Resumen de la tesis que presenta Alexis Crespo Michel como requisito parcial para la obtención
del grado de Doctor en Ciencias en Electrónica y Telecomunicaciones con orientación en Telecomunicaciones
Clasificación automática de imágenes para la identificación de hongos de madera de la vid
Resumen aprobado por:
|
Dr. Miguel Ángel Alonso Arévalo
Director de tesis
|
Resumen en español
Los hongos patógenos asociados a las enfermedades de la madera de la vid (GTD) suponen una grave amenaza para la viticultura mundial, ya que causan importantes pérdidas económicas y resaltan la necesidad de contar con métodos de diagnóstico precisos y accesibles. Esta tesis explora la aplicación del machine learning y el deep learning al análisis de imágenes de microscopia de esporas de hongo con el fin de desarrollar alternativas eficaces y económicas a la identificación molecular. Se adquirió y analizó un conjunto de datos de 2716 imágenes microscópicas que representan 11 especies de 4 géneros, con especial atención a Lasiodiplodia. Inicialmente, se extrajeron características morfológicas, de intensidad de píxeles y de inspiración biológica de esporas segmentadas manualmente y se evaluaron utilizando modelos basados en árboles, lo que reveló el valor discriminatorio de los rasgos morfológicos, pero también su capacidad limitada para captar diferencias sutiles entre especies. Para abordar estas limitaciones, se implementaron enfoques de aprendizaje profundo, incluyendo redes neuronales convolucionales (CNN) y transformadores de visión (ViT). Entre los modelos evaluados, una arquitectura basada en ResNet alcanzó una precisión de validación del 94.01%, lo que confirma la ventaja del aprendizaje de características de end-to-end sobre las representaciones manuales, mientras que los análisis de embeddings proporcionaron información útil sobre las relaciones taxonómicas. Se exploraron más a fondo las estrategias de aprendizaje auto-supervisado, incluyendo autoencoders y SimCLR, para explotar los datos sin etiquetar, lo que dio lugar a mejores representaciones de características y un rendimiento de clasificación mejorado en condiciones de anotación limitadas. Además, se desarrolló un modelo de segmentación de instancias basado en Mask R-CNN para permitir la detección automática de los contornos de las esporas, logrando máscaras de segmentación precisas cuyas características derivadas mostraron solo una diferencia de rendimiento del 1-2.5% en comparación con las de las anotaciones de referencia. Estos métodos se integraron en un prototipo de diagnóstico portátil que combina MobileNetV2 para la clasificación de especies y Mask R-CNN para el análisis morfológico, implementado en una Raspberry Pi 5 con una interfaz basada en PyQt5 para la adquisición, inferencia y visualización de imágenes en tiempo real. En general, este trabajo establece un marco completo y escalable para la identificación y caracterización automatizada de esporas de hongos, lo que demuestra el potencial de la visión por computadora y el aprendizaje auto-supervisado para avanzar en herramientas de diagnóstico de bajo costo para los patógenos GTD.
Palabras clave: Clasificación de esporas de hongo, aprendizaje profundo, aprendizaje auto-supervisado, segmentación de instancias, sistemas de diagnóstico portátiles.
Resumen en inglés
Fungal pathogens associated with grapevine trunk diseases (GTDs) pose a major threat to global viticulture, causing severe economic losses and highlighting the need for accurate and accessible diagnostic methods. This thesis explores the application of machine learning and deep learning for microscopic image analysis of fungal spores to develop efficient, low-cost alternatives to molecular identification. A dataset of 2,716 microscope images representing 11 species across 4 genera, with a strong focus on Lasiodiplodia, was acquired and analyzed. Initially, handcrafted morphological, pixel intensity, and biologically inspired features were extracted from manually segmented spores and evaluated using tree-based models, revealing the discriminative value of morphological traits but limited capacity to capture subtle interspecies differences. To address these limitations, deep learning approaches were implemented, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Among the evaluated models, a ResNet-based architecture achieved a validation accuracy of 94.01%, confirming the advantage of end-to-end feature learning over handcrafted representations, while embedding analyses provided useful insights into taxonomic relationships. Self-supervised learning strategies, including autoencoders and SimCLR, were further explored to exploit unlabeled data, yielding improved feature representations and enhanced classification performance under limited annotation conditions. Additionally, an instance segmentation model based on Mask R-CNN was developed to enable automatic spore boundary detection, achieving precise segmentation masks whose derived features showed only a 1–2.5% performance difference compared to those from ground-truth annotations. These methods were integrated into a portable diagnostic prototype combining MobileNetV2 for species classification and Mask R-CNN for morphological analysis, deployed on a Raspberry Pi 5 with a PyQt5-based interface for real-time image acquisition, inference, and visualization. Overall, this work establishes a comprehensive, scalable framework for automated fungal identification and characterization, demonstrating the potential of computer vision and self-supervised learning to advance affordable diagnostic tools for GTD pathogens.
Fungal pathogens associated with grapevine trunk diseases (GTDs) pose a major threat to global viticulture, causing severe economic losses and highlighting the need for accurate and accessible diagnostic methods. This thesis explores the application of machine learning and deep learning for microscopic image analysis of fungal spores to develop efficient, low-cost alternatives to molecular identification. A dataset of 2,716 microscope images representing 11 species across 4 genera, with a strong focus on Lasiodiplodia, was acquired and analyzed. Initially, handcrafted morphological, pixel intensity, and biologically inspired features were extracted from manually segmented spores and evaluated using tree-based models, revealing the discriminative value of morphological traits but limited capacity to capture subtle interspecies differences. To address these limitations, deep learning approaches were implemented, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Among the evaluated models, a ResNet-based architecture achieved a validation accuracy of 94.01%, confirming the advantage of end-to-end feature learning over handcrafted representations, while embedding analyses provided useful insights into taxonomic relationships. Self-supervised learning strategies, including autoencoders and SimCLR, were further explored to exploit unlabeled data, yielding improved feature representations and enhanced classification performance under limited annotation conditions. Additionally, an instance segmentation model based on Mask R-CNN was developed to enable automatic spore boundary detection, achieving precise segmentation masks whose derived features showed only a 1–2.5% performance difference compared to those from ground-truth annotations. These methods were integrated into a portable diagnostic prototype combining MobileNetV2 for species classification and Mask R-CNN for morphological analysis, deployed on a Raspberry Pi 5 with a PyQt5-based interface for real-time image acquisition, inference, and visualization. Overall, this work establishes a comprehensive, scalable framework for automated fungal identification and characterization, demonstrating the potential of computer vision and self-supervised learning to advance affordable diagnostic tools for GTD pathogens.
Palabras clave: Fungal spore classification, Deep learning, Self-supervised learning, Instance segmentation, Portable diagnostic systems