I want to comment on the fact that the bounding box equal to 141x141x144 instead of 144x144x144 as specified in the documentation. Is this an error in the data or in the documentation? Would the test set have bounding boxes of 141x141x144 also?
what I think there is no error. The boundary box is 141x141x144 pixels/ voxels. But in true dimension wise, it is 144x144x144 mm. Hope it makes sense.
I also found boxes that are not of the specified 144 mm size (141,142,143,144,145). Maybe the order of resampling and cropping (or the other way around) makes this happen. As I use a patch based solution, I do not mind very much about this inconvenience.
Hey Arya, thanks for the comment. If you go to the csv file they provide and make the difference, the size in mm is not 144, but 140.6 or so, which is rounded to 141. Also, as Markjoe is pointing out, there are some other patients that have 142,143 or even 145 (in mm). Since the resampled resolution is 1x1x1, the size in mm is equal to the size in voxel units …
Sorry about that, we didn’t see the discussion before. The test is 144x144x144 but the training bounding boxes were indeed not exactly of this dimension. You can download a csv file with exactly 144x144x144 for the training set here: https://drive.switch.ch/index.php/s/if50vdlbVrM29q0