Generative Adversarial Networks to Synthetically Augment Data for Deep Learning based Image Segmentation. Neff, T., Payer, C., Štern, D., & Urschler, M. In Proceedings of the OAGM Workshop 2018: Medical Image Analysis, Hall/Tyrol, Austria, pages 22-29, 2018.
doi  abstract   bibtex   
In recent years, deep learning based methods achieved state-of-the-art performance in many computer vision tasks. However, these methods are typically supervised, and require large amounts of annotated data to train. Acquisition of annotated data can be a costly endeavor, especially for methods requiring pixel-wise annotations such as image segmentation. To circumvent these costs and train on smaller datasets, data augmentation is commonly used to synthetically generate additional training data. A major downside of standard data augmentation methods is that they require knowledge of the underlying task in order to perform well, and introduce additional hyperparameters into the deep learning setup. With the goal to alleviate these issues, we evaluate a data augmentation strategy utilizing Generative Adversarial Networks (GANs). While GANs have shown potential for image synthesis when trained on large datasets, their potential given small, annotated datasets (as is common in e.g. medical image analysis) has not been analyzed in much detail yet. We want to evaluate if GAN-based data augmentation using state-of-the-art methods, such as the Wasserstein GAN with gradient penalty, is a viable strategy for small datasets. We extensively evaluate our method on two image segmentation tasks: medical image segmentation of the left lung of the SCR Lung Database and semantic segmentation of the Cityscapes dataset. For the medical segmentation task, we show that our GAN-based augmentation performs as well as standard data augmentation, and training on purely synthetic data outperforms previously reported results. For the more challenging Cityscapes evaluation, we report that our GAN-based augmentation scheme is competitive with standard data augmentation methods.
@inproceedings{
 title = {Generative Adversarial Networks to Synthetically Augment Data for Deep Learning based Image Segmentation},
 type = {inproceedings},
 year = {2018},
 pages = {22-29},
 id = {e0d3f062-b87e-3f45-9924-f20342bfe804},
 created = {2018-09-04T04:52:29.681Z},
 file_attached = {false},
 profile_id = {53d1e3c7-2f16-3c81-9a84-dccd45be4841},
 last_modified = {2019-11-08T01:39:29.964Z},
 read = {false},
 starred = {false},
 authored = {true},
 confirmed = {true},
 hidden = {false},
 citation_key = {Neff2018},
 folder_uuids = {0ec41d70-75f1-4a99-820b-0a83ccc37f54},
 private_publication = {false},
 abstract = {In recent years, deep learning based methods achieved state-of-the-art performance in many computer vision tasks. However, these methods are typically supervised, and require large amounts of annotated data to train. Acquisition of annotated data can be a costly endeavor, especially for methods requiring pixel-wise annotations such as image segmentation. To circumvent these costs and train on smaller datasets, data augmentation is commonly used to synthetically generate additional training data. A major downside of standard data augmentation methods is that they require knowledge of the underlying task in order to perform well, and introduce additional hyperparameters into the deep learning setup. With the goal to alleviate these issues, we evaluate a data augmentation strategy utilizing Generative Adversarial Networks (GANs). While GANs have shown potential for image synthesis when trained on large datasets, their potential given small, annotated datasets (as is common in e.g. medical image analysis) has not been analyzed in much detail yet. We want to evaluate if GAN-based data augmentation using state-of-the-art methods, such as the Wasserstein GAN with gradient penalty, is a viable strategy for small datasets. We extensively evaluate our method on two image segmentation tasks: medical image segmentation of the left lung of the SCR Lung Database and semantic segmentation of the Cityscapes dataset. For the medical segmentation task, we show that our GAN-based augmentation performs as well as standard data augmentation, and training on purely synthetic data outperforms previously reported results. For the more challenging Cityscapes evaluation, we report that our GAN-based augmentation scheme is competitive with standard data augmentation methods.},
 bibtype = {inproceedings},
 author = {Neff, Thomas and Payer, Christian and Štern, Darko and Urschler, Martin},
 doi = {10.3217/978-3-85125-603-1-07},
 booktitle = {Proceedings of the OAGM Workshop 2018: Medical Image Analysis, Hall/Tyrol, Austria}
}
Downloads: 0
About
Documentation
Keywords
Search
Users
Terms and Conditions of Use
Privacy Policy
Sitemap
© BibBase.org 2021