Use this link to cite:
https://hdl.handle.net/2183/47694 Análisis automático de cambios en la morfología vascular en imágenes OCT-A asociados a terapias hiperbáricas mediante Deep Learning
Loading...
Identifiers
Publication date
Authors
Varela Martínez, Ainé
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: La oxigenoterapia hiperbárica (HBOT) es un tratamiento médico consolidado que somete al paciente a presiones elevadas de oxígeno puro para favorecer la regeneración tisular. Sin embargo, la exposición a la hiperoxia sistémica plantea interrogantes sobre su impacto fisiológico en tejidos sensibles como la retina, donde puede inducir mecanismos de vasoconstricción o estrés oxidativo. En la práctica clínica actual, la evaluación de estos efectos mediante Angiografía por Tomografía de Coherencia Óptica (OCT-A) depende de análisis manuales subjetivos, tediosos y propensos a la variabilidad inter-observador, lo que dificulta un seguimiento preciso del paciente. Este proyecto presenta el desarrollo e implementación de un sistema automático basado en Deep Learning para la segmentación y cuantificación objetiva de los cambios vasculares retinianos asociados a la HBOT. Metodológicamente, se empleó la arquitectura auto-adaptativa nnU-Net para entrenar modelos de segmentación semántica capaces de discriminar jerarquías vasculares (vasos mayores y capilares) y la Zona Avascular Foveal (ZAF). Para superar la escasez de datos etiquetados, se diseñó una estrategia de generación de Ground Truth mediante etiquetado asistido: se utilizaron modelos pre-entrenados con el dataset público OCTA-500 para inferir pre-etiquetas sobre una cohorte de 29 pacientes, las cuales fueron posteriormente refinadas manualmente. La robustez del sistema se aseguró mediante una validación cruzada de 5 pliegues (5-Fold Cross-Validation). Los resultados validan técnicamente la propuesta, alcanzando un coeficiente Dice global de 0.8771 en la segmentación integral de las estructuras retinianas. Desde la perspectiva clínica, el análisis estadístico longitudinal (mediante la prueba de Wilcoxon) reveló que las alteraciones representativas se manifiestan principalmente en la visita de seguimiento tardía. Se detectó una disminución estadísticamente significativa en la densidad de los vasos mayores (p = 0.0477) y un aumento del área de la ZAF (p = 0.0247), confirmando la teoría de la vasoconstricción reactiva como mecanismo de defensa. No obstante, la densidad capilar (p = 0.7282) y la circularidad de la ZAF (p = 0.1502) se mantuvieron estables, sugiriendo que la microcirculación y la geometría foveal se preservan, lo que respalda la seguridad del tratamiento. Esta herramienta demuestra la viabilidad de integrar la inteligencia artificial en el flujo hospitalario para objetivar biomarcadores y asistir en la toma de decisiones clínicas.
[Abstract]: Hyperbaric Oxygen Therapy (HBOT) is a well-established medical treatment that exposes patients to high pressures of pure oxygen to promote tissue regeneration. However, exposure to systemic hyperoxia raises concerns regarding its physiological impact on sensitive tissues such as the retina, where it may induce vasoconstriction mechanisms or oxidative stress. Current clinical evaluation of these effects using Optical Coherence Tomography Angiography (OCT-A) relies on subjective, time-consuming manual analyses prone to inter-observer variability, hindering precise patient monitoring. This project presents the development and implementation of an automated Deep Learning based system for the objective segmentation and quantification of retinal vascular changes associated with HBOT. Methodologically, the self-configuring nnU-Net architecture was employed to train semantic segmentation models capable of discriminating vascular hierarchies (major vessels and capillaries) and the Foveal Avascular Zone (ZAF). To address the scarcity of labeled data, a Ground Truth generation strategy based on assisted labeling was designed: models pre-trained on the public OCTA-500 dataset were used to infer pre-labels on a cohort of 29 patients, which were subsequently refined manually. System robustness was ensured through a 5-Fold Cross-Validation scheme. The results technically validate the proposal, achieving a Dice coefficient exceeding 0.87 in the integral segmentation of retinal structures. From a clinical perspective, longitudinal statistical analysis (using the Wilcoxon signed-rank test) revealed that representative alterations manifest primarily in the late follow-up visit. A statistically significant decrease in major vessel density (p = 0.0477) and an increase in the ZAF area (p = 0.0247) were detected, confirming the theory of reactive vasoconstriction as a defense mechanism. However, capillary density (p = 0.7282) and ZAF circularity (p = 0.1502) remained stable, suggesting that microcirculation and foveal geometry are preserved, thus supporting the safety of the treatment. This tool demonstrates the feasibility of integrating artificial intelligence into hospital workflows to objective biomarkers and assist in clinical decision-making.
[Abstract]: Hyperbaric Oxygen Therapy (HBOT) is a well-established medical treatment that exposes patients to high pressures of pure oxygen to promote tissue regeneration. However, exposure to systemic hyperoxia raises concerns regarding its physiological impact on sensitive tissues such as the retina, where it may induce vasoconstriction mechanisms or oxidative stress. Current clinical evaluation of these effects using Optical Coherence Tomography Angiography (OCT-A) relies on subjective, time-consuming manual analyses prone to inter-observer variability, hindering precise patient monitoring. This project presents the development and implementation of an automated Deep Learning based system for the objective segmentation and quantification of retinal vascular changes associated with HBOT. Methodologically, the self-configuring nnU-Net architecture was employed to train semantic segmentation models capable of discriminating vascular hierarchies (major vessels and capillaries) and the Foveal Avascular Zone (ZAF). To address the scarcity of labeled data, a Ground Truth generation strategy based on assisted labeling was designed: models pre-trained on the public OCTA-500 dataset were used to infer pre-labels on a cohort of 29 patients, which were subsequently refined manually. System robustness was ensured through a 5-Fold Cross-Validation scheme. The results technically validate the proposal, achieving a Dice coefficient exceeding 0.87 in the integral segmentation of retinal structures. From a clinical perspective, longitudinal statistical analysis (using the Wilcoxon signed-rank test) revealed that representative alterations manifest primarily in the late follow-up visit. A statistically significant decrease in major vessel density (p = 0.0477) and an increase in the ZAF area (p = 0.0247) were detected, confirming the theory of reactive vasoconstriction as a defense mechanism. However, capillary density (p = 0.7282) and ZAF circularity (p = 0.1502) remained stable, suggesting that microcirculation and foveal geometry are preserved, thus supporting the safety of the treatment. This tool demonstrates the feasibility of integrating artificial intelligence into hospital workflows to objective biomarkers and assist in clinical decision-making.
Description
Keywords
Aprendizaje profundo Angiografía por Tomografía de Coherencia Óptica (Angio-OCT) Oxigenoterapia Hiperbárica (OHB) Retina Biomarcadores Vasculares Inteligencia Artificial Segmentación Automática Zona Avascular Foveal (ZAF) Deep Learning Optical Coherence Tomography Angiography (OCT-A) Hyperbaric Oxygen Therapy (HBOT) Vascular Biomarkers Artificial Intelligence Automatic Segmentation Foveal Avascular Zone (FAZ)
Editor version
Rights
Attribution 4.0 International








