Self-Supervised Dementia Prediction From MRI Scans With Metadata Integration

Abstract

We introduce metadata integration in the training process for dementia diagnoses as weak label information using Weakly-Supervised Modified Knowledge Distillation with No Labels (WS-MDINO). Using WSMDINO, we fine-tuned the parameters of the original vision transformer pre-trained with DINO on ImageNet. Our model achieved equivalent to the state-of-the-art epoura rformance of 92% accuracy in the OASIS1 dataset under leave-one-out cross-validation. We visualized the performance of the model by extracting average self-attention maps and average brains from the dataset, showing that the model had learned meaningful structural information about demented brains.