MACHINE LEARNING AND ARTIFICIAL NEURAL NETWORK TECHNOLOGIES IN THE CLASSIFICATION OF POSTKERATOTOMIC CORNEAL DEFORMITY

Abstract

Backgraund: A thorough analysis of both the optical and anatomical properties of the cornea in patients after anterior radial keratotomy is of particular importance in choosing the optical strength of an intraocular lens in the surgical treatment of cataracts and other types of optical correction. The variability of the clinical picture of postkeratotomic deformity (PCRD) determines the need to develop its classification and is an important task of modern ophthalmology. Aims: to develop an automated system of classification of corneal PCRD using machine learning and an artificial neural network based on the analysis of topographic maps of the cornea. Materials and methods: depersonalized results of the analysis of medical records of 250 patients aged 59.63±5.95 (from 46 to 76) years were used as the material. The analysis of 500 maps of the relief-topography of the anterior and posterior surfaces of the cornea and 3 stages of machine learning of the PCRD classification were carried out. Results: Stage 1- analysis of the relief topography of the anterior and posterior surfaces of the cornea allowed us to fix the numerical values of the elevation of the anterior and posterior surfaces of the cornea in three ring-shaped zones. At stage 2, in the course of deep machine learning, a direct distribution neural network was selected and created. 8 auxiliary parameters describing the shape of the anterior and posterior surfaces of the cornea were established. Stage 3 was accompanied by obtaining algorithms for the classification of PCRD depending on the ratio of test and training samples, which ranged from 75 to 91%.. Conclusion: The use of artificial neural network algorithms can become a useful tool for automatic classification of postkeratotomic corneal deformity in patients who have previously undergone radial keratotomy.