Nikonov, Arseniy (2022) Exploring the latent space of the StyleGAN model for the controlled generation of synthetic colonoscopy image. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2022_NikonovA.pdf Download (32MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (123kB) |
Abstract
Machine learning models have successfully solved many tasks in the past years, but they usually require a massive amount of data to succeed. Medical dataset typically doesn’t have enough data to achieve the best performance of machine learning models. The colonoscopy domain is one of the fields that have only limited publicly accessible data, and not all of that data can be used due to the difference in the equipment. At the same time, computer-assisted diagnosis proved helpful in increasing the detection rate during colonoscopy. To solve the data scarcity, generative adversarial networks(GAN) can be used to synthesize fake images. StyleGAN2-Ada model produced realistic images that can be used as an addition to the original training dataset. Even though the images produced by StyleGAN are quite realistic there is a lack of control over the generated image. Several methods of exploring the latent space of StyleGAN were explored to achieve better control over the generated image. Those methods proved that latent exploration is possible as a concept but only achieved such changes as the vertical/horizontal location of a polyp or change in its size. The performance of a detector with the addition of synthesized images achieved a mean average precision score 83.4% compared to the 82.7% obtained by the detector trained on the original dataset.
| 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: | 25 Oct 2022 14:43 |
| Last Modified: | 25 Oct 2022 14:43 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/28796 |
Actions (login required)
 |
View Item |