How to build a content-based retrieval system for 3D brain MR images that encodes each 2D slice independently rather than the whole volume, and then repairs the three things slicing normally breaks. A variational autoencoder supplies a low-dimensional, continuous latent space; in place of an added classifier head, one learned prototype vector per disease class is scored against each slice embedding by cosine similarity (the Baseline++ construction), so the encoder itself is pushed to separate the classes while the reconstruction and Kullback-Leibler terms keep the space complete enough to retrieve on. Because most slices of a diseased brain look healthy, retrieval is done on overlapping blocks of consecutive slices matched block-for-block against the class prototypes, and a scan is called abnormal only when enough blocks agree. Every slice carries its own class probabilities, so aggregating them across the axial, coronal and sagittal cross-sections yields a voxel-level disease probability map that says where the evidence sits. Students learn why a 2D encoder is the data-efficient choice when labelled volumes are scarce, how a prototype-and-similarity objective makes an embedding directly searchable instead of merely classifiable, and how per-slice probabilities become anatomical evidence. Demonstrated (iCBIR-Sli) on ADNI, OASIS and AIBL for Alzheimer disease versus cognitively normal at macro F1 near 0.86, with the maps localising to the hippocampus.
Content-based image retrieval means finding past cases that resemble a new one by comparing the images themselves rather than by matching text labels a radiologist typed. Doing this for brain MRI req…