search

fronay / ceph_xray_keypoints

Try out from-scratch CNN & retrained CNN for cephalometric keypoint detection on skull profile x-ray
14 stars 1 forks source link

about the result #2

Open dusoosoo opened 5 years ago

dusoosoo commented 5 years ago

hello,guys, I also do this cephalometry work now. How about your accuracy rate in this experiment? Do many pictures has good effect? Thank you!

fronay commented 5 years ago

Hi DuSooSoo, overall the accuracy rate was disappointing for the majority of the cephalometric landmarks - with some exceptions (e.g. Sella Turcica, Nasion, Menton seemed to work well and be within 1-2mm/few pixels on image).

I'm guessing that the amount of available annotated images might be a bit low (only around 400 total from the Wang / ISBI dataset) for this non-pretrained deep learning approach, even when applying data augmentation like mirroring etc.

If you had thousands of labeled images, though, I think the accuracy could be improved to the point where it becomes useful.

If you are considering this for any sort of research project into automated cephalometry, I'd suggest checking published papers on the topic on Google Scholar or Pubmed first - some of them claimed reasonable accuracy rates and might at least show what algorithm they used.

dusoosoo commented 5 years ago

fronay,thank you for answered me very much ! And I want to ask that do you have some code about previous work in this topic ,such as ISBI 2014 grand challenges' code or some other method like active shape model ?

fronay commented 5 years ago

Back when I tried this, I didn't find published code (not on github, at least :) ) - it seems that most of the time, only the algorithm but not the implementation is shared (unfortunately!).

If you find anything where the code was shared and that claimed good results, feel free to put a link here...I would also be very curious to see it, even if I'm not working on something at the moment!