Use this link to cite:
http://hdl.handle.net/2183/31922 Segmentación de tejido cancerígeno en imágenes de histopatología tintada con Hematoxilina-eosina usando aprendizaje profundo
Loading...
Identifiers
Publication date
Authors
Pérez Calavia, Daniel
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] Uno de los campos que ha visto más actividad en los últimos años ha sido el uso de técnicas
de deep-learning o aprendizaje profundo en medicina. Gracias a la gran cantidad de datos
disponibles, a un continuo aumento de la capacidad de cálculo y a una mejora continua en los
algoritmos de aprendizaje usados, no es de extrañar que el uso de estas técnicas en sistemas
de diagnóstico asistido por ordenador sea cada vez más común.
Debido a que el cáncer de mama es el cáncer con más prevalencia en las mujeres, cualquier
mejora en el diagnostico tendría un gran efecto beneficioso en la lucha con la enfermedad.
Por ello en este trabajo se desarrolla un sistema de segmentación sobre imágenes histológicas
de cáncer de mama tintadas con H&E sobre un dataset de referencia internacional aplicando
técnicas de deep-learning. En concreto usamos una arquitectura muy popular en el campo
de segmentación de imágenes médicas, además del desarrollo de un sistema de selección de
parches para procesar imágenes de grandes dimensiones, un procesamiento de los datos de
entrada mediante técnicas como data augmentation y un sistema para sustituir parches ya
aprendidos. A través de las iteraciones por las que pasa el desarrollo, se muestra la evolución
de estos y otros sistemas y se analizan los efectos de los cambios propuestos en los resultados
obtenidos.
[Abstract] One of the fields that has seen more activity in recent years has been the use of deeplearning techniques in medicine. Due to the large amount of data available, an increase in computing capacity and improvement in the learning algorithms used, it is not surprising that the use of these techniques in computer-aided diagnostic systems is becoming more and more common. Because breast cancer is the most prevalent cancer in women, any improvement in diagnosis would have a great beneficial effect in fighting the disease. For this reason, in this work we develop a system for the segmentation of cancerous tissue in hystopathology images tinted with H&E on a dataset of international reference using deep-learning techniques. In particular, we use a very popular in the field of medical image segmentation, in addition we develop a patch selection system in order to process large images, a data processing pipeline using data augmentation techniques and a system that substitutes already learned patches. Through multiple iterations we show the evolution of the developed systems and analyze the effects of the proposed changes on the results.
[Abstract] One of the fields that has seen more activity in recent years has been the use of deeplearning techniques in medicine. Due to the large amount of data available, an increase in computing capacity and improvement in the learning algorithms used, it is not surprising that the use of these techniques in computer-aided diagnostic systems is becoming more and more common. Because breast cancer is the most prevalent cancer in women, any improvement in diagnosis would have a great beneficial effect in fighting the disease. For this reason, in this work we develop a system for the segmentation of cancerous tissue in hystopathology images tinted with H&E on a dataset of international reference using deep-learning techniques. In particular, we use a very popular in the field of medical image segmentation, in addition we develop a patch selection system in order to process large images, a data processing pipeline using data augmentation techniques and a system that substitutes already learned patches. Through multiple iterations we show the evolution of the developed systems and analyze the effects of the proposed changes on the results.
Description
Editor version
Rights
Todos os dereitos reservados







