Generación de base de datos y algoritmo de segmentación automática de núcleos hepáticos en imágenes 3D de microscopia mediante redes neuronales.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La comprensión de la organización de los órganos y la coordinación celular para formar tejidos funcionales es fundamental en biología celular y del desarrollo. La histología de los tejidos es crucial para el diagnóstico de diversas patologías, incluyendo el cáncer. Sin embargo, las evaluaciones histopatológicas tradicionales en dos dimensiones (2D) resultan en una pérdida de contexto tridimensional (3D) del tejido. El hígado, un órgano vital con múltiples funciones, depende de una compleja arquitectura 3D para su funcionamiento adecuado. Por ello, es esencial conocer la información espacial exacta de los núcleos y célular para comprender sus funciones e interacciones.
Para superar estas limitaciones, se utilizan modelos geométricos y softwares como Motion Tracking (MT), procesan las imágenes de hígado obtenidas mediante microscopía confocal con fluorescencia, generando una reconstrucción 3D de alta resolución de todos los componentes celulares y subcelulares importantes en el tejido. Un paso clave es la segmentación de núcleos celulares, la cual enfrenta desafíos en núcleos muy cercanos.
Las redes neuronales convolucionales (CNNs) son especialmente efectivas para el procesamiento y análisis de datos visuales. La Residual Attention U-Net 3D es una variante avanzada de CNN diseñada para la segmentación de imágenes volumétricas, combinando la arquitectura U-Net con bloques residuales y mecanismos de atención para mejorar la precisión de la segmentación. En esta memoria, se entrenó un modelo basado en la red mencionada, para lo cual se generó una base de datos de núcleos 3D de alta calidad en imágenes de microscoía de fluorescencia del lobulillo hepático. Además, para entrenar esta red neuronal se decidió un enfoque multitarea, realizando segmentación binaria de los núcleos hepáticos y predicción de un campo vectorial para estimar los centroides de cada núcleos y generar la segmentación por instancia [1, 2, 3, 4, 5].
Para la creación de la base de datos, se segmentaron aproximadamente 2900 núcleos 3D en tres sujetos de control (C1, C2 y C3) en las regiones periportal, central y pericentral del lobulillo hepático. Los sujetos C1 y C2 se utilizaron para el entrenamiento, mientras que el sujeto C3 se empleó para la evaluación. Además del método de segmentación de MT, se utilizó Cellpose [3] para comparar los resultados. Nuestro modelo propuesto mostró una precisión media superior al 17 %, 28% y 33% en comparación con MT en las regiones central, periportal y pericentral, respectivamente. También superó a Cellpose en un 5 %, 17% y 7% en las mismas regiones, respectivamente. En particular, en la región periportal, nuestro modelo logró resultados significativamente mejores, demostrando su capacidad para manejar la segmentación en regiones complejas del tejido hepático.
The understanding of organ organization and cellular coordination to form functional tissues is fundamental in cell and developmental biology. Tissue histology is crucial for the diagnosis of various pathologies, including cancer. However, traditional two-dimensional (2D) histopathological evaluations result in a loss of three dimensional (3D) context of the tissue. The liver, a vital organ with multiple functions, relies on a complex 3D architecture for its proper functioning. Therefore, it is essential to know the exact spatial information of the nuclei and cells to understand their functions and interactions. To overcome these limitations, geometric models and software such as Motion Tracking (MT) are used to process liver images obtained through confocal fluorescence microscopy, generating a highresolution 3D reconstruction of all important cellular and subcellular components in the tissue. A key step is the segmentation of cell nuclei, which faces challenges in very close nuclei. Convolutional neural networks (CNNs) are particularly effective for processing and analyzing visual data. The Residual Attention U-Net 3D is an advanced variant of CNN designed for the segmentation of volumetric images, combining the U-Net architecture with residual blocks and attention mechanisms to improve segmentation accuracy. In this study, a model based on the mentioned network was trained, for which a high-quality 3D nuclei database in fluorescence microscopy images of the hepatic lobule was generated. Additionally, a multitask approach was chosen to train this neural network, performing binary segmentation of hepatic nuclei and predicting a vector field to estimate the centroids of each nucleus and generate instance segmentation [1, 2, 3, 4, 5]. For the database creation, approximately 2900 3D nuclei were segmented in three control subjects (C1, C2, and C3) in the periportal, central, and pericentral regions of the hepatic lobule. Subjects C1 and C2 were used for training, while subject C3 was used for evaluation. In addition to the MT segmentation method, Cellpose [3] was used to compare the results. Our proposed model showed an average precision improvement of 17 %, 28 %, and 33% compared to MT in the central, periportal, and pericentral regions, respectively. It also outperformed Cellpose by 5 %, 17 %, and 7% in the same regions, respectively. Notably, in the periportal region, our model achieved significantly better results, demonstrating its ability to handle segmentation in complex regions of hepatic tissue.
The understanding of organ organization and cellular coordination to form functional tissues is fundamental in cell and developmental biology. Tissue histology is crucial for the diagnosis of various pathologies, including cancer. However, traditional two-dimensional (2D) histopathological evaluations result in a loss of three dimensional (3D) context of the tissue. The liver, a vital organ with multiple functions, relies on a complex 3D architecture for its proper functioning. Therefore, it is essential to know the exact spatial information of the nuclei and cells to understand their functions and interactions. To overcome these limitations, geometric models and software such as Motion Tracking (MT) are used to process liver images obtained through confocal fluorescence microscopy, generating a highresolution 3D reconstruction of all important cellular and subcellular components in the tissue. A key step is the segmentation of cell nuclei, which faces challenges in very close nuclei. Convolutional neural networks (CNNs) are particularly effective for processing and analyzing visual data. The Residual Attention U-Net 3D is an advanced variant of CNN designed for the segmentation of volumetric images, combining the U-Net architecture with residual blocks and attention mechanisms to improve segmentation accuracy. In this study, a model based on the mentioned network was trained, for which a high-quality 3D nuclei database in fluorescence microscopy images of the hepatic lobule was generated. Additionally, a multitask approach was chosen to train this neural network, performing binary segmentation of hepatic nuclei and predicting a vector field to estimate the centroids of each nucleus and generate instance segmentation [1, 2, 3, 4, 5]. For the database creation, approximately 2900 3D nuclei were segmented in three control subjects (C1, C2, and C3) in the periportal, central, and pericentral regions of the hepatic lobule. Subjects C1 and C2 were used for training, while subject C3 was used for evaluation. In addition to the MT segmentation method, Cellpose [3] was used to compare the results. Our proposed model showed an average precision improvement of 17 %, 28 %, and 33% compared to MT in the central, periportal, and pericentral regions, respectively. It also outperformed Cellpose by 5 %, 17 %, and 7% in the same regions, respectively. Notably, in the periportal region, our model achieved significantly better results, demonstrating its ability to handle segmentation in complex regions of hepatic tissue.
Description
Tesis presentada para optar al título de Ingeniero Civil Biomédico
Keywords
Administración de bases de datos, Algoritmos, Imagen tridimensional, Hígado