Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - ARTEMIS (Assessment of Reserve: Translational Evaluation of Medical Images and Statistics-Prediction models for outcomes of brain health)

Teaser

Stroke and cognitive decline are among the leading contributors to disease burden and long-term disability worldwide. However, despite their prevalence, the contributing disease processes are not fully understood. This is in part due to the lack of (early) prediction models...

Summary

Stroke and cognitive decline are among the leading contributors to disease burden and long-term disability worldwide. However, despite their prevalence, the contributing disease processes are not fully understood. This is in part due to the lack of (early) prediction models and ways to characterize protective mechanisms, which can help to distinguish between patients and healthy controls before symptoms show. Such prediction models can facilitate prevention strategies for adverse cognitive and functional outcomes, thereby enriching patients’ life quality and reduce the economic burden on society. Advanced neuroimaging techniques such as MRI have provided additional insight into the underlying disease biology. One major challenge when using neuroimaging techniques lies in the fact that large amounts of data are required to account for variations in clinical presentation and assessment, necessitating the use of dedicated pipelines for extracting phenotypes. However, most pipelines are developed in research settings and tend to fail when applied to clinical cohorts, leading to a subpar use of rich, available datasets.

Here, a fully-automated, translational pipeline for extracting imaging phenotypes from data acquired in clinical and research settings is developed with a particular focus on outlining white matter hyperintensities (WMH). WMH are a common phenotype in aging and across diseases, however, group differences are poorly understood. This makes WMH a prime candidate for extracting additional information, which can be used for outcome prediction. The proposed prediction models in this project utilize newly extracted characteristics, clinical/demographic information and a latent variable construct to predict general cognitive decline and outcome after stroke. In particular, the proposed latent variable has shown promise in acting as a surrogate measure for protective mechanisms in stroke patients, where its biological meaning is assessed as part of this project.

Work performed

\"In the outgoing phase, the automated WMH segmentation pipeline was developed and applied to over 5500 acute ischemic stroke patients. We created a new, deep-learning based brain extraction methodology in clinical magnetic resonance imaging, specifically FLuid Attenuated Inversion Recovery (FLAIR) images, a common processing step in which other elements, such as the skull, are removed from the image. Our methodology significantly outperformed state-of-the-art techniques on these clinical scans. Furthermore, it allowed the estimation of each patient\'s brain volume, which was utilized in an automated quality control step, and also allowed the investigation of the effect of brain volume on long-term stroke outcome. In addition, the processing pipeline only requires a rough registration to a template (linear instead of nonlinear approaches), which increased its utility by reducing the computational cost to under 3.5 minutes per patient (initially 2.5 hours), and by increasing the reliability of the pipeline. Outlines of the WMH have been improved using a deep-learning based framework for segmentation. Finally, we assessed the efficacy of the pipeline by comparing the results to manual segmented outlines, showing good agreement.

To improve nonlinear registrations in clinical image data, which can further help to reduce the complexity of the WMH segmentation task, we created a new multi-template-based registration framework. Each patient\'s FLAIR image is first registered to, i.e. spatially overlapped with, a set of age-specific templates before they are then transferred into a common space. After developing an automated ventricle segmentation algorithm on clinical FLAIR sequences (also based on deep-learning methods), we are able to assess the quality of registration automatically, based on the overlap of the ventricles of the template and the subject.

To investigate WMH patterns beyond a single volumetric measure, we explored the spatial disease patterns in the brain in two complimentary ways. First, as WMH is considered a vascular disease, we utilized information of vascular territories in the brain, i.e. areas in the brain stratified by the major artery which supplies the blood. This resulted in the identification of spatial variations, which are affected/modified by different risk factors (such as hypertension). Additionally, by utilizing the ventricle segmentations mentioned above, we investigated the \'classical\' approach of differentiating WMH burden based on its location with respect to the ventricles (periventricular vs. subcortical). Similarly to the vascular territories, individual disease patterns were driven by different risk factors, however, only the periventricular disease burden contributed in modeling long-term outcome after stroke.

Utilizing a latent variable model, we also investigated the possibility of quantifying the often-observed protective mechanism in the brain leading to better post-stroke functional outcome. Based on the idea of \"\"brain reserve\"\", which is widely studied in populations with cognitive decline, we extended on existing approaches. Including pre-existing disease burden allowed us to describe the remainder of reserve that is available after accounting for the part utilized to compensate for other factors and diseases. The resulting concept characterizes the \"\"effective reserve\"\" and was estimated in a set of stroke patients where we showed its relation with long-term stroke outcome. In these patients a higher reserve was associated with better outcome. Additionally, we demonstrated that the model including effective reserve, explains the observed data better than models without it.\"

Final results

So far, this project has made significant progress beyond the state of the art. We created an effective, high-throughput automated pipeline for clinically relevant MRI phenotype analyses, which has the potential to accelerate the pace of scientific and medical discoveries and to advance the development of clinical applications in risk and outcome modeling in stroke. Importantly, by utilizing clinical data, as it becomes available in the emergency room, we ensured translatability of the investigated approaches and started to close an important gap that currently exists in the translational application of advanced MRI analyses. Additionally, we demonstrated the first conceptualization of the brains capacity to compensate for the negative effects of stroke and identified multiple contributing factors modifying stroke outcome, which helped us to create a more complete picture of the adverse effects of stroke.

Based on our progress so far, we believe that our research will lead to a better understanding of disease processes and patient specific differences in outcomes. With the forthcoming investigation in a population-based cohort, we will be able to create a model for general brain health, which may subsequently lead to effective disease models, help in developing prevention strategies for adverse cognitive and functional outcomes, and ultimately guide us to enriching patients\' quality of life and reducing the economic burden on society.

Website & more info

More info: http://www.markus-schirmer.com/artemis.html.