I am studying the performance of deep learning (DL) models toward abnormality detection in chest X-rays. Due to sparsity of data, I augmented the data using different augmentation strategies including: a) traditional augmentation methods including Gaussian smoothing, unsharp masking, and minimum filtering; and b) Generative Adversarial Networks. Contrary to the existing literature, I find that the DL models showed promising results with traditional augmentation methods (that i have mentioned herewith) than with GAN-generated synthetic images. What brings this performance difference?
To my knowledge, GANs usually can bring the similar information with your orignial data rathar than largely enrich your information. It could be more robust against the adversiral samples or complex noises. This might explain your concerns.
To my knowledge, GANs usually can bring the similar information with your orignial data rathar than largely enrich your information. It could be more robust against the adversiral samples or complex noises. This might explain your concerns.
Danfeng Hong Thanks for your reply. Do you agree to the fact that traditional augmentation like Gaussian smoothing, unsharp masking, and minimum filters try to mimic the biological variance in the test data to improve the classification results?
I would say that traditional augmentation like Gaussian smoothing, unsharp masking, and minimum filters can indeed help improving the performance, since these processing strategies would enrich your training set, thereby further improving the classification performance.
I agree with Danfeng Hong that using GANs would not improve the network's performance. However, I assume that the use of Autoencoders with KL-divergence would help. Recently, many studies use such network representation for cross domain learning by making the distribution of the cross domains similar. However you can generate the images using GANs and process it through such autoencoders for making their distribution similar for data augmentation.
We have studied the effect of utilizing synthetic images produced by GAN for data augmentation purposes. The effect is studied for object detection and segmentation tasks. The paper can be accessed by the link below:
Preprint Red blood cell image generation for data augmentation using ...
https://arxiv.org/abs/1901.06219
- Badges
- Science topic