Перегляд за Автор "Yaroshenko, M. O."

Зараз показуємо 1 - 1 з 1
  • Ескіз
    ДокументВідкритий доступ
    Human Eye Aberrometry Data Generation Using Generative Adversarial Neural Network
    (КПІ ім. Ігоря Сікорського, 2023) Yaroshenko, M. O.
    It’s obvious that for development and improvement of methods and apparatus for diagnosis and treatment of optical flaws of human eye at the modelling stage, it’s necessary to have sets of real measurements. However, data requests to clinics are accompanied by substantial amount of bureaucracy procedures and, at the same time, acquired dataset may be too small, which can be critical, for example, for training of neural networks. According to the analysis of existing publications, publicly available datasets of aberrometry data (sets of eye’s refractive flaws) are rare and consist of relatively low number of measurements. But, due to current development state of neural networks, it is possible to generate data based on real measurements. The most common solutions are methods based on the usage of the Generative Adversarial Networks (GAN). This tendency is also relevant for the modern ophthalmology, but no publications aimed at aberrometry data synthesis were found. For this reason, objective of this work is development of solution for generation of sets of human eye’s refractive errors using neural networks. Proposed solution includes generator and critic networks trained according to the Wasserstein GAN with Gradient Penalty (WGAN GP) algorithm. In order to improve training, the method of data augmentation called Data Augmentation Optimized for GAN (DAG) was used, moreover, the possibility of augmentation of aberrometry data in two forms was implemented — for both Zernike coefficient vectors and wavefront pixel images. According to the result’s evaluation, generated data has the distribution close to the real sample (Fréchet distance equals 0.7) and, at the same time, it is neither a copy of real measurements (92% creativity rate) nor duplication of a few aberration sets (diversity metric equals 3.64 which is close to the optimal 3.83). The direction of further improvement includes enhancement of existing architectures of generator and critic, search or creation of bigger training dataset and refinement of data augmentation technics.