Surface-based tools for characterizing the human brain cortical morphology
- Autores: Alberto Fernández Pena
- Directores de la Tesis: Susana Carmona Cañabate (dir. tes.), Manuel Desco Menéndez (codir. tes.)
- Lectura: En la Universidad Carlos III de Madrid ( España ) en 2023
- Idioma: inglés
- Tribunal Calificador de la Tesis: Juan Domingo Gispert López (presid.), Norberto Malpica González de la Vega (secret.), Gemma Cristina Monte Rubio (voc.)
- Programa de doctorado: Programa de Doctorado en Ciencia y Tecnología Biomédica por la Universidad Carlos III de Madrid
- Materias:
- Enlaces
- Tesis en acceso abierto en: e-Archivo
- Resumen
The cortex of the human brain is highly convoluted. Folded brains present significant advantages over brains without convolutions. For instance, gyrification allows an expansion of cortical surface area without greatly increasing the cranial volume, thus facilitating the pass of the head through the birth channel. Additionally, the lissencephalic brain would require longer axons to connect different cortical regions, thus increasing the metabolic costs and reducing the speed of information exchange. It is essential to state that the forces governing the gyrification process are far from random. There might be variability between subjects and even between hemispheres of the same subject; still, the primary sulci are preserved with remarkable consistency across subjects and are highly reproducible at the intra-species level.
The exact mechanisms leading to cortical gyrification have yet to be well understood. Multiple theories have been proposed and analyzed from molecular, cellular, and mechanical perspectives. However, due to the relationship between the cortical organization and the underlying cytoarchitecture, studying the human brain's cortical morphology and the processes leading to the cortical folds has been critical for an increased understanding of the pathological processes driving psychiatric disorders such as schizophrenia, bipolar disorders, autism, or major depression. Furthermore, charting the normal developmental changes in cortical morphology during adolescence or aging can be of great importance for detecting deviances that may be precursors for pathology.
The accurate characterization of the neurodevelopment processes is challenging. Multiple mechanisms co-occur at a molecular or cellular level and can only be studied through the analysis of ex-vivo samples, usually of animal models. This poses a major constraint for these investigations, as the available sample is limited and complicates longitudinal studies. Magnetic resonance imaging can partially fill the breach, allowing the portrayal of the macroscopic processes surfacing on in-vivo samples. Structural magnetic resonance imaging is a modality that aims to characterize the gross anatomy of the brain, specifically the three primary components that form it: gray matter, white matter, and cerebrospinal fluid. The basis of this modality is to exploit the magnetic properties of the protons, which are abundant in the brain due to its high content in water and, therefore, the presence of hydrogen nuclei non-invasive medical imaging modality that generates 3D representations of the internal anatomy of the head.
Using structural magnetic resonance imaging, different metrics have been defined to measure cortical structure to describe the brain's morphological changes and infer the associated microstructural events. Metrics such as cortical thickness, surface area, or cortical volume help establish a relation between the measured voxels on a magnetic resonance image and the underlying biological processes. However, the existing methods present limitations or room for improvement.
One of the most widely used is cortical thickness, defined as the distance from the cortical interface between white matter and gray matter to the one between cerebrospinal fluid. Precisely this dependence on the gray matter/white matter interface can be a limitation for this technique, as it is known that the contrast between gray and white matter declines with age, making this measure age-dependent. This drawback can also be associated with cortical volume, as it relies on the same gray matter/white matter contrast.
Another important metric is the gyrification index. This ingenious morphometric feature has great potential for assessing brain maturation and aiding in prognosis, patient stratification, and diagnosis in different pathologies. However, its interpretability is severely limited as it is a composite measure made up of sulcal depth, sulcal width, and area.
On their side, methods extracting the lines representing the gyral and sulcal morphology tend to over- or underestimate the total length. These lines can provide important information about how sulcal and gyral regions function differently due to their distinctive ontogenesis. Nevertheless, some methods label every small fold on the cortical surface as a sulcal fundus, thus losing the perspective of lines that travel through the deeper zones of a sulcal basin. On the other hand, some methods are too restrictive, labeling sulcal fundi only for a bunch of primary folds.
To overcome this issue, we have proposed a Laplacian-collapse-based algorithm that can delineate the lines traversing the top regions of the gyri and the fundi of the sulci avoiding anastomotic sulci. For this, the cortex, represented as a surface composed of triangular faces, is segmented into gyral and sulcal surfaces attending to the curvature and depth at every point of the mesh. Each of these resulting surfaces is spatially filtered, obtaining a surface that maintains the original morphometry and topology but with smoother boundaries, i.e., removing the high-frequency noise. Then, a Laplacian-collapse-based algorithm is applied to obtain a thinned representation of the morphology of each structure. This algorithm relies on two terms, a Laplacian term and a mass term. The Laplacian term induces the collapse of the surface, while the mass term induces the vertices to keep the original position. In each iteration of the algorithm, the Laplacian and mass matrices are updated, determining which regions are kept fixed and which ones keep collapsing.
The output of this process is a thin curve for each of the resulting surfaces of the segmentation step. These thin curves are processed to detect where the extremities or endpoints lie. The detection step is performed by making the best effort to avoid the delineation of small anastomotic sulci. For each point in the thinned curve, we establish a neighborhood and perform a principal component analysis on the points inside the neighborhood. The points detected at the extremities of each neighborhood are selected as candidate endpoints. Only those points that appear as candidates in every neighborhood they are included are the ones labeled as endpoints.
As a final step, the sulcal fundi and gyral crown lines are obtained by eroding the surfaces while preserving the structure topology and connectivity between the endpoints. The developed algorithm was assessed by comparing the method against three well-known sulcal lines extraction algorithms in a test-retest database. Additionally, the total length values for five sulci of interest were compared to the reference values obtained using the BrainVISA pipeline. The results showed that the sulcal lines labeled with the developed method were closer to this proposed ground truth length. The statistical analyses also showed that the proposed method crossed deeper and more curved areas for most of the sulci. Finally, the reproducibility was measured using the maximum and mean Hausdorff distances. The developed tool obtained reproducibility scores better or similar to those of previous algorithms. The conducted analyses demonstrated that the designed algorithm was the most balanced approach in the measured features, proving it as a reliable method to extract sulcal lines with an accurate representation of the sulcal topology while avoiding anastomotic branches and the overestimation of the sulcal fundi lines.
A second limitation of the existing metrics concerns the measurement of sulcal width. This metric, understood as the physical distance between the points on opposite sulcal banks, can come in handy in detecting cortical flattening or complementing the information provided by cortical thickness or cortical volume. Another benefit of sulcal width metrics is that it relies on the gray matter/cerebrospinal fluid interface. This interface is much more robust to aging than the one between gray matter and white matter, as the difference in proton densities of gray matter and cerebrospinal fluid remains more stable through aging. Hence, sulcal width is more age-independent than cortical thickness or volume. Additionally, sulcal width presents a metric that, in combination with depth and surface area, can provide a more biological interpretation of the gyrification index. Nevertheless, the existing methods fail to adequately characterize sulcal width as the available algorithms only estimate averaged values for different regions of interest, ignoring the possible intra-sulcal variability and, in some cases failing in small sulci.
Regarding this, we developed a method that estimates the distance from each sulcal point in the cortex to its corresponding opposite, thus providing a per-vertex map of the physical sulcal distances. For this, in the first place, the pial surface is processed to estimate the depth values at each point of the mesh. Then, using the depth map is used to sample the surface at different depth values. This is achieved by generating isolines at regular distances. These isolines are a set of closed curves, one for each depth level. On each of these curves, we need to find the points where the sulcal wall changes in order to match points from different sulcal banks. This is done by applying a line simplification algorithm on the curve and then detecting the segments forming very acute angles, assuming that those segments are where the wall change is happening. Then, the points corresponding to each sulcal wall are labeled and matched with the closest point on a different sulcal bank. Finally, as a correction step, the algorithm traces the line that joins each matched pair of points, checking whether the segment crosses the surface or is enclosed inside the volume formed by the pial. Those connections are recalculated, indicating that the selected points are not in opposing sulcal banks but crossing a gyrus. The sulcal width map is created by assigning the estimated distances to each of the vertices on the surface.
The developed algorithm had a twofold validation. First, it was validated vertex-wise against a simulated sulcus. This allowed us to estimate the performance for every point of a sulcus that resembles a simple fold. Second, the tool was assessed by comparing it with two widely-used sulcal width estimation methods on a test-retest database. This allowed us to validate the output metrics' accuracy and the results reproducibility. As existing tools provide averaged values for different regions of interest, we followed the same approach, averaging the obtained sulcal width values in five sulci defined using BrainVISA. The output of these analyses reported that the developed algorithm behaved very similarly to the reference tools in terms of reproducibility (measuring intraclass correlation and root mean squared error). Also, correlation values between tools were promising, showing Pearson's r values around 0.96 for the central sulcus. The results of the validation analyses demonstrated that the developed method produces a reliable direct measure of the sulcal width, providing vertex-wise maps while tackling the main problems found in the previous literature.
Finally, having developed and validated these two tools, the remaining step was to assess their potential in a clinical use case. For this, we selected a sample of 620 cognitively normal adults ranging from 42 to 95 years old from the Open Access Series of Imaging Studies (OASIS) database. The main idea of this study was to elucidate fine-grained morphological changes in the human cortex with aging by conducting three analyses: a comparison of the age-dependencies of cortical thickness in gyral and sulcal lines, an analysis of how the sulcal and gyral length changes with age, and a vertex-wise study of sulcal width and cortical thickness.
These analyses showed a general flattening of the cortex with aging. In concrete, there were interesting findings, such as a differential age-dependency of thickness thinning in the sulcal and gyral regions, where the cortical thickness in sulcal fundi showed a stronger age-dependency than the same metric in gyri. By demonstrating that our method can detect this difference, our results can pave the way for future in vivo studies focusing on macro- and microscopic changes specific to gyri or sulci, as similar differences have been detected in patients with schizophrenia. Our method can generate new brain-based biomarkers specific to sulci and gyri, and these can be used on large samples to establish normative models to which patients can be compared. Sulcal length analyses showed a pattern of increased length in both gyral and sulcal lines with aging for most of the studied lobes. This increment is compatible with a general flattening of the cortex, as a plain cortex would include longer tails in the sulcal and gyral extremities. In parallel, the vertex-wise analyses show that sulcal width is very sensitive to changes during aging, independent of cortical thickness. This corroborates the concept of sulcal width as a metric that explains, in the least, the unique variance of morphology not fully captured by existing metrics. Our method allows for sulcal width vertex-wise analyses that were not possible previously, potentially changing our understanding of how changes in sulcal width shape cortical morphology.
The two developed methods have been conceived to be fully automated and easily employed by any user. Unlike other published articles, our methods are free and publicly available in their respective repositories (https://github.com/HGGM-LIM/ABLE, https://github.com/HGGM-LIM/SWiM). This makes the algorithms open to community contributions. These methods are designed to work with standard FreeSurfer output as well as output from other toolkits. This means that for the hundreds of existing databases already processed with FreeSurfer, using these tools does not require additional analysis, just executing those methods, potentially reaching a broader user base.
In summary, this thesis presents two new tools, open source and publicly available, for estimating cortical surface-based morphometrics, expanding the existing capabilities in terms of analyses of structural magnetic resonance imaging. The methods have been validated and assessed against existing algorithms. They have also been tested on a real dataset, providing new, exciting insights into cortical morphology and showing their potential for defining innovative biomarkers.
