Research department have published an article describing an automated system capable of scanning the retina of the eye and identifying diabetic retinopathy. To do this, retinal images were uploaded to the program, which were manually checked by ophthalmologists.
Diabetes mellitus is widespread all over the world. Every diabetic is at risk of developing diabetic retinopathy, a retinal lesion that can lead to terrible outcomes up to complete and irreversible loss of vision.
The insidiousness of diabetic retinopathy lies in the fact that at an early stage the patient himself does not feel any changes from the side of vision. As the disease progresses, it becomes more and more difficult to keep it "in check". Therefore, timely detection of retinopathy and constant monitoring of the condition of the retina plays a critical role. To this end, all diabetics are recommended to undergo a full ophthalmological examination annually, even in the absence of complaints.
In an earlier paper, Peng and colleagues described the use of neural networks– complex mathematical models for identifying patterns–to recognize diabetic retinopathy. Thousands of retinal images were uploaded to a special program to "teach" these neural networks to identify microbleeds and other initial signs of diabetic retinal damage. Then the authors wrote that their program works no worse than the average ophthalmologist.
Nevertheless, the classification of the degree of diabetic retinopathy is very difficult even for a practicing physician. And the patient's management tactics directly depend on the degree of the lesion, which may be limited only to dynamic observation or require laser or interventional surgery.
The purpose of the new work is to improve the diagnostics carried out by artificial intelligence from Google. To do this, the researchers involved a group of narrow specialists who have devoted years of practice to retinal diseases. They manually sorted through all the images uploaded to the program and corrected the descriptions of the images, making each decision collectively after discussion. Thus, it was possible to exclude the influence of artifacts such as dust, blurring of the image and others, as well as to clarify and identify the smallest changes that were missed earlier.
The corrected data were uploaded to the program, which significantly increased its diagnostic accuracy, which, according to experts who described the images, exceeds the average level of diagnostics in medical institutions.
This work provides a basis for further research, it is designed to help practitioners and their patients, and also significantly increases the level of standards in the field of automated medical systems.
Article by J. Krause et al. Grader Variability and the Importance of Reference Standards for Evaluating Machine Learning Models for Diabetic Retinopathy is published in the journal Ophthalmology.
Aminat Adzhieva, portal "Eternal Youth" http://vechnayamolodost.ru based on the materials of the American Academy of Ophthalmology: Google's AI Program: Building Better Algorithms for Detecting Eye Disease.