list index out of range while testing. about medicaldetectiontoolkit

[pfjaeger]

hi thanks for reaching out. can you post your configs please?

[cammu00]

Thanks for response. I used brat18 dataset with T1, T1ce, flair, and T2 . In preprocessing, I created 4ch data with this 4 files. So, 4ch dataset was combined with this 4 data file. Tumor segment file was used for ROIs. I applied mask rcnn for data training. Training stage was completed well.
I attached config.py files which was used for this.
config.txt
Best,
David.

[pfjaeger]

hey yes so like you assumed, the problem lies withing your channel dimension. As LIDC has only one channel, the LIDC data_loader does is not written for data with a multiple channels. So the image you pass to get_patch_crop_coords has 4 dimensions instead of the expected 3. The fixes you need to make should be easy to implement, just check to shift the dimensions by one where needed in the PatientBatchIterator class.

[pfjaeger]

I will write this class more generic to avoid the problem in the future, but it might take a couple of days till I find the time. Good Luck!

[cammu00]

Thank you for your comments. I'll try it for this. I have a wondering. If I do test with shifting the dimensions by one from 4ch data, is there any difference with each separately trained data?(e.g. when I do train each dataset(T1, T1ce, T2, Flair) as 1ch dataset)

Thank you for your help.
David.

[pfjaeger]

Generally you would want to train multi-modal models so as to grasp the correlations between modalities in your model paramters. In other words, if you train separately your model will not learn relational semantics between modalities.