Quantitative neuroimaging with handcrafted and deep radiomics in neurological diseases

The motivation behind this thesis is to explore the potential of "radiomics" in the field of neurology, where early diagnosis and accurate treatment selection are crucial for improving patient outcomes. Neurological diseases are a major cause of disability and death globally, and there is a pressing need for reliable imaging biomarkers to aid in disease detection and monitoring. While radiomics has shown promising results in oncology, its application in neurology remains relatively unexplored. Therefore, this work aims to investigate the feasibility and challenges of implementing radiomics in the neurological context, addressing various limitations and proposing potential solutions. The thesis begins with a demonstration of the predictive power of radiomics for identifying important diagnostic biomarkers in neuro-oncology. Building on this foundation, the research then delves into radiomics in non-oncological neurology, providing an overview of the pipeline steps, potential clinical applications, and existing challenges. Despite promising results in proof-of-concept studies, the field faces limitations, mostly data-related, such as small sample sizes, retrospective nature, and lack of external validation. To explore the predictive power of radiomics in non-oncological tasks, a radiomics approach was implemented to distinguish between multiple sclerosis patients and normal controls. Notably, radiomic features extracted from normal-appearing white matter were found to contain distinctive information for multiple sclerosis detection, confirming the hypothesis of the thesis. To overcome the data harmonization challenge, in this work quantitative mapping of the brain was used. Unlike traditional imaging methods, quantitative mapping involves measuring the physical properties of brain tissues, providing a more standardized and consistent data representation. By reconstructing the physical properties of each voxel based on multi-echo MRI acquisition, quantitative mapping produces data that is less susceptible to domain-specific biases and scanner variability. Additionally, the insights gained from quantitative mapping are building the bridge toward the physical and biological properties of brain tissues, providing a deeper understanding of the underlying pathology. Another crucial challenge in radiomics is robust and fast data labeling, particularly segmentation. A deep learning method was proposed to perform automated carotid artery segmentation in stroke at-risk patients, surpassing current state-of-the-art approaches. This novel method showcases the potential of automated segmentation to enhance radiomics pipeline implementation. In addition to addressing specific challenges, the thesis also proposes a community-driven open-source toolbox for radiomics, aimed at enhancing pipeline standardization and transparency. This software package would facilitate data curation and exploratory analysis, fostering collaboration and reproducibility in radiomics research. Through an in-depth exploration of radiomics in neuroimaging, this thesis demonstrates its potential to enhance neurological disease diagnosis and monitoring. By uncovering valuable information from seemingly normal brain tissues, radiomics holds promise for early disease detection. Furthermore, the development of innovative tools and methods, including deep learning and quantitative mapping, has the potential to address data labeling and harmonization challenges. Looking to the future, embracing larger, diverse datasets and longitudinal studies will further enhance the generalizability and predictive power of radiomics in neurology. By addressing the challenges identified in this thesis and fostering collaboration within the research community, radiomics can advance toward clinical implementation, revolutionizing precision medicine in neurology.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.