Scalable Unsupervised Subtle Anomaly Detection from Longitudinal MR Imaging Data: Application to Parkinson’s Disease

This thesis presents a novel approach to unsupervised anomaly detection for longitudinal medical imaging data, with a specific focus on Parkinson’s disease. The increasing prevalence of neurodegenerative disorders, especially in an aging population and the challenges lying in their early diagnosis require innovative methods for analyzing medical images over time. We propose a pipeline that leverages the temporal aspect of MRI data to improve the detection of subtle brain anomalies. Our methodology combines a longitudinal variational autoencoder (LVAE) with a mixed effects model to capture both spatial and temporal features of brain MRIs. This approach allows for the generation of time-aware encodings that can be used for anomaly detection. The pipeline is designed to work with multiple MRI modalities, including T1-weighted, Fractional Anisotropy (FA), and Mean Diffusivity (MD) images. We evaluate our method on synthetic data, demonstrating its ability to reconstruct and project images across time points accurately. While the full implementation on Parkinson’s disease data was not completed, our results on synthetic data show promise. This work lays a foundation for future research in unsupervised anomaly detection using longitudinal medical imaging datasets. It offers potential for earlier and more accurate detection of neurodegenerative diseases, as well as improved monitoring of disease progression. Future work may explore the integration of additional MRI modalities, the use of positional encoding for image patches, and the application of diffusion models to further enhance anomaly detection capabilities.