Development of new regularization approaches for deep learning and application to medical images

Improving the prediction performance of artificial neural network (ANN)-based approaches is an important topic discussed in the literature. Regularization is one of the methods used to enhance prediction performance. In this thesis, three different innovative regularization methods are proposed for ANN-based models. These methods, applied in various deep networks, have improved prediction performance in the classification of medical image data. In the first application, regularization was performed with data. It was discovered that the capsule network (Ca) failed to achieve performance in statistical classification metrics when trained using classical data augmentation. Training data were generated in chest x-ray images using the proposed generative adversarial network (GAN) architecture. The classification performance in changes in the number of filters, size, and the number of training data parameters in the Ca training was explained in detail. The proposed method achieved an accuracy performance of 94%. Additionally, the GAN-based data generation approach was adapted for the classification of skin lesion images. Classification was performed on a seven-class skin lesion dataset using the proposed data augmentation model with transfer learning. This method achieved approximately a 10% increase in performance in statistical metrics.In the second application, a new loss augmentation function was proposed to increase the prediction performance of ANN and CNN models. With this method, a generalized mathematical model was presented to learn network weights as unit vectors. The proposed method increased the generalization performance by forcing the weight matrix of the models to learn as binary vectors. In experimental studies, the proposed method was applied to the training of seven different datasets. The proposed method increased the classification performance of ANN and CNN by 1% to 5%. In the final application, a network-based regularization approach was presented. A new network was proposed for the resolver network layer of the contrastive learning architecture, simclr method. Additionally, the model was reconstructed with a supervised learning strategy. The method applied to colon cancer histopathological datasets achieved high classification performance compared to other methods in the literature. Keywords: Regularization, Orthogonal Regularization, Medical Image Classification, Deep Learning, Contrastive Learning