Clasificación de núcleos celulares en 3D basada en deep learning para la reconstrucción digital de tejido hepático.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
El hígado, crucial para la digestión y el sistema circulatorio, contiene diversos tipos de células con variaciones morfológicas y nucleares. Comprender su funcionamiento a nivel celular es esencial para el tratamiento de enfermedades hepáticas, por lo que es necesario contar con un sistema que permita diferenciar estos tipos celulares.
A pesar de los desafíos recientes, la informática ha permitido avances significativos en biología celular, como la creación de modelos 3D del lobulillo hepático a partir de imágenes de microscopía confocal. Aunque estos modelos 3D son prometedores, presentan limitaciones como la dependencia de múltiples tipos de imágenes para una clasificación precisa y la clasificación errónea debido a las limitaciones de los algoritmos actuales. En este contexto, una estrategia prometedora es el uso de Deep Learning (DL) con Redes Neuronales Convolucionales (CNN), que ha demostrado éxito en clasificación en biología e imágenes médicas. El DL puede superar las limitaciones actuales al reducir la dependencia de múltiples imágenes y al identificar patrones complejos para una clasificación más precisa.
El objetivo principal de esta memoria es utilizar Deep Learning (DL) para clasificar diferentes tipos de núcleos hepáticos en 3D mediante el entrenamiento de Redes Neuronales Convolucionales (CNN). Se creó un Ground Truth (GT) en 3D a partir de imágenes de microscopía confocal y se entrenaron tres CNN reconocidas adaptadas a 3D (ResNet50, InceptionV3 e Inception-ResNet-v2). Se evaluaron distintos prototipos de clasificación, tanto multicategoría como binarios, usando métricas como F1-Score, Recall, Accuracy y Precision para seleccionar el mejor modelo.
Como resultado, se obtuvo un algoritmo robusto para clasificar diferentes tipos de núcleos hepáticos, con resultados comparables o incluso superiores a los del software MotionTracking (MT). Se desarrollaron modelos binarios específicos para distintos tipos de núcleos hepáticos, destacando el modelo para hepatocitos, que logró una correcta clasificación del 95 % de estos núcleos tanto en tejido hepático normal como en tejido en regeneración.
En resumen, la investigación subraya la importancia de mejorar la reconstrucción en 3D del tejido hepático con menos inmunomarcadores. Esto permite una clasificación más precisa de los núcleos hepáticos, facilita el estudio de otras estructuras biológicas en el hígado, optimiza recursos y tiempos de laboratorio, y reduce la dependencia de marcadores específicos, promoviendo investigaciones más detalladas sobre las complejidades del tejido hepático.
The liver, crucial for digestion and the circulatory system, contains various types of cells with morphological and nuclear variations. Understanding its cellular function is essential for treating liver diseases, thus a system that allows differentiation between these cell types is necessary. Despite recent challenges, computing has enabled significant advances in cellular biology, such as the creation of 3D models of the liver lobule from confocal microscopy images. Although these 3D models are promising, they have limitations, such as dependence on multiple types of images for accurate classification and misclassification due to current algorithm limitations. In this context, a promising strategy is the use of Deep Learning (DL) with Convolutional Neural Networks (CNN), which have shown success in classification in biology and medical imaging. DL can overcome current limitations by reducing the dependency on multiple images and identifying complex patterns for more accurate classification. The main objective of this thesis is to use Deep Learning (DL) to classify different types of liver nuclei in 3D by training Convolutional Neural Networks (CNNs). A 3D Ground Truth (GT) was created from confocal microscopy images, and three well-known CNNs adapted to 3D (ResNet50, InceptionV3, and Inception-ResNet-v2) were trained. Various classification prototypes, both multi-category and binary, were evaluated using metrics such as F1-Score, Recall, Accuracy, and Precision to select the best model. As a result, a robust algorithm was developed to classify different types of liver nuclei, with results comparable to or even exceeding those of the MotionTracking (MT) software. Specific binary models were developed for different types of liver nuclei, with the hepatocyte model achieving a correct classification rate of 95 % for these nuclei in both normal liver tissue and regenerating tissue. In summary, the research highlights the importance of improving 3D reconstruction of liver tissue with fewer immunomarkers. This allows for more accurate classification of liver nuclei, facilitates the study of other biological structures in the liver, optimizes laboratory resources and time, and reduces dependence on specific markers, promoting more detailed investigations into the complexities of liver tissue.
The liver, crucial for digestion and the circulatory system, contains various types of cells with morphological and nuclear variations. Understanding its cellular function is essential for treating liver diseases, thus a system that allows differentiation between these cell types is necessary. Despite recent challenges, computing has enabled significant advances in cellular biology, such as the creation of 3D models of the liver lobule from confocal microscopy images. Although these 3D models are promising, they have limitations, such as dependence on multiple types of images for accurate classification and misclassification due to current algorithm limitations. In this context, a promising strategy is the use of Deep Learning (DL) with Convolutional Neural Networks (CNN), which have shown success in classification in biology and medical imaging. DL can overcome current limitations by reducing the dependency on multiple images and identifying complex patterns for more accurate classification. The main objective of this thesis is to use Deep Learning (DL) to classify different types of liver nuclei in 3D by training Convolutional Neural Networks (CNNs). A 3D Ground Truth (GT) was created from confocal microscopy images, and three well-known CNNs adapted to 3D (ResNet50, InceptionV3, and Inception-ResNet-v2) were trained. Various classification prototypes, both multi-category and binary, were evaluated using metrics such as F1-Score, Recall, Accuracy, and Precision to select the best model. As a result, a robust algorithm was developed to classify different types of liver nuclei, with results comparable to or even exceeding those of the MotionTracking (MT) software. Specific binary models were developed for different types of liver nuclei, with the hepatocyte model achieving a correct classification rate of 95 % for these nuclei in both normal liver tissue and regenerating tissue. In summary, the research highlights the importance of improving 3D reconstruction of liver tissue with fewer immunomarkers. This allows for more accurate classification of liver nuclei, facilitates the study of other biological structures in the liver, optimizes laboratory resources and time, and reduces dependence on specific markers, promoting more detailed investigations into the complexities of liver tissue.
Description
Tesis presentada para optar al título de Ingeniero Civil Biomédico
Keywords
Núcleo celular, Hígado, Imágenes 3D