Publication: Análisis y propuesta de métricas de calidad de imagen médica que mimetizan al observador humano
Loading...
Files
Official URL
Full text at PDC
Publication Date
2017-02-09
Authors
Prieto Renieblas, Gabriel
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Complutense de Madrid
Citations
Exportar
Abstract
La investigación que se presenta en este documento se centra en el paradigma de la percepción automática de la calidad de imagen médica, y en la correlación de dicha percepción con la percepción humana. El análisis de la calidad de imagen médica tiene un lugar central en el diseño de sistemas de imagen para diagnóstico. El objetico de este análisis es, usualmente, el de diseñar una métrica capaz de evaluar la calidad de imagen percibida por un observador, una IQM por sus siglas en inglés (Image Quality Metric). Más aún, el objetivo de un gran número de investigadores es el de desarrollar métricas automatizadas capaces de reproducir los resultados que produciría un observador humano ante dichas imágenes. De forma prácticamente universal, estas métricas se desarrollan como programas informáticos, desarrollados en uno u otro lenguaje de programación. Hasta el momento solo se han obtenido éxitos parciales. El número existente de aproximaciones a este problema y, por tanto, el número de algoritmos desarrollado es elevado; sin embargo, sigue siendo una cuestión abierta. En la literatura médica se encuentran dos aproximaciones claramente diferenciadas; una de ellas está basada en modelos de la función visual humana o en modelos ideales de observador (bien juntos o por separado). Estos modelos tratan de reproducir el procesado de la imagen en el observador desde su captación en el ojo hasta su procesado de alto nivel en el cerebro. Son modelos muy complejos, con una validez limitada y no han mostrado respuestas satisfactorias y, sobre todo, generalizables. Son estudios y modelos típicos en el campo de la imagen médica. Por otro lado, los especialistas del mundo de las Telecomunicaciones han analizado la calidad de imagen desde un punto de vista más amplio, más enfocado en estudios de imágenes naturales (aquellas presentes en el entorno natural humano), y tanto en estudios de imagen fija como en vídeo. Muchos de estos análisis están basados en aproximaciones “top‐down” al sistema visual humano. Estos modelos proponen hipótesis de carácter general acerca del funcionamiento del sistema visual humano y construyen modelos del mismo basándose en dichas hipótesis. Algunos de estos estudios han propuesto y desarrollado métricas muy bien correlacionadas con la percepción humana. Es quizá sorprendente que, hasta hace unos años, ha habido muy pocos estudios sobre la aplicación de estas métricas al campo de la imagen médica. Dentro de este acercamiento, la métrica que ha tenido más éxito ha sido, sin ningún género de dudas, SSIM, presentada por Wang, Bovik y Simoncelli en el año 2004. Esta métrica se basa en la teoría propuesta por Wang y Bovik sobre el funcionamiento del sistema visual humano. Esta teoría afirma que nuestro sistema visual está especialmente adaptado para extraer información estructural de una imagen. Es una aproximación en la que se parte de una teoría del funcionamiento general del sistema visual humano, en lugar de deducir un esquema de funcionamiento a partir de sus elementos funcionales. A partir de esta métrica se ha desarrollado una amplia familia de índices que comparte la estructura básica con SSIM y que ha obtenido correlaciones crecientes entre los resultados de dichas métricas y los resultados del observador humano. Actualmente es la métrica más usada para medir la calidad de imagen percibida en la industria del vídeo por cable y por satélite...
The research shown in this document is focused on the automated perception of the image quality in medical imaging and its correlation with the human perception of this image quality. Image quality analysis plays a central role in the design of imaging systems for medical diagnosis. The final objective of this image quality analysis is usually to design a metric able to score the perceived quality of a medical image: an image quality metric (IQM). Moreover, the goal of a great number of researches is to develop an automated metric capable of mimic the results produced by a human observer. Usually these metrics are developed as computerized algorithms. So far, only partial success has been achieved. The number of algorithms and approaches in the literacy is high and the problem is still an open question. Remarkably, in the scientific literacy there are two approaches clearly different; one is based on human visual function models or on ideal observer models, (considered together or not). They are models that try to reproduce the image process from the eye to the perception centers in the brain, even modelling (in some approaches) the neuronal response or the whole visual system. These models are usually very complex, with a limited validity and they have not shown generalized and satisfactory responses. These models are typical in the medical image field. For years, many specialists in Telecommunications have analyzed the quality image problem from a broader point of view, more focused on the studies of natural images (those from the natural environment), in still image as well as in video image. Many of these studies have been based on “top‐down” models of the human visual system. These models propose some hypotheses about the general function of the human visual system and build their human visual models according to these hypotheses. Some of these studies have proposed metrics that correlate very well with the human perception. Surprisingly, so far, there are a few studies and applications of these metrics in the medical imaging field. Attending to this approach, the most successful metric developed lately has been the Structural SIMilarity Index (SSIM) proposed by Wang et al. in 2004. This metric is based on the Wang and Bovik’s theory about the human visual system. This theory states that our visual system is highly adapted to extract structural information from an image (note the “top‐down” approach). A broad family of metrics has been developed based on this metric, with increased correlations between the metrics and the human perception. One of the most promising members of this family is the cross correlation multiscale coefficient of SSIM, the so called R*, developed by Rouse & Hemami in 2009. Its design was focused around the problem of perception near the limit of visibility. This task is of great importance in medical phantom analysis and, generally speaking, in the field of medical imaging.
The research shown in this document is focused on the automated perception of the image quality in medical imaging and its correlation with the human perception of this image quality. Image quality analysis plays a central role in the design of imaging systems for medical diagnosis. The final objective of this image quality analysis is usually to design a metric able to score the perceived quality of a medical image: an image quality metric (IQM). Moreover, the goal of a great number of researches is to develop an automated metric capable of mimic the results produced by a human observer. Usually these metrics are developed as computerized algorithms. So far, only partial success has been achieved. The number of algorithms and approaches in the literacy is high and the problem is still an open question. Remarkably, in the scientific literacy there are two approaches clearly different; one is based on human visual function models or on ideal observer models, (considered together or not). They are models that try to reproduce the image process from the eye to the perception centers in the brain, even modelling (in some approaches) the neuronal response or the whole visual system. These models are usually very complex, with a limited validity and they have not shown generalized and satisfactory responses. These models are typical in the medical image field. For years, many specialists in Telecommunications have analyzed the quality image problem from a broader point of view, more focused on the studies of natural images (those from the natural environment), in still image as well as in video image. Many of these studies have been based on “top‐down” models of the human visual system. These models propose some hypotheses about the general function of the human visual system and build their human visual models according to these hypotheses. Some of these studies have proposed metrics that correlate very well with the human perception. Surprisingly, so far, there are a few studies and applications of these metrics in the medical imaging field. Attending to this approach, the most successful metric developed lately has been the Structural SIMilarity Index (SSIM) proposed by Wang et al. in 2004. This metric is based on the Wang and Bovik’s theory about the human visual system. This theory states that our visual system is highly adapted to extract structural information from an image (note the “top‐down” approach). A broad family of metrics has been developed based on this metric, with increased correlations between the metrics and the human perception. One of the most promising members of this family is the cross correlation multiscale coefficient of SSIM, the so called R*, developed by Rouse & Hemami in 2009. Its design was focused around the problem of perception near the limit of visibility. This task is of great importance in medical phantom analysis and, generally speaking, in the field of medical imaging.
Description
Tesis inédita de la Universidad Complutense de Madrid, Facultad de Medicina, Departamento de Radiología y Medicina Física, leída el 19-01-2016