Automatic detection of outliers in longitudinal data

[valosekj]

A single patient commonly undergoes several MRI sessions across time (i.e., first session at month 0, second session at month X, third session at month Y, etc.). Images are then analyzed to explore longitudinal trends (e.g., spinal cord compression progression, MS lesion monitoring, etc.).
It is crucial to ensure that image quality is satisfactory across all sessions. However, for some acquisitions, such as diffusion-weighted imaging (DWI) or functional MRI (fMRI), the visual quality control of the data can be tricky due to its low SNR and a large number of volumes (DWI and fMRI are 4D data).
An automatic (deep-learning based) detection of outliers (i.e., session(s) with poor data quality) would dramatically facilitate the process.

[valosekj]

Relevant papers:

• Transformer-based out-of-distribution detection for clinically safe segmentation
• Unsupervised Brain Anomaly Detection and Segmentation with Transformers

Mark Graham had a relevant presentation about "out-of-distribution" detection during "Generative Al for Medical Imaging with MONAI [SE52190]", NVIDIA GTC 2023.

MONAI implementation of "out-of-distribution" detection using VQ-VAE here.