Deep learning achieves encouraging performance in natural image classification. It has huge potential for detecting neural degenerative anomalies, but this is often limited by the availability of well-segmented neuroimages and computational resources. One way to address this problem is to apply self-supervised learning methods that utilise unsegmented neuroimages and artificial labels. Another solution is developing surrogate tasks to learn the representations of neuroimages for classification. This thesis reports the investigations of different variants of self-supervised learning and pretext tasks to train feature extractors for downstream Alzheimer’s Disease classification. Firstly, this thesis reviews the literature regarding Alzheimer’s Disease classification and possible data leakage issues. Then, a lightweight 3D CNN-based ensemble is trained to predict brain age using the 3D MRI data of cognitively normal subjects from the OASIS-3 dataset. The extracted features are evaluated in the binary classification of CN vs. AD patients from their brain MRI scans. This approach achieved competitive performance compared with state-of-the-art methods in the literature. The next part of this thesis developed four different self-supervised learning pretext tasks for feature extractor training: brain age prediction, brain sMRI reconstruction, brain sMRI rotation classification, as well as a combination of all three approaches into one single multi-task predictor. To further explore the feasibility of employing synthetic neuroimaging data in the self-supervised learning setting, the proposed approaches are trained on the LDM100K dataset followed by evaluation using real-world OASIS and ADNI datasets. The real-world data training and testing leads to the best classification performance. The random cropping data augmentation technique can improve feature extractor training on 3D MRI data. Due to high computational expense and time limitations, the results of the training using synthetic data are not as satisfactory as those using real-world data. Future research is needed to develop more advanced feature extractor architectures and more complex pretext tasks that can learn more discriminative features. Another area of research to improve training efficiency would involve developing specialised software and hardware for processing 3D neuroimaging data.