Development and Testing of an Artificial Neural Network Model to Analyse Data, Image Data

The relevance of the study is related to the increased needs of medical personnel in qualitative solution of image data analysis tasks, in particular lung scans during the period of controlling the consequences of coronavirus infection. An artificial neural network model was developed and tested to improve the quality of research results, reduce time, and optimise the work of medical staff. A convolutional neural network was chosen as a tool to solve this problem. The tools and software products most commonly used to build and design artificial neural networks were analysed. Python programming language and PyQt5 framework were used as tools for conducting software product development and testing. Three architectures were involved in the experiment, and the number of layers, epochs, average epoch training time and network accuracy were taken into account to evaluate their efficiency. It was found that the optimal architecture is the one consisting of 9 layers, 10 epochs, giving an accuracy of 72.39%. The effectiveness of the regularisation methods for modifying the convolutional neural network on the publicly available CIFAR-10 dataset was evaluated. Among the six modification methods considered, the modification using a combination of Weight Initialisation and Batch Normalisation methods was preferred. Next, Data Augmentation method was applied to the base models and modifications to reduce overfitting. However, using Data Augmentation in conjunction with other regularisers resulted in worse results. The interface of convolutional neural network was developed. The result of the study was a convolutional neural network with 90% accuracy.