Voxel and streamline based multimodal variational autoencoder for fiber segmentation in tractography.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Las imágenes de resonancia magnética por difusión (dMRI) son una técnica no invasiva que mide la difusión de las moléculas de agua en el cerebro. En la materia blanca, esta difusión es anisotrópica y sigue predominantemente la dirección de los axones neuronales. Utilizando modelos de difusión, es posible estimar la orientación local de las fibras en cada vóxel. Mediante algoritmos de tractografía es posible generar reconstrucciones tridimensionales de los tractos de materia blanca, denominados datos de tractografía. Estos datos pueden incluir cientos de miles o incluso millones de fibras, lo que hace necesaria su segmentación para identificar aquellas relevantes para el estudio. La segmentación manual por expertos, aunque considerada el estándar de oro, no es viable en grandes bases de datos debido al tiempo requerido y a la variabilidad entre expertos. Por ello, se han desarrollado distintos algoritmos de segmentación automática de datos de tractografía a lo largo de los años.
Esta tesis presenta un nuevo enfoque para la segmentación de fibras cerebrales basado en el uso de un autoencoder variacional multimodal, que integra tanto información espacial de las tractografías como información anatómica proveniente de distintas modalidades de MRI. El método fue entrenado utilizando conjuntos de datos de tractografía probabilística de la base de datos pública HCP y evaluado comparando los resultados de la segmentación con los fascículos de referencia empleados en la creación del algoritmo TractSeg. El método de segmentación propuesto opera en el espacio latente del modelo entrenado, donde se aplica un algoritmo de clasificación basado en radio. Este algoritmo fue optimizado para permitir el uso eficiente de radios personalizados para cada clase, y se desarrolló un procedimiento automático para la optimización de dichos radios. Además, se analizaron distintas combinaciones de datos de entrada, encontrándose que el modelo que incluye información de tractografías, imágenes de anisotropía fraccional (FA), imágenes T1w y los primeros picos de la fODF obtuvo el mejor desempeño.
El desempeño del método propuesto fue evaluado en la segmentación de 66 fascículos de 30 sujetos, comparándose con el método de segmentación TractSeg. Los resultados indican que el método propuesto permite segmentar una mayor cantidad de fibras con una menor distancia entre ellas y una mayor co rrelación de densidad de los fascículos, aunque con un desempeño ligeramente inferior en las métricas basadas en máscaras.
Diffusion magnetic resonance imaging (dMRI) is a non-invasive technique that measures the diffusion of water molecules in the brain. In white matter, this diffusion is anisotropic and predominantly follows the direction of neuronal axons. Using diffusion models, it is possible to estimate the local orientation of fibers in each voxel and, through tractography algorithms, generate three-dimensional reconstructions of white matter tracts, known as tractography data. These datasets can include hundreds of thousands or even millions of fibers, making segmentation necessary to identify those relevant to the study. Manual segmentation by experts, although considered the gold standard, is not feasible for large databases due to the time required and the variability among experts. For this reason, various automatic segmentation algorithms for tractography data have been developed over the years. This thesis presents a novel approach for segmenting brain fibers using a multimodal variational au toencoder, which integrates spatial information from tractographies with anatomical information derived from various MRI modalities. The method was trained using probabilistic tractography datasets from the public HCP database and evaluated by comparing its segmentation results to the reference fascicles used in the creation of the TractSeg algorithm. The proposed segmentation method operates in the la tent space of the trained model, where a radius-based classification algorithm is applied. This algorithm was optimized to efficiently use custom radii for each class, and an automatic procedure for optimizing these radii was developed. Additionally, various combinations of input data were analyzed, with the best performance achieved by the model incorporating tractography data, fractional anisotropy (FA) images, T1w images, and the first peaks of the fODF. The performance of the proposed method was evaluated on the segmentation of 66 fascicles from 30 subjects, compared to the TractSeg segmentation method. Results indicate that the proposed method allows for the segmentation of a greater number of fibers, with shorter distances between them and a higher density correlation of the fascicles, although with slightly lower performance in mask-based metrics.
Diffusion magnetic resonance imaging (dMRI) is a non-invasive technique that measures the diffusion of water molecules in the brain. In white matter, this diffusion is anisotropic and predominantly follows the direction of neuronal axons. Using diffusion models, it is possible to estimate the local orientation of fibers in each voxel and, through tractography algorithms, generate three-dimensional reconstructions of white matter tracts, known as tractography data. These datasets can include hundreds of thousands or even millions of fibers, making segmentation necessary to identify those relevant to the study. Manual segmentation by experts, although considered the gold standard, is not feasible for large databases due to the time required and the variability among experts. For this reason, various automatic segmentation algorithms for tractography data have been developed over the years. This thesis presents a novel approach for segmenting brain fibers using a multimodal variational au toencoder, which integrates spatial information from tractographies with anatomical information derived from various MRI modalities. The method was trained using probabilistic tractography datasets from the public HCP database and evaluated by comparing its segmentation results to the reference fascicles used in the creation of the TractSeg algorithm. The proposed segmentation method operates in the la tent space of the trained model, where a radius-based classification algorithm is applied. This algorithm was optimized to efficiently use custom radii for each class, and an automatic procedure for optimizing these radii was developed. Additionally, various combinations of input data were analyzed, with the best performance achieved by the model incorporating tractography data, fractional anisotropy (FA) images, T1w images, and the first peaks of the fODF. The performance of the proposed method was evaluated on the segmentation of 66 fascicles from 30 subjects, compared to the TractSeg segmentation method. Results indicate that the proposed method allows for the segmentation of a greater number of fibers, with shorter distances between them and a higher density correlation of the fascicles, although with slightly lower performance in mask-based metrics.
Description
Tesis presentada para optar al grado de Magíster en Ciencias de la Ingeniería con Mención en Ingeniería Eléctrica
Keywords
Ingeniería eléctrica, Neuroimagen, Redes neurales (neurobiología)