AI-based optimization of fetal MR image processing to characterize brain development

Magnetic Resonance Imaging (MRI) of the fetal brain is a noninvasive tool used to study the characteristics of central nervous system development. Over the last two decades, its use has become widespread thanks to a combination of advances in imaging and analysis technologies. The development of the fetal brain is a fascinating yet extremely complex process encompassing different stages of growth and organization that are crucial for proper anatomical and functional formation. The increasing number of studies using MRI to investigate the fetal brain reflects the growing interest within fetal research communities to enhance our understanding of the neurodevelopment process. My doctoral research focused on the optimization of MR image processing methods to assess fetal central nervous system development during pregnancy. Moreover, we concentrated on two specific fetal brain structures that are crucial for monitoring neurodevelopment, the cortex and the ganglionic eminence (GE), and discussed novel methodological solutions (both Artificial Intelligence-based and traditional) to study them throughout a highly extended range of gestational weeks. First, we surveyed the existing literature on fetal brain MRI, investigating which techniques are used for fetal brain segmentation from MR images using deep learning (DL) methods and the available fetal brain atlases and datasets. We then optimized the fetal brain reconstruction pipeline. We performed a qualitative and quantitative characterization of the high-resolution reconstruction of fetal brain from different acquisition sequences obtained using publicly available volume reconstruction toolkits and we introduced a light-weight convolutional neural network to automatically identify potential nondiagnostic slices that could compromise the volume reconstruction process. We introduced novel tools to analyze high-resolution reconstructed fetal brain MR images and provide insights into the architecture of the human brain. The first tool involves the definition of a spatio-temporal MRI atlas of the fetal brain with 0.5 mm isotropic resolution. The second tool is an automatic DL-based approach for GE extraction from high-resolution reconstructed T2w MR images. This tool originated from a synergic collaboration between three clinical centers (IRCCS Eugenio Medea La Nostra Famiglia; Medical University of Vienna; Fondazione Ca' Granda - Ospedale Maggiore Policlinico), with the aim of developing an early diagnostic automated image analysis tool that can be used by different institutions. Moreover, we described three cutting-edge DL-based algorithms participating in the Fetal Tissue Annotation and Segmentation (FeTA) Challenge, an international competition held in the framework of MICCAI 2024. Notably, two different DL-based approaches are proposed to segment the brain into seven tissues, and a DL-based approach is proposed to predict the measurements of five fetal brain structures in the three orthogonal planes. Finally, we quantitatively assessed brain development during pregnancy to investigate structural changes associated with different pathological conditions. We specifically focused on the cerebral cortex and the GE. Global multidimensional point-wise shape signatures were used alongside scalar point-wise curvature-based signatures to study the gyrification process of the fetal cerebral cortex. Furthermore, volumetric analysis of the GE structure and biometric measurements of brain structure growth were performed to investigate the possible influence of maternal depression on fetal brain development. In-utero fetal imaging is a powerful tool that may support the identification of high-risk fetuses. The proposed MR processing methods will improve the monitoring of brain development, providing new critical insights into prodromal signs of potential clinical conditions which may play an important role in promoting early postnatal habilitative/rehabilitative therapy.