import os, torch, nibabel as nib, numpy as np from glob import glob from torch.utils.data import DataLoader import torch # function that takes four slices from 4 modalities and stack them def stack_slices(t1c, t2f, t1n, t2w): stacked_slices = np.stack((t1c, t2f, t1n, t2w), axis=0) return stacked_slices # function that takes the segmentation mask and turn it into a 4 channel mask def convert_to_multichannel(mask): """ Convert labels to multi channels based on brats classes: The provided segmentation labels have values of: 1 for NCR (necrotic) 2 for ED (edema) 3 for ET (enhancing tumor) 0 for background. The possible classes are TC (Tumor core == NCR and ET), WT (Whole tumor) , ET (Enhancing tumor) and background. """ results = [] # merge label 1 and label 3 to construct TC results.append( np.logical_or(mask == 1, mask == 3) ) # merge labels 1, 2 and 3 to construct WT results.append( mask != 0 ) # merge label 3 to keep ET results.append( mask == 3 ) # merge label 0 to keep background results.append( mask == 0 ) return np.stack(results, axis=0).astype(np.uint8) class MyIterableDataset(torch.utils.data.IterableDataset): def __init__(self, images_t1c, images_t2f, images_t1n, images_t2w, segs): self.images_t1c = images_t1c self.images_t2f = images_t2f self.images_t1n = images_t1n self.images_t2w = images_t2w self.segs = segs def stream(self): for i in range(self.start, self.end): t1c = nib.load(self.images_t1c[i]).get_fdata() t2f = nib.load(self.images_t2f[i]).get_fdata() t1n = nib.load(self.images_t1n[i]).get_fdata() t2w = nib.load(self.images_t2w[i]).get_fdata() seg = nib.load(self.segs[i]).get_fdata() for j in range(t1c.shape[2]): if np.sum(t1c[:,:,j]) != 0: yield stack_slices(t1c[:,:,j], t2f[:,:,j], t1n[:,:,j], t2w[:,:,j]), convert_to_multichannel(seg[:,:,j]) else: continue def __iter__(self): worker_info = torch.utils.data.get_worker_info() if worker_info is None: self.start = 0 self.end = len(self.images_t1c) else: per_worker = int(np.ceil(len(self.images_t1c) / float(worker_info.num_workers))) self.worker_id = worker_info.id self.start = self.worker_id * per_worker self.end = min(self.start + per_worker, len(self.images_t1c)) return self.stream() def get_MyIterableDataset(folder_name): # check if the folder exists if not os.path.exists(folder_name): raise FileNotFoundError(f"Folder {folder_name} not found,current working directory: {os.getcwd()}") images_t1c = sorted(glob(os.path.join(folder_name, "*/*-t1c.nii.gz"))) images_t2f = sorted(glob(os.path.join(folder_name, "*/*-t2f.nii.gz"))) images_t1n = sorted(glob(os.path.join(folder_name, "*/*-t1n.nii.gz"))) images_t2w = sorted(glob(os.path.join(folder_name, "*/*-t2w.nii.gz"))) segs = sorted(glob(os.path.join(folder_name, "*/*seg.nii.gz"))) number_of_scans = len(images_t1c) print(f"Number of scans: {number_of_scans}") number_of_slices = nib.load(images_t1c[0]).get_fdata().shape[2] # do a train test split train_size = int(0.8 * number_of_scans) train_images_t1c = images_t1c[:train_size] train_images_t2f = images_t2f[:train_size] train_images_t1n = images_t1n[:train_size] train_images_t2w = images_t2w[:train_size] train_segs = segs[:train_size] test_images_t1c = images_t1c[train_size:] test_images_t2f = images_t2f[train_size:] test_images_t1n = images_t1n[train_size:] test_images_t2w = images_t2w[train_size:] test_segs = segs[train_size:] train_dataset = MyIterableDataset(train_images_t1c, train_images_t2f, train_images_t1n, train_images_t2w, train_segs) test_dataset = MyIterableDataset(test_images_t1c, test_images_t2f, test_images_t1n, test_images_t2w, test_segs) return train_dataset, test_dataset