Quality-Driven Optimization of QSIPrep Pipelines for Diffusion MRI Preprocessing

Abstract

Diffusion magnetic resonance imaging (dMRI) characterizes the microscopic displacement of water molecules in biological tissues, but remains sensitive to various artifacts. These artifacts can affect image geometry, signal consistency across volumes, and subsequent analyses. Preprocessing pipelines can reduce some of these effects, but several correction tools and strategies may be considered, and their relevance depends on the characteristics of the acquired data. This motivates the development of methods capable of predicting the effect of a preprocessing strategy. The objective of this thesis is to assess whether raw dMRI data contain sufficient information to predict preprocessing-induced changes in several image quality metrics (IQMs). The study is based on the COFITAGE dataset. Two preprocessing configurations were defined using a software framework for diffusion MRI (QSIPrep): a Baseline configuration, applying a limited number of diffusion-specific corrections, and an Advanced configuration, incorporating several artifact-correction tools. Seven IQMs were then used to characterize the raw data and the outputs of both configurations. Based on these measurements, deep learning models were developed to predict the relative changes in IQMs between the Advanced and Raw conditions using a multi-task ordinal classification framework. The dMRI volumes were represented using a pretrained encoder whose parameters were kept frozen, and three mechanisms for aggregating these representations were then compared. On the COFITAGE dataset, the selected IQMs made it possible to quantitatively characterize the differences induced by the preprocessing strategies under study. However, the performance of the deep learning models was insufficient, in their current form, to support the use of this approach as a reliable preprocessing recommendation tool for new acquisitions. Nevertheless, this work provides an exploratory framework for investigating whether preprocessing strategies could be adapted to properties already present in raw dMRI data.