Popescu, Mihai (2021) Generating Synthetic Training Data using Deep Generative Adversarial Networks in Medical Endoscopy Images. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2021_MihaiP.pdf Download (59MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (98kB) |
Abstract
Building automated detection tools for endoscopy procedures has been a pursued interest in the field of machine learning to reduce the number of omitted polyps during endoscopies. Training such systems is difficult in the current landscape as the availability of medical images containing polyps is low. This thesis attempts to solve data scarcity by synthesizing images containing polyps using generative adversarial networks in order to augment existing polyp datasets used by detection models. The most promising model based on StyleGAN2-Ada produced realistic images that, when augmented into the original training set of the detector, obtained a mean average precision score of 92.13\% compared to the 92.44\% obtained by the detector trained on a non-augmented dataset. Although the performance of the model did not increase, the quality of the generated images was impressive from a realism standpoint and promising conclusions could be drawn regarding the possibility of manipulating the latent space and generative conditional embedding of the network to generate custom types of polyps.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Cnossen, F. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 29 Nov 2021 12:40 |
| Last Modified: | 29 Nov 2021 12:40 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/26279 |
Actions (login required)
 |
View Item |