Javascript must be enabled for the correct page display

Classifying Infection Stages in Patients with Ventricular Assist Devices

Luneburg, Noel (2019) Classifying Infection Stages in Patients with Ventricular Assist Devices. Master's Thesis / Essay, Artificial Intelligence.

[img]
Preview
Text
mAI_2019_LuneburgN.pdf

Download (6MB) | Preview
[img] Text
toestemming.pdf
Restricted to Registered users only

Download (55kB)

Abstract

Deep learning has made large steps forward in the image processing field, including in highly specific medical cases. Skin infections are a common problem in patients with a heart assist device. The research in this thesis focuses on classifying skin infections by making use of recent deep learning advancements. A semantic segmentation convolutional neural network was trained to identify objects in the images unrelated to infection status. A small and imbalanced data set formed one of the challenges. In addition to affine transformations, generative adversarial networks (GANs) were used to generate training data as a form of data set augmentation, artificially increasing the size of the data set. Results show that applying segmentation masks of irrelevant objects to images during the classification step does not improve classification performance. Images similar to the training data were generated by a GAN. However, the available data set was likely too small for GANs to be used as data augmentation. Classification results show that infection class prediction based solely on external photos is a difficult problem, as confirmed by a validation experiment with medical experts.

Item Type: Thesis (Master's Thesis / Essay)
Supervisor:
Supervisor nameSupervisor E mail
Wiering, M.A.M.A.Wiering@rug.nl
Supervisor (outside RUG):
Supervisor outside RUG nameSupervisor outside RUG E mail
Jansen, SybrenUNSPECIFIED
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 02 Oct 2019
Last Modified: 04 Oct 2019 09:04
URI: http://fse.studenttheses.ub.rug.nl/id/eprint/21061

Actions (login required)

View Item View Item