fixed a bug

• the transformation of patch will change the content of original image and label volume! This was fixed, and now the training looks stable.

used to train a Tbar detection net

• replaced TorchIO with our own patch provider. It is much more faster. over 100x speed up.

initial working version for T-bar prediction

• use TorchIO for data augmentation. The patch sampling seems pretty slow. it tasks about 6-7 seconds to sample a patch without any augmentation! With some augmentation, the time could be increased to 100-200 seconds. Besides, the dependency of SimpleITK needs to be installed using conda rather than pip. I plan to remove this dependency in the next release.