I have been experimenting with greyscale CXRs and also with different colormaps to observe a possible difference in the performance of deep learning models. Has anyone identified the best colormap (like jet, HSV, summer, hot etc.) to convert the greyscale to RGBs exclusively for CXRs? Can you also point to the references in this regard?
This decision probably depends strongly on the data as well as the planned deep learning model. As such, it would be wise to test out many variants instead of assuming that one is best beforehand.
Could you explain the motivation behind converting grayscale data to color maps in the first place?
It depends on your purpose. I think colormaps like hot, or cold are better than others, because they preserve the pattern of the gray image. I mean dark parts are mapped to dark, and light parts to light. While the results of colormaps such as spring, or jet, are not compatable to the gray image.
Ideally, it should not matter in deep learning.
With deep learning, what we do is we make the machines do both feature engineering and learning task.
All we do when converting from gray scale to color map is applying some function to the image. With sufficient expressive power deep networks should be able to do this; may be even better. This is what exactly deep learning is good at. So from my view point, if you have a sufficiently expressive network, your input format should not make much difference.
But one rational way to compare these different color maps will be to freeze the network architecture and then compare the results. But it brings up one question - are you looking to improve the method? Or are you looking to just compare different ways of inputs.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification