Use this link to cite:
https://hdl.handle.net/2183/45751 Estudio de funciones de pérdida centradas en la estructura local para el entrenamiento de redes de neuronas en tareas de regresión de imagen
Loading...
Identifiers
Publication date
Authors
Iglesias Sánchez, Aldara
Other responsabilities
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: En el ámbito oftalmológico es común utilizar imágenes de fondo de ojo para apoyar el diagnóstico clínico. El análisis de estas imágenes permite detectar y monitorear enfermedades oculares y sistémicas que se manifiestan en forma de lesiones o alteraciones anatómicas en el fondo de ojo. Con el creciente volumen de imágenes generadas en clínicas y hospitales, y la necesidad de diagnósticos rápidos y precisos, se ha vuelto cada vez más importante aplicar métodos automáticos basados en inteligencia artificial para analizar imágenes del fondo de ojo. Estas técnicas se orientan principalmente a tareas de detección y segmentación precisa de las estructuras anatómicas y patológicas. Sin embargo, uno de los principales desafíos en este ámbito es la escasez de imágenes etiquetadas, especialmente cuando se requiere una delineación precisa de las estructuras, como es el caso en tareas de segmentación y detección. Esto ha motivado el desarrollo de tareas que puedan entrenarse sin etiquetas explícitas, pero que generen representaciones útiles para tareas clínicas posteriores mediante aprendizaje transferido. Aunque muchas tareas de preentrenamiento se enfocan en la interpretación global de la imagen, recientemente han ganado interés los modelos que se centran en información local. Estos modelos se centran en tareas de predicción densa (píxel a píxel) y ofrecen representaciones más adecuadas para segmentación y detección. Un ejemplo es la reconstrucción de imágenes de angiografía a partir de fotografías de fondo de ojo, lo cual implica generar imágenes sintéticas mediante el reconocimiento de patrones vasculares que indican por dónde circularía el marcador de fluoresceína inyectado, y predecir su apariencia de forma densa. Para este tipo de tareas de reconstrucción y síntesis de imagen, es habitual utilizar funciones de pérdida de regresión punto a punto. Sin embargo, estudios previos muestran que incluir información del contexto local y estructura de la imagen a la hora de entrenar - como en el caso de SSIM (Structural Similarity Index Measure) - mejora los resultados de reconstrucción. Este trabajo se centrará en desarrollar y analizar nuevas funciones de pérdida basadas en la estructura local de la imagen, con el objetivo de evaluar si su codificación y uso mejora la calidad de reconstrucción y si dichas mejoras pueden trasladarse a otras tareas mediante aprendizaje transferido.
[Abstract]: In the field of ophthalmology, fundus images are commonly used to support clinical diagnosis. The analysis of these images enables the detection and monitoring of ocular and systemic diseases that manifest as lesions or anatomical abnormalities in the fundus. With the increasing volume of images generated in clinics and hospitals, and the need for fast and accurate diagnoses, the application of automated methods based on artificial intelligence has become increasingly important for the analysis of fundus images. These techniques are primarily aimed at the accurate detection and segmentation of anatomical and pathological structures. However, one of the main challenges in this domain is the scarcity of labeled images, particularly when precise delineation of structures is required, as is the case in segmentation and detection tasks. This limitation has motivated the development of tasks that can be trained without explicit labels, but that generate useful representations for downstream clinical tasks through transfer learning. While many pretraining tasks focus on the global interpretation of the image, models that emphasize local information have recently attracted increasing interest. These models target dense prediction tasks (pixel-wise) and provide more suitable representations for segmentation and detection. One example is the reconstruction of angiography images from fundus photographs, which involves generating synthetic images by recognizing vascular patterns that indicate where the injected fluorescein marker would travel, and predicting its appearance in a dense manner. For this type of image reconstruction and synthesis task, point-wise regression loss functions are commonly used. However, previous studies have shown that incorporating local context and structural information into training—such as through the Structural Similarity Index Measure (SSIM)—can enhance reconstruction performance. This work will focus on the development and analysis of novel loss functions based on the local structure of the image, with the aim of evaluating whether their encoding and application improve reconstruction quality, and whether such improvements can be transferred to other tasks via transfer learning.
[Abstract]: In the field of ophthalmology, fundus images are commonly used to support clinical diagnosis. The analysis of these images enables the detection and monitoring of ocular and systemic diseases that manifest as lesions or anatomical abnormalities in the fundus. With the increasing volume of images generated in clinics and hospitals, and the need for fast and accurate diagnoses, the application of automated methods based on artificial intelligence has become increasingly important for the analysis of fundus images. These techniques are primarily aimed at the accurate detection and segmentation of anatomical and pathological structures. However, one of the main challenges in this domain is the scarcity of labeled images, particularly when precise delineation of structures is required, as is the case in segmentation and detection tasks. This limitation has motivated the development of tasks that can be trained without explicit labels, but that generate useful representations for downstream clinical tasks through transfer learning. While many pretraining tasks focus on the global interpretation of the image, models that emphasize local information have recently attracted increasing interest. These models target dense prediction tasks (pixel-wise) and provide more suitable representations for segmentation and detection. One example is the reconstruction of angiography images from fundus photographs, which involves generating synthetic images by recognizing vascular patterns that indicate where the injected fluorescein marker would travel, and predicting its appearance in a dense manner. For this type of image reconstruction and synthesis task, point-wise regression loss functions are commonly used. However, previous studies have shown that incorporating local context and structural information into training—such as through the Structural Similarity Index Measure (SSIM)—can enhance reconstruction performance. This work will focus on the development and analysis of novel loss functions based on the local structure of the image, with the aim of evaluating whether their encoding and application improve reconstruction quality, and whether such improvements can be transferred to other tasks via transfer learning.
Description
Editor version
Rights
Attribution 4.0 International








