1
$\begingroup$

I am studying the performance of deep learning models toward abnormality detection in chest X-rays.

Due to sparsity of data, I augment the data using different augmentation strategies including:

  • Traditional augmentation methods (Gaussian smoothing, unsharp masking, and minimum filtering)
  • Generative Adversarial Networks

Contrary to the existing literature, I find that the models showed promising results with traditional augmentation methods (that i have mentioned herewith) than with GAN-generated synthetic images.

What brings this performance difference?

$\endgroup$
1
$\begingroup$

In general, you have to be careful when using data augmentation.

For example, doing rotation for this kind of image makes sense, we expect to see any of these images as potential 'real-life' example : enter image description here


However, doing rotation for this kind of image is less meaningful. We don't expect to see this in 'real-life' example : enter image description here


And GAN potentially makes generated image meaningless. If your GAN produce 'thrash' augmented-data, then your network will train and learn 'thrash', which you don't want.

When you are training your model on GAN generated images, you're actually training your model to recognize GAN-generated image, not real-life example.

Sources: - towardsdatascience - quora question

|improve this answer|||||
$\endgroup$
  • $\begingroup$ I performed traditional and gan augmentation as individual experiments to test the performance of the DL model. I never perform rotation or flipping as it makes no sense for a chest x ray data. The issues boils down to what kind of variation does the traditional and gan augmented data have when training the DL model. Could you explain this variance? $\endgroup$ – shiva Nov 30 '18 at 13:04
  • $\begingroup$ Rotation / flipping is just an example. If your GAN makes new data which does not make a lot of sense, your model will be trained on these data, and learn how to recognize data that does not make a lot of sense. And when you test your model on real data (validation / test set), the data having sense, your model will not perform well on it. $\endgroup$ – Astariul Dec 3 '18 at 0:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.