Advanced Methods for Robust Neuroimage Analysis

Abstract

Magnetic resonance imaging (MRI) is the central modality for acquiring images of the brain in biomedical research. Advanced computational pipelines aid researchers along the way, from initially acquiring MRI to ultimately reaching research conclusions. The researchers' interaction with these pipelines impacts the power and validity of research findings. Software tools can mitigate error and bias, but inadequate analysis methods can also introduce them. To aid the research community, we propose three novel methods that make neuroimage analysis more robust and intervene at different points in the analysis process. <br/> First, we propose a method to track head motion during the MRI acquisition. Motion artifacts are the most prominent MRI artifacts and are known to cause biased morphometric analysis results. Our measurements allow researchers to reduce this bias by including head motion in statistical modeling. The method introduces an improved robust registration approach for depth images acquired by an optical tracking camera, along with head motion metrics and novel evaluation approaches. Additionally, we show that the acquired measurements can be used to train a deep-learning model that estimates the head motion from images alone. Second, we address a major challenge in the robustness of neuroimaging methods by providing a novel approach for analyzing lesion-affected scans. We propose a lesion inpainting tool that replaces damaged tissue with healthy-looking tissue in MRI scans. This tool allows consecutive analyses to run as if no lesions were present, enabling new analyses on data that would previously be discarded. It leverages a denoising diffusion probabilistic model (DDPM) combined with a novel inference scheme, allowing for mask-independent, memory-efficient volumetric inpainting. <br/> Finally, we address multiple gaps in current corpus callosum analysis methods. The corpus callosum (CC) is the central connection between the brain's hemispheres and of high interest for research of disease and aging. Previous work has successfully developed automated tools for CC analysis but has shortcomings in accuracy, reliability, and performance. To close this gap, we propose a comprehensive corpus callosum analysis tool that performs segmentation and surface reconstruction and then derives suitable metrics for statistical modeling. We show that building a state-of-the-art analysis method extends far beyond simple volumetric segmentation. Our method standardizes the head position specifically for unbiased segmentation, then uses the CC segmentation to derive curvature, length, thickness, and other advanced CC-specific metrics. The combination of method improvements in every step results in a method that detects disease effects where they would be missed by previous methods. <br/> We place special attention on the thorough evaluation of the proposed methods, testing generalization to unseen datasets and diseases. All proposed methods were released to the research community as open-source tools (on https://github.com/Deep-MI). The motion-tracking tools are currently deployed in the Rhineland Study, a large population study with over 10,000 participants. Furthermore, the general image processing tools are integrated into the established FastSurfer neuroimage toolbox, allowing the global research community to seamlessly interface with these novel analysis techniques.