ON TRANSFER LEARNING METHODS IN BIOMEDICAL IMAGES CLASSIFICATION TASKS [О МЕТОДАХ ПЕРЕНОСА ГЛУБОКОГО ОБУЧЕНИЯ В ЗАДАЧАХ КЛАССИФИКАЦИИ БИОМЕДИЦИНСКИХ ИЗОБРАЖЕНИЙ]

Computer studies of the effectiveness of deep transfer learning methods for solving the problem of human brain tumors recognition based on magtetic resonance imaging (MRI) are carried out. Various strategies of transfer learning and fine-tuning of the models are proposed and implemented. Deep convolutional networks VGG-16, ResNet-50, Xception, and MobileNetV2 were used as the baseline models, pre-trained on ImageNet. Also, a deep convolutional neural network 2D CNN was built and trained from scratch. Computer analysis of their performance metrics showed that when using the strategy of fine-tuning models on augmented MRI-scans data set, Xception model demonstrated higher accuracy values compared to other deep learning models. For Xception model, the accuracy of classification of MRI-scans with brain tumors was 96%, precision 99.43%, recall 96.03%, f1-score 97.7%, and AUC 98.92%. © 2021 Federal Research Center "Computer Science and Control" of Russian Academy of Sciences. All rights reserved.