FabianIsensee
commented
5 years ago Hi there, I am happy to hear that nnU-net works well for your data!
After training and test phase, how could I get those quantitative metrics results? I try to use the evaluator.py and collect_results_files.py, but they do not work.
What is it you are trying to do exactly? Do you want to know what the results on your cross-validation are? Do you want to know what models to use for test set prediction?
I don't know whether the Data with different channels can input into the network for training simultaneously, namely, if i have some data format like(328, 400, 400) and (290, 400, 400), can I train them in the same network?
This sounds like you have 3D data. In this case 328 and 290 are NOT channels (as in color channels) but spatial dimensions. You can process heterogeneous data sizes. That is no problem at all ;-)
By the way, it would be wonderful if their have a data augmentation Implementation with nifti or other 3D medical data format.
Data loading and augmentation with nifti is a VERY bad idea. I use numpy arrays specifically because I can read them in a very particular way that is (as far as I know) not possible with niftis. Imagine your data is 512x512x512 voxels large (like an abdominal CT) and your patch size is 128x128x128. That means that your patch size covers 1/64 of a case. If you use nifti, you would need to read the entire case (if it is compressed you also need to decompress it) and then you throw away 63/64th of your work because you only need 1/64. That is a lot of wasted I/O and CPU time. Numpy arrays can be read as memmaps, meaning that if I need 128x128x128 from a specific location that is all I need to read, not the entire file. This problem becomes even more apparent when running 2D networks. Imagine having to read multiple >2GB files, one for each slice in the batch. Infeasible. The data augmentation as it is is highly optimized. nnU-Net takes niftis as input at the very start and produces niftis at the very end. I don't see a good reason why using niftis in the middle would be required :-)
Best, Fabian
li-pengcheng
Hi @FabianIsensee , I'm learned and inspired a lot from your work! And I trained on my own data and got some promising results using your code. But there still some questions bothers me: 1) After training and test phase, how could I get those quantitative metrics results? I try to use the
evaluator.pyandcollect_results_files.py, but they do not work. 2) I don't know whether the Data with different channels can input into the network for training simultaneously, namely, if i have some data format like(328, 400, 400) and (290, 400, 400), can I train them in the same network? 3) By the way, it would be wonderful if their have a data augmentation Implementation with nifti or other 3D medical data format. Forgive my poor English, have a good day!