КЛАСИФІКАЦІЯ ГІСТОЛОГІЧНИХ ЗОБРАЖЕНЬ НА ОСНОВІ ЗГОРТКОВИХ НЕЙРОННИХ МЕРЕЖ

Context. Automated classification of histological images is of great importance for speeding up and improving the accuracy ofdiagnostics in medicine. Taking into account the complexity and high variability of histological structures, the use of deep learning and convolutional neural networks, in particular, is a promising direction in solving this problem.The object of study is the process of classifying histological images using convolutional neural networks to determine and optimize the architecture with the highest accuracy rate.Objective. The aim of this work is to develop an effective approach to histological image classification using convolutional neural networks that provides high accuracy through step-by-step optimization of the architecture and application of data expansion methods using affine transformations and synthesis based on diffusion models.Method. The study includes four main stages. The first stage involves a comparative analysis of 12 well-known convolutional neural network architectures on a basic histological dataset. The second and third stages involve comparative analysis on extended data, including affine transformations and synthetic images generated by the diffusion model, respectively. The final, fourth stage involves a neuroevolutionary search for the optimal architectural cell. Once it is found, it is integrated into the model architecture, where for each layer a choice is made between certain blocks and the found cell. This approach allows to automatically form the optimal sequence of blocks in the model, which ensures the highest classification accuracy.Results. The proposed approach improved the accuracy of histological image classification compared to the initial architectures. The addition of synthetic images to the training set provided an increase in model performance. The search for the optimal cell and its integration into the model with further optimization demonstrated an additional improvement in classification quality, increasing the accuracy to 94.9, 96.1, and 99.8% on each dataset, respectively.Conclusions. The proposed approach allows achieving high accuracy of histological image classification through a step-by-step process that includes the use of classical convolutional neural network architectures, generation of synthetic data, and search for optimal architectural and hyperparametric configurations. A software module for classifying histological images has been developed that can be used in an automatic diagnostic system.