Abstract | Keywords | Export | Availability | Bookmark

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Magnetic Resonance Imaging (MRI) has become a cornerstone in diagnosing and monitoring brain pathologies. However, the manual segmentation of pathological regions and the parcellation of healthy brain tissues are labor-intensive, prone to errors, and demand high expertise. This thesis explores the potential of deep learning techniques, particularly those based on a U-Net architecture, to automate these tasks, focusing on both pathology segmentation and healthy tissue parcellation. This research emphasizes the superior performance and adaptability of nnU-Net in segmenting brain pathologies and parcellating healthy brain tissues. nnU-Net’s automated pipeline configuration significantly enhances segmentation accuracy and efficiency across diverse datasets, demonstrating its robustness and clinical applicability. The thesis introduces a novel two-step pipeline designed for clinical use, combining lesion segmentation and cortical parcellation. This pipeline begins with the acquisition of T1-weighted and T2-weighted MRI images, processed together through two distinct neural networks. The first network performs lesion segmentation, creating a binary mask to accurately identify pathological tissues. The second network focuses on parcellating the healthy brain regions, dividing the brain into 84 distinct areas. A key innovation of this work is the creation of a virtual database by integrating lesioned areas from real brain images into healthy brain images from the Human Connectome Project. This technique allows the training of the parcellation network on a sufficiently large and diverse dataset, facilitating the development of robust deep learning models capable of handling complex brain structures. The results demonstrate significant enhancements in the accuracy and efficiency of brain segmentation and parcellation processes, highlighting the critical role of automated pipeline configuration in achieving superior performance. Future research should aim to further optimize the nnU-Net configuration, explore fine-tuning techniques, and investigate hierarchical approaches to brain parcellation. Collaborations to acquire larger and more diverse datasets will enhance the robustness and generalizability of the proposed models, ultimately advancing the capabilities of automated brain imaging analysis.