Comprehensive analysis of synthetic learning applied to neonatal brain MRI segmentation.

Brain segmentation from neonatal MRI images is a very challenging task due to large changes in the shape of cerebral structures and variations in signal intensities reflecting the gestational process. In this context, there is a clear need for segmentation techniques that are robust to variations in image contrast and to the spatial configuration of anatomical structures. In this work, we evaluate the potential of synthetic learning, a contrast-independent model trained using synthetic images generated from the ground truth labels of very few subjects. We base our experiments on the dataset released by the developmental Human Connectome Project, for which high-quality images are available for more than 700 babies aged between 26 and 45 weeks postconception. First, we confirm the impressive performance of a standard UNet trained on a few volumes, but also confirm that such models learn intensity-related features specific to the training domain. We then confirm the robustness of the synthetic learning approach to variations in image contrast. However, we observe a clear influence of the age of the baby on the predictions. We improve the performance of this model by enriching the synthetic training set with realistic motion artifacts and over-segmentation of the white matter. Based on extensive visual assessment, we argue that the better performance of the model trained on real T2w data may be due to systematic errors in the ground truth. We propose an original experiment allowing us to show that learning from real data will reproduce any systematic bias affecting the training set, while synthetic models can avoid this limitation. Overall, our experiments confirm that synthetic learning is an effective solution for segmenting neonatal brain MRI. Our adapted synthetic learning approach combines key features that will be instrumental for large multisite studies and clinical applications.

A framework based on sulcal constraints to align preterm, infant and adult human brain images acquired in vivo and post mortem.

Robust spatial alignment of post mortem data and in vivo MRI acquisitions from different ages, especially from the early developmental stages, into standard spaces is still a bottleneck hampering easy comparison with the mainstream neuroimaging results. In this paper, we test a landmark-based spatial normalization strategy as a framework for the seamless integration of any macroscopic dataset in the context of the Human Brain Project (HBP). This strategy stems from an approach called DISCO embedding sulcal constraints in a registration framework used to initialize DARTEL, the widely used spatial normalization approach proposed in the SPM software. We show that this strategy is efficient with a heterogeneous dataset including challenging data as preterm newborns, infants, post mortem histological data and a synthetic atlas computed from averaging the ICBM database, as well as more commonly studied data acquired in vivo in adults. We then describe some perspectives for a research program aiming at improving folding pattern matching for atlas inference in the context of the future HBP's portal.

Spatial normalization of brain images and beyond.

The deformable atlas paradigm has been at the core of computational anatomy during the last two decades. Spatial normalization is the variant endowing the atlas with a coordinate system used for voxel-based aggregation of images across subjects and studies. This framework has largely contributed to the success of brain mapping. Brain spatial normalization, however, is still ill-posed because of the complexity of the human brain architecture and the lack of architectural landmarks in standard morphological MRI. Multi-atlas strategies have been developed during the last decade to overcome some difficulties in the context of segmentation. A new generation of registration algorithms embedding architectural features inferred for instance from diffusion or functional MRI is on the verge to improve the architectural value of spatial normalization. A better understanding of the architectural meaning of the cortical folding pattern will lead to use some sulci as complementary constraints. Improving the architectural compliance of spatial normalization may impose to relax the diffeomorphic constraint usually underlying atlas warping. A two-level strategy could be designed: in each region, a dictionary of templates of incompatible folding patterns would be collected and matched in a way or another using rare architectural information, while individual subjects would be aligned using diffeomorphisms to the closest template. Manifold learning could help to aggregate subjects according to their morphology. Connectivity-based strategies could emerge as an alternative to deformation-based alignment leading to match the connectomes of the subjects rather than images.

On the Influence of Confounding Factors in Multisite Brain Morphometry Studies of Developmental Pathologies: Application to Autism Spectrum Disorder.

Pooling data acquired on different MR scanners is a commonly used practice to increase the statistical power of studies based on MRI-derived measurements. Such studies are very appealing since they should make it possible to detect more subtle effects related to pathologies. However, the influence of confounds introduced by scanner-related variations remains unclear. When studying brain morphometry descriptors, it is crucial to investigate whether scanner-induced errors can exceed the effect of the disease itself. More specifically, in the context of developmental pathologies such as autism spectrum disorders (ASD), it is essential to evaluate the influence of the scanner on age-related effects. In this paper, we studied a dataset composed of 159 anatomical MR images pooled from three different scanners, including 75 ASD patients and 84 healthy controls. We quantitatively assessed the effects of the age, pathology, and scanner factors on cortical thickness measurements. Our results indicate that scan pooling from different sites would be less fruitful in some cortical regions than in others. Although the effect of age is consistent across scanners, the interaction between the age and scanner factors is important and significant in some specific cortical areas.

Deep sulcal landmarks: algorithmic and conceptual improvements in the definition and extraction of sulcal pits.

Recent interest has been growing concerning points of maximum depth within folds, the sulcal pits, that can be used as reliable cortical landmarks. These remarkable points on the cortical surface are defined algorithmically as the outcome of an automatic extraction procedure. The influence of several crucial parameters of the reference technique (Im et al., 2010) has not been evaluated extensively, and no optimization procedure has been proposed so far. Designing an appropriate optimization framework for these parameters is mandatory to guarantee the reproducibility of results across studies and to ensure the feasibility of sulcal pit extraction and analysis on large cohorts. In this work, we propose a framework specifically dedicated to the optimization of the parameters of the method. This optimization framework relies on new measures for better quantifying the reproducibility of the number of sulcal pits per region across individuals, in line with the assumptions of one-to-one correspondence of sulcal roots across individuals which is an explicit aspect of the sulcal roots model (Régis et al., 2005). Our procedure benefits from a combination of improvements, including the use of a convenient sulcal depth estimation and is methodologically sound. Our experiments on two different groups of individuals, with a total of 137 subjects, show an increased reliability across subjects in deeper sulcal pits, as compared to the previous approach, and cover the entire cortical surface, including shallower and more variable folds that were not considered before. The effectiveness of our method ensures the feasibility of a systematic study of sulcal pits on large cohorts. On top of these methodological advances, we quantify the relationship between the reproducibility of the number of sulcal pits per region across individuals and their respective depth and demonstrate the relatively high reproducibility of several pits corresponding to shallower folds. Finally, we report new results regarding the local pit asymmetry, providing evidence that the algorithmic and conceptual approach defended here may contribute to better understanding of the key role of sulcal pits in neuroanatomy.

Benchmark data for sulcal pits extraction algorithms.

This article contains data related to the research article Auzias et al. (2015) [1]. This data can be used as a benchmark for quantitative evaluation of sulcal pits extraction algorithm. In particular, it allows a quantitative comparison with our method, and the assessment of the consistency of the sulcal pits extraction across two well-matched populations.

Atypical sulcal anatomy in young children with autism spectrum disorder.

Autism spectrum disorder is associated with an altered early brain development. However, the specific cortical structure abnormalities underlying this disorder remain largely unknown. Nonetheless, atypical cortical folding provides lingering evidence of early disruptions in neurodevelopmental processes and identifying changes in the geometry of cortical sulci is of primary interest for characterizing these structural abnormalities in autism and their evolution over the first stages of brain development. Here, we applied state-of-the-art sulcus-based morphometry methods to a large highly-selective cohort of 73 young male children of age spanning from 18 to 108 months. Moreover, such large cohort was selected through extensive behavioral assessments and stringent inclusion criteria for the group of 59 children with autism. After manual labeling of 59 different sulci in each hemisphere, we computed multiple shape descriptors for each single sulcus element, hereby separating the folding measurement into distinct factors such as the length and depth of the sulcus. We demonstrated that the central, intraparietal and frontal medial sulci showed a significant and consistent pattern of abnormalities across our different geometrical indices. We also found that autistic and control children exhibited strikingly different relationships between age and structural changes in brain morphology. Lastly, the different measures of sulcus shapes were correlated with the CARS and ADOS scores that are specific to the autistic pathology and indices of symptom severity. Inherently, these structural abnormalities are confined to regions that are functionally relevant with respect to cognitive disorders in ASD. In contrast to those previously reported in adults, it is very unlikely that these abnormalities originate from general compensatory mechanisms unrelated to the primary pathology. Rather, they most probably reflect an early disruption on developmental trajectory that could be part of the primary pathology.

Local landmark alignment for high-resolution fMRI group studies: toward a fine cortical investigation of hand movements in human.

Converging evidence conclusively demonstrates the robust relationship between anatomical landmarks and underlying functional organization in primary cortical regions. In consequence, a precise alignment across subjects of such specific individual landmarks should improve the overlap of the corresponding functional areas and thus the detection of active clusters at the group level. In an effort to define a dedicated processing pipeline for a fine non-invasive exploration of the motor cortex in human, we evaluated four recent non-linear registration methods based on anatomical and functional indexes. We used high-resolution functional MRI data to finely reveal the impact of the registration on the cortical assignment of the detected clusters. Our results first demonstrate that the quality of registration strongly affects the statistical significance and the assignment of activated clusters to specific anatomical regions, here in the primary motor area. Our results also illustrate the bias induced by the chosen reference template on the detected clusters. The analysis of the Jacobian of the deformation field informs us about how each method deforms the anatomical structures and functional maps. The methodology we propose, combining high resolution fMRI and non-linear registration method, allows a robust non-invasive exploration of the motor cortex.

Model-driven harmonic parameterization of the cortical surface: HIP-HOP.

In the context of inter subject brain surface matching, we present a parameterization of the cortical surface constrained by a model of cortical organization. The parameterization is defined via an harmonic mapping of each hemisphere surface to a rectangular planar domain that integrates a representation of the model. As opposed to previous landmark-based registration methods we do not match folds between individuals but instead optimize the fit between cortical sulci and specific iso-coordinate axis in the model. This strategy overcomes some limitation to sulcus-based registration techniques such as topological variability in sulcal landmarks across subjects. Experiments on 62 subjects with manually traced sulci are presented and compared with the result of the Freesurfer software. The evaluation involves a measure of dispersion of sulci with both angular and area distortions. We show that the model-based strategy can lead to a natural, efficient and very fast (less than 5 min per hemisphere) method for defining inter subjects correspondences. We discuss how this approach also reduces the problems inherent to anatomically defined landmarks and open the way to the investigation of cortical organization through the notion of orientation and alignment of structures across the cortex.

Automatic sulcal line extraction on cortical surfaces using geodesic path density maps.

We present here a method that is designed to automatically extract sulcal lines on the mesh of any cortical surface. The method is based on the definition of a new function, the Geodesic Path Density Map (GPDM), within each sulcal basin (i.e. regions with a negative mean curvature). GPDM indicates at each vertex the likelihood that a shortest path between any two points of the basins boundary goes through that vertex. If the distance used to compute shortest path is anisotropic and constrained by a geometric information such as the depth, the GPDM indicates the likelihood that a vertex belongs to the sulcal line in the basin. An automatic GPDM adaptive thresholding procedure is proposed and sulcal lines are then defined. The process has been validated on a set of 25 subjects by comparing results to the manual segmentation from an expert and showed an average error below 2mm. It is also compared to our previous reference method in the context of inter-subject cortical surface registration and shows an significant improvement in performance.