Fetal magnetic resonance imaging (MRI) has been increasingly used as a powerful complement imaging modality to ultrasound imaging (US) for the clinical evaluation of prenatal abnormalities. Specifically, clinical application of fetal MRI has been significantly improved in the nineties by hardware and software advances with the development of ultrafast multi-slice T2-weighted (T2w) acquisition sequences able to freeze the unpredictable fetal motion and provide excellent soft-tissue contrast. Fetal motion is indeed the major challenge in fetal MRI and slice acquisition time should be kept as short as possible. As a result, typical fetal MRI examination involves the acquisition of a set of orthogonally planned scans of thick two-dimensional slices, largely free of intra-slice motion artifacts. The poor resolution in the slice-select dimension as well as possible motion occurring between slices limits further quantitative data analysis, which is the key for a better understanding of the developing brain but also the key for the determination of operator-independent biomarkers that might significantly facilitate fetal diagnosis and prognosis.
To this end, several research groups have developed in the past ten years advanced image processing methods, often denoted by motion-robust super-resolution (SR) techniques, to reconstruct from a set of clinical low-resolution (LR) scans, a high-resolution (HR) motion-free volume. SR problem is usually modeled as a linear inverse problem describing the imaging degradation due to acquisition and fetal motion. Typically, such approaches consist in iterating between slice motion estimation that estimates the motion parameters and SR that recovers the HR image given the estimated degradation model. This thesis focuses on the development of novel advanced image processing methods, which have enabled the design of a completely automated reconstruction pipeline for fetal MRI. The proposed techniques help in improving state-of-the-art fetal MRI reconstruction in terms of efficiency, robustness and minimized user-interactions, with the ultimate goal of being translated to the clinical environment.
The first part focuses on the development of a more efficient Total Variation (TV)-regularized optimization algorithm for the SR problem. The algorithm uses recent advances in convex optimization with a novel adaptive regularization strategy to offer simultaneously fast, accurate and robust solutions to the fetal image recovery problem. Extensive validations on both simulated fetal and real clinical data show the proposed algorithm is highly robust in front of motion artifacts and that it offers the best trade-off between speed and accuracy for fetal MRI recovery as in comparison with state-of-the art methods.
The second part focuses on the development of a novel automatic brain localization and extraction approach based on template-to-slice block matching and deformable slice-totemplate registration. Asmost fetal brain MRI reconstruction algorithms rely only on brain tissue-relevant voxels of low-resolution (LR) images to enhance the quality of inter-slice motion correction and image reconstruction, the fetal brain needs to be localized and extracted as a first step. These tasks generally necessitate user interaction, manually or semi-automatically done. Our methods have enabled the design of completely automated reconstruction pipeline that involves intensity normalization, inter-slice motion estimation, and super-resolution.
Quantitative evaluation on clinical MRI scans shows that our approach produces brain masks that are very close to manually drawn brain masks, and ratings performed by two expert observers show that the proposed pipeline achieves similar reconstruction quality to reference reconstruction based on manual slice-by-slice brain extraction without any further effort. The third part investigates the possibility of automatic cortical folding quantification, one of the best biomarkers of brain maturation, by combining our automatic reconstruction pipeline with a state-of-the-art fetal brain tissue segmentation method and existing automated tools provided for adult brain’s cortical folding quantification. Results indicate that our reconstruction pipeline can provide HR MR images with sufficient quality that enable the use of surface tessellation and active surface algorithms similar to those developed for adults to extract meaningful information about fetal brain maturation.
Finally, the last part presents new methodological improvements of the reconstruction pipeline aiming at improving the quality of the image for quantitative data analysis, whose accuracy is highly dependent on the quality and resolution of the reconstructed image. In particular, it presents a more consistent and global magnetic bias field correction method which takes advantage of the super-resolution framework to provide a final reconstructed image quasi free of the smooth bias field. Then, it presents a new TV SR algorithm that uses the Huber norm in the data fidelity term to be more robust to non-Gaussian outliers. It also presents the design of a novel joint reconstruction-segmentation framework and the development of a novel TV SR algorithm driven by segmentation to produce images with enhanced edge information that could ultimately improve their segmentation. Finally, it preliminary investigates the capability of increasing the resolution in the in-plane dimensions using SR to ultimately reduce the partial volume effect.