I designed an algorithm for MRI data denoising which has good properties under heavy Rician noise (sigma is greater than 80). The method was tested on some phantoms and Rician noise generators. Now I want to test the algorithm on real MRI data. In the MRI data, which I have, the noise level is low. Is it methodologically correct adding noise to a real MR image to increase its level based on Rician noise generator?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Using real data is a way to validate the assumption that your algorithm will behave with real data in a similar way to the behavior with generated data.
In a general sense.
When we want to test an algorithm we can create tests, that are other algorithms that provide inputs and evaluate outputs of the algorithm we are design (your denoising algorithm).
Sometimes acquiring real data is difficult, but we can, making some assumptions generate data with similar distribution from data generator (the procedure by which you apply noise using your Rician noise generators), this makes easier to test your algorithm in much more data, and get more confidence about the algorithm.
When you write your data generator you are making one assumption about the types of distortion the images may have. Testing with real data you may find effects that are not documented, not studied, not expected. This is why in the real world the performance of algorithm may be degraded with measured data compared to generated data. Testing with real data will give you one idea of what is the impact of the neglected effects.
Different neglected effects may occur. The response may be non-linear and the noise do not follow Rice distribution. The parameters of the noise distribution may depend on the position differently from what you assumed (you may have assumed that the distribution of the noise is independent of the position). Your scanner(s) may have different responses due to different temperature. Different scanners (same of different models/manufacturer) may produce data with different characteristics.
When you add noise to your images you are testing how the algorithm performs with the same real images and different Rician additional noise. There is nothing wrong in doing so, but you must know that worse quality images may have more of the distortions you are not considering.
