使用pix2pix-gan做医学图像合成的时候,如果把nii数据转成png格式会损失很多信息,以为png格式图像的灰度值有256阶,因此直接使用nii的医学图像做输入会更好一点。
但是Pythorch中的Dataloader是不能直接读取nii图像的,因此加一个CreateNiiDataset的类。
先来了解一下pytorch中读取数据的主要途径——Dataset类。在自己构建数据层时都要基于这个类,类似于C 中的虚基类。
自己构建的数据层包含三个部分
代码语言:javascript复制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])
根据自己的需要编写CreateNiiDataset子类:
因为我是基于https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
做pix2pix-gan的实验,数据包含两个部分mr 和 ct,不需要标签,因此上面的 def getitem(self, index):中不需要index这个参数了,类似地,根据需要,加入自己的参数,去掉不需要的参数。
代码语言:javascript复制class CreateNiiDataset(Dataset):
def __init__(self, opt, transform = None, target_transform = None):
self.path1 = opt.dataroot # parameter passing
self.A = 'MR'
self.B = 'CT'
lines = os.listdir(os.path.join(self.path1, 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.path1, self.A, file))
img2 = sitk.ReadImage(os.path.join(self.path1, 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.path1, self.A, file)] # should be a list, with path inside
data['B_paths'] = [os.path.join(self.path1, self.B, file)]
return data
def load_data(self):
return self
def __len__(self):
return len(self.imgs)
注意:最后输出的data是一个字典,里面有四个keys=[‘A’,‘B’,‘A_paths’,‘B_paths’], 一定要注意数据要转成FloatTensor。
其次是data[‘A_paths’] 接收的值是一个list,一定要加[ ] 扩起来,要不然测试存图的时候会有问题,找这个问题找了好久才发现。
然后直接在train.py的主函数里面把数据加载那行改掉就好了
data_loader = CreateNiiDataset(opt) dataset = data_loader.load_data()
Over!
补充知识:nii格式图像存为npy格式
我就废话不多说了,大家还是直接看代码吧!
代码语言:javascript复制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
以上这篇Pytorch 使用 nii数据做输入数据的操作就是小编分享给大家的全部内容了,希望能给大家一个参考。