Hello,
I have 3D CT images (and the corresponding segmentation masks), from each of which I would like to retrieve a variable number of 3D patches, centered around a ROI (based on the labels of the segmentation masks).
I would like to train on these patches at random, independently of which CT they came from (i.e. shuffle in the DataLoader should be True).
For this purpose, the Dataset's __getitem__ should give me a single patch + patch_mask pair.
I can do this by loading the same full_size CT image every time I want to crop a patch from it, like so:
class myDataset(Dataset):
def __init__(self, df, size):
# df is a dataframe with the paths to the full-size CT's and masks, like so:
# index | ID | ID_mask | label
# 0 'path_to_CT1' 'path_to_CT1_mask' 1
# 1 'path_to_CT1' 'path_to_CT1_mask' 2
# 2 'path_to_CT2' 'path_to_CT2_mask' 1
# 3 'path_to_CT2' 'path_to_CT2_mask' 2
# 4 'path_to_CT2' 'path_to_CT2_mask' 3
# 5 'path_to_CT2' 'path_to_CT2_mask' 4
# 6 'path_to_CT3' 'path_to_CT3_mask' 1
# ... .... .... ....
self.df = df
self.size = size # size of the patches
def __len__(self):
return len(self.df)
def __getitem__(self, idx):
path_to_CT = self.df.at[idx,'ID']
path_to_CT_mask = self.df.at[idx,'ID_mask']
CT = sitk.ReadImage(path_to_CT) # loads entire CT (512x512x512)
CT_mask = sitk.ReadImage(path_to_CT_mask) # loads entire CT_mask (512x512x512)
label = self.df.at[idx,'label']
# crop a single patch + patch mask around label (ROI) of size e.g. 128x128x128
patch, patch_mask = crop_roi(CT, CT_mask , label, self.size)
return {'imgs': patch, 'masks': patch_mask}
I can also retrieve all of the patches from the same CT at once and store them for example in a dictionary in the __init__ and then access them in the next __getitem__. However, this would either force me to go through all of the patches in order (with no shuffle) or store all of the patches of all of the CT’s in a dictionary which doesn’t work well in terms of memory.
I’ve read in a previous post that I can “stack the crops into the batch dimension”. I’ve tried it, but because I have a variable number of patches per image, I don’t think it works.
My question: is there a way to avoid loading the same CT every time I need a patch from it (considering that the number of patches of each CT can be up to 150)?