Skip to content

Latest commit

 

History

History
81 lines (52 loc) · 3.55 KB

README.md

File metadata and controls

81 lines (52 loc) · 3.55 KB
Raw

dataset

dataset ped-ct-seg https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=89096588

notes

python gen_args.py $DATADIR $RESULTSDIR
  • (optional) clean up generated data from prior run (add -delete to delete)
cd $DATADIR
find . -name "liver.nii.gz" -print
find . -name "segmentations.nii.gz" -print
find . -name "scores.json" -print
  • submit inference and dice computation, and results aggregation DAG jobs.
rm condor.dag.* docker_stderror log/* results/*.csv
condor_submit_dag condor.dag

results

** DISCLAIMER, results to be verified, likely bugs present in my scripts. **

Dice computed using manual vs totalsegmentator (with and without --fast flag) contours using dataset ped-ct-seg (n=~359) (csvs available in results folder).

I don't quite believe the results yet (again see Dislaimer above), nevertheless, some observations are listed below:

png

  • As advertised, --fast indeed have a lower accuracy (assuming manual contours are good - i have not yet reviewed the ped-ct-seg contours) - not sure who got the adrenal l/r correct.

  • Lung & kidneys have fairly decent dice (>0.8), implying

  • That said, a decent amount of organs (say bone, heart, spleen) that I think should have a dice > 0.6 have dice < 0.5! If model got the liver location "right" (>0.5), why is the spleen dice so low (<.2)? Could it be due to contrast phases. Or is it just that organ size varies too much due to age, thus model is unable to "generalize" - sets of adults organs is "vastly" different from younger organs...

    • lo and behold, below shows the age vs dice plot (liver and also all organs), dice does indeed drop as age decreases.

    • Interestingly the drop varies among organ. For kidney and lung, dice plumets at age < ~4, whyile you have a more gradual drop for intestine, esophgus, liver, and surprisingly bone (?).

    • If we ignore adrenal and other small organs, how come we don't see the dice/age trend in heart, bladder? Further, why is deviation of dice relatively large? (** AGAIN, still not ruling out potential bugs on my end **, and manual countour quality - ) What's up with stomach? how is bladder dice worse/comparable to adrenal? Likely, there is still bug/factors that contribute to the relatively lower dice in some organs.

    • why would dice in bone drop with age? Maybe there is difference in contour SOP / how i'm aggregating the bone segment outputs from totalsegmentator.

png png

notes

something doesn't seem right...

http://gtx.local/compare?case_id=Pediatric-CT-SEG-03A54FD1&dataset=ped-ct-seg

http://gtx.local/compare?case_id=Pediatric-CT-SEG-02AC04B6&dataset=ped-ct-seg

containers used in condor jobs

https://github.com/wasserth/TotalSegmentator/blob/master/Dockerfile

https://github.com/pangyuteng/totalsegmentator-and-friends/blob/main/assess/docker-with-weights/Dockerfile

docker pull pangyuteng/totalsegmentator:latest

https://github.com/pangyuteng/public-scratch/tree/master/ml-docker

docker pull pangyuteng/ml:latest