Deep Learning para la generación de imágenes histopatológicas realistas mediante aritmética de vectores conceptuales
Generative Adversarial Networks (GANs) can offer a way to tackle chronic lack of labeled samples in the medical imaging field, not only using unbounded generation but also by using conditional generation on different attributes of our choice or by modifying the results of this generation with the ap...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2019 |
| País: | España |
| Institución: | Universitat Oberta de Catalunya (UOC) |
| Repositorio: | O2, repositorio institucional de la UOC |
| OAI Identifier: | oai:openaccess.uoc.edu:10609/100086 |
| Acceso en línea: | http://hdl.handle.net/10609/100086 |
| Access Level: | acceso abierto |
| Palabra clave: | aprendizaje profundo redes generativas antagónicas histopatología aritmética de vectores conceptuales deep learning generative adversarial network histopathology latent space arithmetic aprenentatge profund xarxes generatives antagòniques histopatologia aritmètica de vectors conceptuals Machine learning -- TFM Aprenentatge automàtic -- TFM Aprendizaje automático -- TFM |
| Sumario: | Generative Adversarial Networks (GANs) can offer a way to tackle chronic lack of labeled samples in the medical imaging field, not only using unbounded generation but also by using conditional generation on different attributes of our choice or by modifying the results of this generation with the application of conceptual arithmetic operations between images. We train a Deep Convolutional GAN and a conditional DCGAN on a breast cancer dataset and we make use of some of its properties in order to edit histopathological images by arithmetically operating its latent vectors. The breast cancer dataset is created out of several patients WSI images, and it contains annotations for positive and negative samples. By using these arithmetical properties, we are able to perform several operations on the images, like inversion of the class of a sample (from tumorous to normal or vice versa), smooth interpolations of two samples that can show the transition between two states or two types of them or transference of features from one sample to another by combining additions and subtractions of latent vectors. Aside from its utility to augment labeled datasets for supervised algorithms, this type of edition may be useful in other situations, for example being used as didactic or informative material or as a way to deal with the frequent privacy and anonymity problems that can be encountered when working with this type of medical data. |
|---|