Generating realistic microscopy images of blood cells using Generative adversarial networks

Berghuis, Rutger (2021). Generating realistic microscopy images of blood cells using Generative adversarial networks. Thesis (Master thesis), E.T.S. de Ingenieros Informáticos (UPM).

Description

Title: Generating realistic microscopy images of blood cells using Generative adversarial networks
Author/s:
  • Berghuis, Rutger
Contributor/s:
  • Alonso Calvo, Raúl
Item Type: Thesis (Master thesis)
Masters title: Ciencia de Datos
Date: June 2021
Subjects:
Faculty: E.T.S. de Ingenieros Informáticos (UPM)
Department: Lenguajes y Sistemas Informáticos e Ingeniería del Software
Creative Commons Licenses: Recognition - No derivative works - Non commercial

Full text

[img]
Preview
PDF - Requires a PDF viewer, such as GSview, Xpdf or Adobe Acrobat Reader
Download (4MB) | Preview

Abstract

Most haematologic diseases are still diagnosed manually using microscopic images of blood. To diagnose these diseases the different blood cells on the microscopic images have to be classified by their cell type. Manually diagnosing these diseases is a very time consuming and subjective process. A system that can automatically segment and classify the different blood cells on a microscopic image with high accuracy, can save experts a lot of time in making diagnoses and make the process a lot more reliable. The recent advances in deep learning have highly increased the possibilities of automatically diagnosing diseases. The most promising technique for classifying the different cell types is the Convolutional Neural Network(CNN). The main limiting factor for most of the applications of deep learning in the medical field is the lack of training data. Deep learning techniques need very large datasets to properly train the model and avoid overfitting. In the medical field it can be hard to obtain a large dataset necessary to train a neural network. The annotations have to be made by medical experts, which is a very time consuming process and restrictions on sharing patient data make it even harder to obtain and share medical datasets. One way to solve the problem of datasets being too small is to use data augmentation techniques. A new method for data augmentation is generating synthetic data using Generative Adversarial Networks(GANs). GANs consist of two artificial neural networks, a generative model and a discriminative model. The generative model generates synthetic data and the discriminative model estimates the probability that the sample data came from the training data. The goal is that the generative model eventually generates synthetic data that is indistinguishable from real data. In this work an image-to-image translation with a conditional GAN is used to generate synthetic microscopic images of blood cells. The method used to generate these pictures is by training the GAN with a dataset containing annotated images of different types of blood cells. These images are used to create the input images that are used to train the GAN. The input images to train the GAN consist of an input image in which the shape of the cell is a certain colour that corresponds to the cell type and an image with the ground truth that consists of the cell itself. After training the GAN with these images of single cell types, first images of single cell types are generated to confirm the quality of the generated different cell types. The final goal is to generate a microscopic image with multiple cells on it of different cell types. The method used to achieve this goal is by creating an input image with the shapes of the to be generated cells in the colour that corresponds to their cell type. The distribution of the cells in the image is determined using a Dorogovtsev-Goltsev-Mendes network. The x and y coordinates of the nodes in the network are used as the x and y coordinates of the cells in the image and the z coordinate is used to determine the size of the cell. For the shapes of the cells a shape is taken from the input training images of the same category. The shape is then randomly rotated and flipped to create more randomness. To generate the synthetic microscopic image, first the background is generated procedurally by generating it block by block. Next the cells are generated cell by cell and the generated cells are pasted onto the background. The result is a synthetic microscopic image of different blood cells. The generated images are not perfect yet and there is still room for a lot of improvement. However, for a first work the images show a lot of promise to be used in the future for medical applications such as training physicians or improving the performance of a CNN.

More information

Item ID: 68758
DC Identifier: https://oa.upm.es/68758/
OAI Identifier: oai:oa.upm.es:68758
Deposited by: Biblioteca Facultad de Informatica
Deposited on: 07 Oct 2021 14:19
Last Modified: 07 Oct 2021 14:19
  • Logo InvestigaM (UPM)
  • Logo GEOUP4
  • Logo Open Access
  • Open Access
  • Logo Sherpa/Romeo
    Check whether the anglo-saxon journal in which you have published an article allows you to also publish it under open access.
  • Logo Dulcinea
    Check whether the spanish journal in which you have published an article allows you to also publish it under open access.
  • Logo de Recolecta
  • Logo del Observatorio I+D+i UPM
  • Logo de OpenCourseWare UPM