Dear All,
Im wondering if hypothesis testing is needed in the following case:
I have a set of image medical images (it is relevant here to understand, that medical images are normally corrected during the reconstruction process for systematic errors introduced by detectors or because correct, direct measurements of the required quantity is normally not possible, e.g. effects like beam hardening, scatter, etc. need to be corrected) which I want to use for testing of how much new methods for such corrections can improve the image. To test this, I would reconstruct the very same data set with both correction variants and evaluate/compare image based figure of merits, e.g the contrast of two regions of interest.
Now my question is: do I need hypothesis testing here, e.g location tests of the both mean values over the voxels of the ROI, distribution tests, etc. to confirm that there is an effect of the changed data correction methods? (It can be assumed, that the reconstruction process and all correction methods are deterministic, i.e. I will obtain exactly the same image when I reconstruct the same data set twice). My understanding is, that hypothesis testing is meaningless in this case, as any difference that I may observe in the reconstructed images necessarily must be due to the different correction method and there is no possibility, that these difference occur by pure chance, as for instance in comparison of groups, where I cannot run the same experiment twice modifying only one condition and leaving all others untouched.
I would appreciate, if someone with sound expertise in statistics and hypothesis testing could comment on this, ideally recommending references, where this case is discussed.
Many thanks, Christoph
Hey Christoph,
Intuitively, to me, it seems that the only thing worthwhile checking here is that the deterministic process operates as expected on all images (i.e., that all/most images exhibit an improvement in the real signal, rather than some false-positive). Whereas this could be formulated as some p-value test,I think that signal-to-noise tests are more straightforward in this case. (e.g., https://iopscience.iop.org/article/10.1088/0031-9155/30/6/001/meta?casa_token=N7FhWEpqin4AAAAA:jZi2QyGNl7Z_mjKTziIUuVZ4KVY7BVH3Zoi0HpkaNWK0oI-1Wft9KSaI8lSFBd0VwRI8AV08_QGJZZY)
Hope this helps.
Good day. Because your empirical data is deterministic, and the behavior of a deterministic process is predictable and can be precisely defined. So I don't think it needs to be applied.
You apply your algorithms to a sample of images. You might need a test to see if the improvements can also be expected on other images not in your sample. This might be relevant if images come with different structures that are addressed differently by your algorithms.
And it would be good to test some relevant hypotheses, e.g. that the rate of correct diagnoses based on the altered images is improved. It clearly makes no sense to test hypotheses about changed criteria of the images that were changed.
Your thinking is right if you are talking about the common "null hypothesis". When you are comparing two different populations, there is no need to do a hypothesis test - you know that they are different (what is the chance of the same mean to 17 digits?). It is commonly done, but silly in most cases. The question is the size of the difference and its importance. A confidence interval should do the trick, and "equivalence testing" is the right approach.
Good afternoon,
When we need to take decision among a sample, have two way one is deterministic second probabilistic, if you choose the second one, need to compare with table value. This situation we use hypothesis.
When interpreting research findings, researchers need to assess whether these findings may have occurred by chance. Hypothesis testing is a systematic procedure for deciding whether the results of a research study support a particular theory which applies to a population.
- Badges
- Science topic
- Similar topics
- Biological Science
- Molecular Biology
- Biomolecules
- Nucleic Acids
- RNA