Interpretable Deep Learning for MRI-Based Content-Based Image Retrieval in Alzheimer’s Disease

Alzheimer’s disease is a neurodegenerative disorder in which structural brain alterations, observable through magnetic resonance imaging (MRI), play a key role in both diagnosis and disease monitoring. In recent years, deep learning approaches have shown great potential for automatically extracting meaningful patterns from MRI data. However, many of these models rely on heavily preprocessed images and behave as black boxes, limiting their clinical applicability and interpretability. In this context, this thesis develops and evaluates a deep learning framework capable of reproducing NeuroCBIR target embeddings directly from minimally preprocessed T1-weighted MRI scans. A conventional 3D CNN was used as a baseline and compared with a prototype- based PIP-Net model adapted for embedding regression. Both models were evaluated through subject-level image retrieval on OASIS-3 and on three external datasets. Across the four validation folds, PIP-Net achieved substantially higher and more stable retrieval performance than the CNN baseline, with an mS@1 of 0.865 ± 0.007, compared with 0.613 ± 0.157 for the CNN. On the external datasets, PIP-Net achieved mS@1 values of 0.994 on MIRIAD, 0.827 on AIBL, and 0.604 on SLIM, compared with 0.995, 0.696, and 0.381 for the CNN baseline, indicating strong generalization to datasets with characteristics similar to the training data, with reduced performance under larger domain shifts. Overall, the findings suggest that NeuroCBIR-like embeddings can be approximated from minimally preprocessed MRI scans, while prototype activations provide an additional way to inspect the image regions contributing to the predicted representation.