详解Pytorch如何使用nii数据做输入数据操作

class Dataset(object):
"""An abstract class representing a Dataset.
All other datasets should subclass it. All subclasses should override
``__len__``, that provides the size of the dataset, and ``__getitem__``,
supporting integer indexing in range from 0 to len(self) exclusive.
"""
def __getitem__(self, index):
 raise NotImplementedError
def __len__(self):
 raise NotImplementedError
def __add__(self, other):
 return ConcatDataset([self, other])

class CreateNiiDataset(Dataset):
 def __init__(self, opt, transform = None, target_transform = None):
  self.path2 = opt.dataroot # parameter passing
  self.A = 'MR' 
  self.B = 'CT'
  lines = os.listdir(os.path.join(self.path2, self.A))
  lines.sort()
  imgs = []
  for line in lines:
   imgs.append(line)
  self.imgs = imgs
  self.transform = transform
  self.target_transform = target_transform

 def crop(self, image, crop_size):
  shp = image.shape
  scl = [int((shp[0] - crop_size[0]) / 2), int((shp[1] - crop_size[1]) / 2)]
  image_crop = image[scl[0]:scl[0] + crop_size[0], scl[1]:scl[1] + crop_size[1]]
  return image_crop

 def __getitem__(self, item):
  file = self.imgs[item]
  img1 = sitk.ReadImage(os.path.join(self.path2, self.A, file))
  img2 = sitk.ReadImage(os.path.join(self.path2, self.B, file))
  data1 = sitk.GetArrayFromImage(img1)
  data2 = sitk.GetArrayFromImage(img2)

  if data1.shape[0] != 256:
   data1 = self.crop(data1, [256, 256])
   data2 = self.crop(data2, [256, 256])
  if self.transform is not None:
   data1 = self.transform(data1)
   data2 = self.transform(data2)

  if np.min(data1)<0:
   data1 = (data1 - np.min(data1))/(np.max(data1)-np.min(data1))

  if np.min(data2)<0:
   #data2 = data2 - np.min(data2)
   data2 = (data2 - np.min(data2))/(np.max(data2)-np.min(data2))

  data = {}
  data1 = data1[np.newaxis, np.newaxis, :, :]
  data1_tensor = torch.from_numpy(np.concatenate([data1,data1,data1], 1))
  data1_tensor = data1_tensor.type(torch.FloatTensor)
  data['A'] = data1_tensor # should be a tensor in Float Tensor Type

  data2 = data2[np.newaxis, np.newaxis, :, :]
  data2_tensor = torch.from_numpy(np.concatenate([data2,data2,data2], 1))
  data2_tensor = data2_tensor.type(torch.FloatTensor)
  data['B'] = data2_tensor # should be a tensor in Float Tensor Type
  data['A_paths'] = [os.path.join(self.path2, self.A, file)] # should be a list, with path inside
  data['B_paths'] = [os.path.join(self.path2, self.B, file)]
  return data

 def load_data(self):
  return self

 def __len__(self):
  return len(self.imgs)

import nibabel as nib
import os
import numpy as np
 
img_path = '/home/lei/train/img/'
seg_path = '/home/lei/train/seg/'
saveimg_path = '/home/lei/train/npy_img/'
saveseg_path = '/home/lei/train/npy_seg/'
 
img_names = os.listdir(img_path)
seg_names = os.listdir(seg_path)
 
for img_name in img_names:
 print(img_name)
 img = nib.load(img_path + img_name).get_data() #载入
 img = np.array(img)
 np.save(saveimg_path + str(img_name).split('.')[0] + '.npy', img) #保存
 
for seg_name in seg_names:
 print(seg_name)
 seg = nib.load(seg_path + seg_name).get_data()
 seg = np.array(seg)
 np.save(saveseg_path + str(seg_name).split('.')[0] + '.npy