Applying deep learning/GANs to histology image for data augmentation: a general study
In medical imaging tasks, annotations are made by radiologists with expert knowledge on the data and task. Therefore, Histology images are especially difficult to collect as they are: expensive, time consuming and information can not be always disclosed for research. To tackle all these issues data...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2018 |
| País: | España |
| Institución: | Universitat Oberta de Catalunya (UOC) |
| Repositorio: | O2, repositorio institucional de la UOC |
| OAI Identifier: | oai:openaccess.uoc.edu:10609/91330 |
| Acceso en línea: | https://hdl.handle.net/10609/91330 |
| Access Level: | acceso abierto |
| Palabra clave: | data augmentation GAN deep learning histology aprenentatge profund histologia dades d'augment aprendizaje profundo histología datos de aumento Artificial intelligence -- TFM Intel·ligència artificial -- TFM Inteligencia artificial -- TFM |
| Sumario: | In medical imaging tasks, annotations are made by radiologists with expert knowledge on the data and task. Therefore, Histology images are especially difficult to collect as they are: expensive, time consuming and information can not be always disclosed for research. To tackle all these issues data augmentation is a popular solution. Data augmentation, consist of generating new training samples from existing ones, boosting the size of the dataset. When applying any type of artificial neural network, the size of the training is key factor to be successful. especially when employing supervised machine learning algorithms that require labelled data and large training examples. We present a method for generating synthetic medical images using recently presented deep learning Generative Adversarial Networks (GANs). Furthermore, we show that generated histology images can be used for synthetic data augmentation and improve the performance of CNN for medical image classification. The GAN is a non-supervised machine learning technique where one network generates candidates (generative) and the other evaluates them (discriminative) to generate new sample like the original. In our case we will focus in a type of GAN called Deep Convolution Generative Convolutional Network (DCGAN) where the CNN architecture is used in both networks and the discriminator is reverting the process created by the generator. Finally, we will apply this technique, for data augmentation, with two different datasets: Narrow bone and Breast tissue histology image. To check the result, we will classify the synthetic images with a pre-trained CNN with real images and labelled by specialist. |
|---|