NADERAN M. Hybrid convolutional neural network for image processing and medical diagnostics

The aim of the dissertation research is to develop a new architecture of a hybrid convolutional network and a classification model to improve the quality of breast cancer recognition and reduce training time. The research is driven by the need to develop new and improve existing models and methods for image processing and medical diagnostics. The hybrid convolutional network should provide the selection of informative features that improve the quality criterion of the model for the tasks of diagnosing breast cancer. In the proposed hybrid convolutional model, the convolutional autoencoder was used to search for informative features, and the convolutional neural network DenseNet - for classification. Experimental studies of the developed model of breast cancer recognition were performed, the following indicators were obtained: sensitivity, accuracy (precision), F1-Score and accuracy (accuracy) models which are 93.5%, 93.2%, 93.3% and 93%, respectively, which is much higher than in known convolutional networks that have been used for this task. The scientific novelty of the dissertation is: - A model is proposed that, unlike existing models, allows the diagnosis of breast cancer in a minimum time compared to known methods. - Developed a hybrid convolutional network based on the encoder, which improves the quality of classification of breast cancer and in particular to achieve a minimum percentage of false negative error (FN) in comparison with known works in the classification of breast cancer. - Modified the architecture of the Inception V3 model by expanding the number of fully connected layers.