search

sct-pipeline / csa-atrophy

Evaluate the sensitivity of atrophy detection with SCT
https://csa-atrophy.readthedocs.io/
MIT License
1 stars 0 forks source link

Add graph showing intra-subject variability (STD) as a function of atrophy estimation error #86

Closed jcohenadad closed 3 years ago

jcohenadad commented 4 years ago

Once the graph is done, we should identify the 5-10 subjects with worse STD, and see if they are associated with strong movements/artifacts.

If that's the case, it should be reported as a new finding in the manuscript.

PaulBautin commented 4 years ago

Graph without removing any subjects

The first graph shows COV in function of error % without removing any subjects. fig_err_in_function_of_std

The 2nd graph shows that T1w and T2w outliers seem not correlated. Subjects in red are the worst outliers of T1w images and subjects in green are the worst outliers of T2w images. Screenshot from 2020-10-24 15-16-41

After removing subjects

Presently if CSA is not computed the vertebrae levels are removed from the dataframe. It looks like the segmentation on images with missing CSA does not cover C3-C5 cord, this could be related to image transformation. We could try to re-integrate padding to see if it resolves this issue. example: Cropped and transformed image: sub-barcelona04_T1w_RPI_r_crop_r0.98_t1.zip segmentation of ^ image sub-barcelona04_T1w_RPI_r_crop_r0.98_t1_seg.zip

In the following graph all subjects with a missing CSA were removed (not only the vertebra level). It seems to eliminate values with high COV and low error % improving the correlation between both metrics.

fig_err_in_function_of_std_without_error

In this last graph, i tried to identify outlier subjects and looked at their image quality. It is difficult to affirm that these are the worst images, but it is obvious that if the subject is an outlier for one rescaling it has good chances to be an outlier for the other rescalings. : fig_err_in_function_of_std

jcohenadad commented 4 years ago

Very cool investigations @PaulBautin ! Too bad these results don't follow our intuitions (about the correlation). Could you open a PR so I can look at the code (and understand exactly how you generated those figures)?

PaulBautin commented 4 years ago

@jcohenadad. The merge of PR #89, has not modified the results and the number of outliers. No correlation between COV and percentage can be deduced. Used repo for t1: https://github.com/PaulBautin/csa-atrophy/tree/compute_canada Used repo for t2: https://github.com/sct-pipeline/csa-atrophy

Screenshot from 2020-11-04 16-28-49

jcohenadad commented 4 years ago

thank you for looking into it @PaulBautin , but as a sanity check i'd still like to see a side-by-side results comparison as mentioned in https://github.com/sct-pipeline/csa-atrophy/issues/83#issuecomment-721986071

PaulBautin commented 4 years ago

As predicted PR #89 eliminates values with very high COV:

repo for results before #89: https://github.com/sct-pipeline/csa-atrophy/tree/stat_cov_err repo for results after #89: https://github.com/PaulBautin/csa-atrophy/tree/compute_canada

Screenshot from 2020-11-04 17-21-02

jcohenadad commented 4 years ago

it's reinsuring. Even though there are still effects that we don't fully understand (the offset for T1w), we can trust these results better, now that the bug with the crop is fixed. I think we can move forward with the article.