Reconstrucción de imágenes a partir de mediciones de tomografía por impedancia eléctrica usando algoritmos de Deep Learning.
No Thumbnail Available
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Las enfermedades respiratorias representan una carga crítica de salud pública mundial, lo que motiva la búsqueda de herramientas de monitoreo que superen las limitaciones de costo y accesibilidad de los estándares clínicos actuales. En este contexto, la Tomografía por Impedancia Eléctrica (TIE) surge como una alternativa funcional y no ionizante; sin embargo, los sistemas comerciales existentes suelen ser voluminosos y de alto costo. Por ello, existe una necesidad de desarrollar soluciones portátiles y económicas capaces de resolver con precisión el problema inverso de la TIE, el cual es intrínsecamente no lineal, mal condicionado y sensible al ruido. El presente trabajo de tesis aborda este desafío mediante la reconstrucción de imágenes con algoritmos de deep learning aplicados a un dispositivo portátil diseñado en laboratorio, utilizando un fantoma pulmonar para su validación experimental. La metodología se inició con el diseño y construcción de un fantoma físico equipado con 16 electrodos de cobre para un sistema de adquisición TIE portátil y de bajo costo. Dada la alta demanda temporal que implica la toma de muestras experimentales masivas para el entrenamiento de modelos de inteligencia artificial, se implementó un gemelo digital en COMSOL Multiphysics. Este permitió generar un dataset sintético robusto de 7.992 configuraciones distintas, proporcionando la base necesaria para que las redes aprendieran la física del problema. Posteriormente, se evaluaron tres arquitecturas de deep learning: una Red Neuronal Convolucional (CNN), una U-Net y un modelo KAN-ViT. Para cerrar la brecha entre la simulación y la realidad, se aplicó una estrategia de fine-tuning utilizando 7.115 datos experimentales recolectados en el fantoma físico a partir de 150 configuraciones distintas, adaptando así los modelos al ruido electrónico y las no-idealidades del hardware real. Los resultados obtenidos validan la hipótesis de investigación al demostrar que es posible realizar reconstrucciones precisas y de alta calidad estructural en tiempo real. La arquitectura U-Net destacó como la más robusta, alcanzando un índice de similitud estructural (SSIM) de 0.9453 y un coeficiente de correlación de 0.8785 en datos reales, resultados que superan a los métodos clásicos. El sistema integrado logró una tasa de procesamiento de 1.94 cuadros por segundo, con una latencia de inferencia de apenas 15 ms. Este trabajo no solo aporta una metodología rigurosa para compensar las limitaciones del hardware portátil mediante inteligencia artificial, sino que demuestra que la combinación de un fantoma físico y un gemelo digital permite obtener imágenes de alta fidelidad a través de estrategias de transfer learning. Estos hallazgos sientan las bases tecnológicas para el futuro desarrollo de herramientas de monitoreo clínico que sean verdaderamente portátiles, seguras y accesibles.
Respiratory diseases represent a critical global public health burden, motivating the search for monitoring tools that overcome the cost and accessibility limitations of current clinical standards. In this context, Electrical Impedance Tomography (EIT) emerges as a functional and non-ionizing alternative; however, existing commercial systems are typically bulky and expensive. Therefore, there is a need to develop portable and affordable solutions capable of accurately solving the EIT inverse problem, which is inherently non-linear, ill-posed, and sensitive to noise. This thesis addresses this challenge through image reconstruction using deep learning algorithms applied to a lab-designed portable device, utilizing a lung phantom for experimental validation. The methodology began with the design and construction of a physical phantom equipped with 16 copper electrodes for a portable, low-cost EIT acquisition system. Given the high temporal demand involved in collecting massive experimental samples for training artificial intelligence models, a digital twin was implemented in COMSOL Multiphysics. This allowed for the generation of a robust synthetic dataset of 7,992 distinct configurations, providing the necessary foundation for the networks to learn the physics of the problem. Subsequently, three architectures were evaluated: a Convolutional Neural Network (CNN), a U-Net, and a KAN-ViT model. To bridge the gap between simulation and reality, a fine-tuning strategy was applied using 7,115 experimental data points collected from the physical phantom across 150 distinct configurations, thus adapting the models to electronic noise and real-world hardware nonidealities. The results validate the research hypothesis by demonstrating that accurate, high-structural-quality real-time reconstructions are possible. The U-Net architecture stood out as the most robust, achieving a Structural Similarity Index (SSIM) of 0.9453 and a correlation coefficient of 0.8785 on real data results that outperform classical methods. The integrated system achieved a processing rate of 1.94 frames per second, with an inference latency of just 15 ms. This work not only contributes a rigorous methodology for compensating the limitations of portable hardware through artificial intelligence but also demonstrates that the combination of a physical phantom and a digital twin enables high-fidelity imaging through transfer learning strategies. These findings lay the technological foundations for the future development of clinical monitoring tools that are truly portable, safe, and accessible.
Respiratory diseases represent a critical global public health burden, motivating the search for monitoring tools that overcome the cost and accessibility limitations of current clinical standards. In this context, Electrical Impedance Tomography (EIT) emerges as a functional and non-ionizing alternative; however, existing commercial systems are typically bulky and expensive. Therefore, there is a need to develop portable and affordable solutions capable of accurately solving the EIT inverse problem, which is inherently non-linear, ill-posed, and sensitive to noise. This thesis addresses this challenge through image reconstruction using deep learning algorithms applied to a lab-designed portable device, utilizing a lung phantom for experimental validation. The methodology began with the design and construction of a physical phantom equipped with 16 copper electrodes for a portable, low-cost EIT acquisition system. Given the high temporal demand involved in collecting massive experimental samples for training artificial intelligence models, a digital twin was implemented in COMSOL Multiphysics. This allowed for the generation of a robust synthetic dataset of 7,992 distinct configurations, providing the necessary foundation for the networks to learn the physics of the problem. Subsequently, three architectures were evaluated: a Convolutional Neural Network (CNN), a U-Net, and a KAN-ViT model. To bridge the gap between simulation and reality, a fine-tuning strategy was applied using 7,115 experimental data points collected from the physical phantom across 150 distinct configurations, thus adapting the models to electronic noise and real-world hardware nonidealities. The results validate the research hypothesis by demonstrating that accurate, high-structural-quality real-time reconstructions are possible. The U-Net architecture stood out as the most robust, achieving a Structural Similarity Index (SSIM) of 0.9453 and a correlation coefficient of 0.8785 on real data results that outperform classical methods. The integrated system achieved a processing rate of 1.94 frames per second, with an inference latency of just 15 ms. This work not only contributes a rigorous methodology for compensating the limitations of portable hardware through artificial intelligence but also demonstrates that the combination of a physical phantom and a digital twin enables high-fidelity imaging through transfer learning strategies. These findings lay the technological foundations for the future development of clinical monitoring tools that are truly portable, safe, and accessible.
Description
Tesis presentada para optar al grado de Magíster en Ciencias de la Ingeniería con mención en Ingeniería Eléctrica.
Keywords
Tomógrafos, Imágenes digitales, Enfermedades respiratorias, Deep learning