Publication: HERRAMIENTA DE RECUPERACIÓN DE IMÁGENES MÉDICAS BASADO EN CONTENIDO PARA EL SISTEMA DE SALUD CHILENO
Date
2021-06
Authors
MOLINA BARRA, GABRIEL ANDRES
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
En el área de la salud, las imágenes médicas han sido un recurso fundamental para el diagnostico oportuno de patologías, la investigación de enfermedades y herramienta de estudio para los futuros profesionales de la salud. Por esta misma causa, sistemas de búsqueda de imágenes se han desarrollado a lo largo de los años para ayudar a la recuperación de dicha información. Estos sistemas funcionan transformando una imagen médica en un vector multidimensional para así construir índices en bases de datos, generándose de esta forma buscadores rápidos y automáticos. Los sistemas mas tradicionales utilizan descriptores de una imagen para transformarla en un vector multidimensional, pero en la actualidad, esto es poco factible ya que la gran mayoría de las imágenes médicas no están correctamente etiquetadas y/o estudiadas. Esto ocurre debido a que el volumen de datos generados es tan grande que no existen suficientes expertos para estudiar todas las imágenes, resultando en que los sistemas tradicionales de búsquedas sean abrumados por el rápido crecimiento de datos y la falta de evaluación de expertos. Se han propuesto diferentes métodos para resolver el problema denominado Content Base Image Retrieval System (CBIRS), en el cual, se agrupan todos los sistema de búsqueda de imágenes basado en contenido, incluyendo búsqueda de reconocimiento facial, imágenes naturales y el tema central de esta tesis, las imágenes médicas; sien do estas: radiografías, tomografías computacionales, examen histológico, entre otros. Existen autores que han intentado mejorar estos sistemas por el área de desarrollo de software, incrementando la infraestructura, perfeccionando los métodos de indexación y actualizando sistemas basados en la tecnología más moderna. A pesar de que este esfuerzo es suficiente para hacer los sistemas de búsqueda más robustos al crecimiento actual de datos en la gran mayoría de problemas, no son lo suficientemente eficaces para transformar en sistemas suficientemente escalables para problemas más complejos como son las imágenes médicas. Por estas razón, otros autores han preferido el área del deep learning para mejorar los vectores asociados al sistema, obteniendo mejores resultados en recuperación de imágenes, obteniendo el estado del arte en muchos tipos de búsquedas de enfermedades, pero a pesar de que estos sistemas basados en redes neuronales son más modernos, sufren de la misma falencia de su contraparte tradicional, lo cual es la falta de datos etiquetados, llevando a estos sistemas a tener una gran dificultad a ser implementados en un ambiente de trabajo médico real. En este documento se propone un CBMIRS basado en deep learning, el cual estará compuesto por dos secciones de redes neuronales que en conjunto construirán un vector más robusto para el sistema y además, este nuevo CBMIRS podrá ser ocupado a pesar de que las imágenes no lleven etiqueta, permitiendo una mejor escalabilidad en un ambiente clínico real. Esta propuesta se basa en el uso de combinar lo aprendido por un modelo de segmentación y un modelo de clasificación, para transformar una imagen en un vector con mayor información que utilizando una única red neuronal, permitiendo que un sistema de búsqueda de imágenes médica tenga mejores resultados y escalabilidad. La evidencia experimental, muestra que comparado con otros métodos del estado del arte, en recuperación de imágenes médicas, la propuesta muestra una gran ventaja al buscar imágenes que contengan una cierta enfermedad en escenarios clínicos reales. Además, se demuestra la capacidad efectiva de transferencia de información median te la utilización de los embeddings resultantes del segmentador en el reentrenamiento del clasificador o mediante la concatenación de embeddings resultantes. Finalmente, se demostró que la utilización de un algoritmo de eliminación de near-duplicate tiene un efecto regularizante en la de búsqueda de imágenes, permitiendo que sistemas que utilizan modelos más simples sean igual de competitivos que sistemas que utilizan modelos más complejos.
In the health area, medical images have been a fundamental resource for the timely diagnosis of pathologies, the investigation of diseases and a study tool for future health professionals. For this same reason, image search systems have been developed over the years to aid in the retrieval of such information. These systems work by transforming a medical image into a multidimensional vector in order to build indexes in databases, thus generating fast and automatic search engines. The most traditional systems use descriptors of an image to transform it into a multidimensional vector, but at present, this is not very feasible since the vast majority of medical images are not correctly labeled and/or studied. This, because the volume of data generated is so great that there are not enough experts to study all the images, resulting in traditional search systems being overwhelmed by the rapid growth of data and the lack of expert evaluation. Different methods have been proposed to solve the problem called Content Base Image Retrieval System (CBIRS), in which all content-based image search systems are grouped, including facial recognition search, natural images and the central theme of this thesis, medical images; These being: X-rays, computational tomography, histological examination, among others. There are authors who have tried to improve these systems in the area of software development, increasing the infrastructure, perfecting indexing methods and updating systems based on the most modern technology. Although this effort is enough to make search systems more robust to the current growth of data in the vast majority of problems, these efforts are not effective enough to transform into systems scalable enough for more complex problems such as medical images. For these reasons, other authors have preferred the area of deep learning to improve the vectors associated with the system, obtaining better results in image retrieval, obtaining the state of the art in many types of disease searches, but despite the fact that these systems based In neural networks they are more modern, they suffer from the same shortcoming of their traditional counterpart, which is the lack of labeled data, leading these systems to have great difficulty in being implemented in a real medical work environment. In this document, a CBMIRS based on deep learning is proposed, which will be composed of two sections of neural networks that together will build a more robust vector for the system and also, this new CBMIRS can be occupied even though the images do not have label, allowing better scalability in a real clinical environment. This proposal is based on the use of combining what has been learned by a segmentation model and a classification model, to transform an image into a vector with more information than using a single neural network, allowing a medical image search system to have better results and scalability. The experimental evidence shows that compared to other state-of-the-art methods in medical image retrieval, the proposal shows a great advantage when looking for images that contain a certain disease in real clinical settings. Furthermore, the effective information transfer capacity is demonstrated by using the resulting embeddings of the segmenter in the retraining of the classifier or by concatenating the resulting embeddings. Finally, it was shown that the use of a near-duplicate elimination algorithm has a regularizing effect on image search, allowing systems that use simpler models to be just as competitive as systems that use more complex models.
In the health area, medical images have been a fundamental resource for the timely diagnosis of pathologies, the investigation of diseases and a study tool for future health professionals. For this same reason, image search systems have been developed over the years to aid in the retrieval of such information. These systems work by transforming a medical image into a multidimensional vector in order to build indexes in databases, thus generating fast and automatic search engines. The most traditional systems use descriptors of an image to transform it into a multidimensional vector, but at present, this is not very feasible since the vast majority of medical images are not correctly labeled and/or studied. This, because the volume of data generated is so great that there are not enough experts to study all the images, resulting in traditional search systems being overwhelmed by the rapid growth of data and the lack of expert evaluation. Different methods have been proposed to solve the problem called Content Base Image Retrieval System (CBIRS), in which all content-based image search systems are grouped, including facial recognition search, natural images and the central theme of this thesis, medical images; These being: X-rays, computational tomography, histological examination, among others. There are authors who have tried to improve these systems in the area of software development, increasing the infrastructure, perfecting indexing methods and updating systems based on the most modern technology. Although this effort is enough to make search systems more robust to the current growth of data in the vast majority of problems, these efforts are not effective enough to transform into systems scalable enough for more complex problems such as medical images. For these reasons, other authors have preferred the area of deep learning to improve the vectors associated with the system, obtaining better results in image retrieval, obtaining the state of the art in many types of disease searches, but despite the fact that these systems based In neural networks they are more modern, they suffer from the same shortcoming of their traditional counterpart, which is the lack of labeled data, leading these systems to have great difficulty in being implemented in a real medical work environment. In this document, a CBMIRS based on deep learning is proposed, which will be composed of two sections of neural networks that together will build a more robust vector for the system and also, this new CBMIRS can be occupied even though the images do not have label, allowing better scalability in a real clinical environment. This proposal is based on the use of combining what has been learned by a segmentation model and a classification model, to transform an image into a vector with more information than using a single neural network, allowing a medical image search system to have better results and scalability. The experimental evidence shows that compared to other state-of-the-art methods in medical image retrieval, the proposal shows a great advantage when looking for images that contain a certain disease in real clinical settings. Furthermore, the effective information transfer capacity is demonstrated by using the resulting embeddings of the segmenter in the retraining of the classifier or by concatenating the resulting embeddings. Finally, it was shown that the use of a near-duplicate elimination algorithm has a regularizing effect on image search, allowing systems that use simpler models to be just as competitive as systems that use more complex models.
Description
Keywords
PROCESAMIENTO DE DATOS , BASES DE DATOS , SISTEMAS DE IMAGENES EN MEDICINA