which code to generate candidate

[shartoo]

Hi,is there any code to generate candidate from CT scans?LUNA2016 provide mask annotation file,but is there any code to generate it?There are three papers on how to do this,but no open source.

[gzuidhof]

Hi, I am not sure whether you are talking about nodule annotations or lung segmentation masks.

For nodules there's the src/data_processing/create_same_spacing_data_NODULE.py script which draws circles on each of the slices of its output. This does however save it as a gzipped pickled images, not a very common format. I believe the same is done in the tutorial of the DSB2017 competition on Kaggle. After the DSB2017 I can share a script which outputs the actual (polygon) annotations of the radiologists on the LIDC-IDRI dataset (which is a superset of the LUNA dataset).

For segmentations, I don't think there are sophisticated open source methods available. In my preprocessing tutorial at the DSB2017 challenge I demonstrate a simple approach using region growing and thresholding, perhaps that is a good starting point.

Of course you can also do a machine learning approach. I have tried this on a slice per slice basis with a simple Unet architecture (5x512x512 input), which worked pretty well.

[shartoo]

Yeah,i'm on DSB2017 too and have forked your kernel,that's a excellent work. Your work has made lung segmentation done,but we could take a step further like refinement to generate nodules candidates and nodule segmentation,cause feeding the whole lung structure to CNN sounds not a good idea. Actually ,there do exists methods to do nodule auto segmentation and show a nice sensitivity 94.2% with thresholding ,morphology and some other machine learning methods.Your can visit LUNA2016 data references ,there are papers on the bottom of the site.