Diseño e implementación de un sistema para la detección de neumonía basado en pulsioximetría y algoritmos de Deep Learning aplicados al análisis de radiografías
No Thumbnail Available
Date
2024-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Nacional de Trujillo
Abstract
En el presente proyecto de ingeniería se diseñó un sistema para la detección de
neumonía en niños menores de 5 años basándose en la medición de saturación de oxígeno
en la sangre y del procesamiento de imágenes radiográficas por medio de algoritmos de
Deep Learning. Al comparar diversos algoritmos a partir de los antecedentes consultados,
se identificaron a las arquitecturas MobileNet y VGG19 como las más aptas para esta
aplicación. Luego, se identificó la base de datos pública del Centro Médico de Mujeres y
Niños de Guangzhou, la cual sirvió para realizar el entrenamiento, la validación y la
prueba de cada uno de los modelos propuestos. Para iniciar con el desarrollo, se
redimensionó las imágenes a 224x224 píxeles y se mejoró el contraste de estas por medio
de la ecualización de histograma y se aplicó el método de aumento de datos para
generalizar la información de entrada al modelo de Deep Learning. Es así como, se
realizaron diferentes comparaciones para determinar qué modelo propuesto contaba con
las mejores prestaciones. Gracias a ello se pudo identificar mejores resultados al no
utilizar pesos pre entrenados para el modelo, utilizar el algoritmo de optimización SGD,
hacer uso de la arquitectura MobileNet y valerse de capas de Dropout a 50% entre cada
una de las capas densas añadidas al modelo. El modelo realizado fue entrenado por 20
épocas y obtuvo 90.66 % y 95.48 % para los valores de exactitud y sensibilidad,
respectivamente. Finalmente, se construyó un sistema de pulsioximetría con errores bajo
1% y una desviación estándar de 0.688%, que, integrado en una interfaz gráfica de usuario
(GUI) con el modelo de Deep Learning, disminuyó los falsos positivos y casos no
detectados, alcanzando una exactitud del sistema integrado del 97.00%.
At this engineering project, a system for the detection of pneumonia in children under 5 years was designed based on the measurement of oxygen saturation in the blood and the processing of chest x-ray images through Deep Learning algorithms. By comparing different algorithms based on the information consulted, the MobileNet and VGG19 architectures was identified as the most suitable for this application. Then, the public database of the Guangzhou Women's and Children's Medical Centre was found, which served to carry out the training, validation and testing of each of the proposed models. To start with the development, the images were resized to 224x224 pixels, and their contrast was improved by histogram equalization and the data augmentation method was applied to generalize the input information to the Deep Learning model. Therefore, different comparisons were made to decide which proposed model had the best performance. Thanks to this, it was possible to find better results by not using pre-trained weights for the model, using the SGD optimization algorithm, making use of the MobileNet architecture, and using Dropout layers at 50 % between each of the dense layers added to the model. The model performed was trained for 20 epochs and obtained 90.66 % and 95.48 % for accuracy and recall values, respectively. Finally, a pulse oximetry system was built with errors below 1% and a standard deviation of 0.688%, which, when combined with a user graphical interface (GUI) and the Deep Learning model, decreased the number of false positives, and missed cases, achieving an accuracy of 97.00% for the integrated system.
At this engineering project, a system for the detection of pneumonia in children under 5 years was designed based on the measurement of oxygen saturation in the blood and the processing of chest x-ray images through Deep Learning algorithms. By comparing different algorithms based on the information consulted, the MobileNet and VGG19 architectures was identified as the most suitable for this application. Then, the public database of the Guangzhou Women's and Children's Medical Centre was found, which served to carry out the training, validation and testing of each of the proposed models. To start with the development, the images were resized to 224x224 pixels, and their contrast was improved by histogram equalization and the data augmentation method was applied to generalize the input information to the Deep Learning model. Therefore, different comparisons were made to decide which proposed model had the best performance. Thanks to this, it was possible to find better results by not using pre-trained weights for the model, using the SGD optimization algorithm, making use of the MobileNet architecture, and using Dropout layers at 50 % between each of the dense layers added to the model. The model performed was trained for 20 epochs and obtained 90.66 % and 95.48 % for accuracy and recall values, respectively. Finally, a pulse oximetry system was built with errors below 1% and a standard deviation of 0.688%, which, when combined with a user graphical interface (GUI) and the Deep Learning model, decreased the number of false positives, and missed cases, achieving an accuracy of 97.00% for the integrated system.
Description
Keywords
TECHNOLOGY::Engineering physics::Other engineering physics