Abstract : Recent advancements in deep leaming techniques have significantly improved the diagnosis of brain tumors by enabling the development of powerful models. This thesis aims to contribute to the advancement of brain tumor diagnosis systems by exploring and improving various recent deep learning techniques, with a focus on two tasks: classification and segmentation. The first task comprises a sequence of three extensive studies, involving a thorough evaluation, analysis of results, and comparisons. In the first study, the performance of pre- trained and non-pretrained benchmark models such as VGG and ResNet is compared to custom-created CNN and Vision Transformer models. The second study applies the Model Soup algorithm to select highly performing models from the first study to improve accuracy. In the third study, a pre-trained CoAtNet model is employed for brain tumor classification and enhanced using techniques such as augmentations, large batch sizing, normalization, and exponentially decaying learning rate. Our model achieved the highest accuracy of 99.16%, surpassing state-of-the-art models, and our results were compared with previous literature results. In the second task, we evaluate the performance of multiple U-Net models with a minimal skip-connections configuration on the segmentation of glioblastoma, a specific type of malignant brain tumor. The results showed that climinating the first path results in the optimal U-Net skip-connection configuration. Overall, this thesis employs a rigorous approach to explore the potential of deep learning techniques for brain tumor diagnosis, and the results have significant implications for improving the accuracy of brain tumor diagnosis.