monai学习
任小云
MetaTensor
这是monai特有的类型
metaTensor=tensor+metaObj
也是原本的tensor，带上meta data，例如affine，spacing，direction，origin
#a: MetaTensor
a.meta['spatial_orgin_shape']
a.meta['affine']
a.meta['original_affine']
a.affine
两个MetaTensor相加，如果is_batch为False，则复制第一个的metadata
数据增强
一般都在monai.transforms底下
有两种版本，一种是增强一张图片，另一种是增强一组图片（例如图片+标签），这种结尾会带一个d
增强完了类型都会变成metatensor
结尾带d的需要传入dict，会返回dict
再训练的时候，可以考虑
data['image'].as_tensor()
来转为普通的tensor，减少计算开销
加载图片
class monai.transforms.LoadImaged(keys, reader=None, dtype=<class 'numpy.float32'>,
meta_keys=None, meta_key_postfix='meta_dict', overwriting=False, image_only=False,
ensure_channel_first=False, simple_keys=False, prune_meta_pattern=None, prune_meta_sep='.',
allow_missing_keys=False, *args, **kwargs)
例如
trans=LoadImaged(['flair', 't1', 't1ce', 't2', 'seg'], image_only=True, allow_missing_keys=True)
trans({'flair':'/mnt/data/xxx', 't1':'/mnt/data/xxxx'})
其中image_only表示只返回图片，False的话，会额外往返回的dict中添加key_meta_dict
allow_missing_keys表示有存在的key也不会报错
读取图片的reader，默认如下
(nii, nii.gz -> NibabelReader), (png, jpg, bmp -> PILReader), (npz, npy -> NumpyReader), (dcm, DICOM series and others -> ITKReader).
类型转换
class monai.transforms.CastToTyped(keys, dtype=<class 'numpy.float32'>,
allow_missing_keys=False)
例如
CastToTyped(keys=['seg'], dtype=torch.long, allow_missing_keys=True)
添加通道
class monai.transforms.EnsureChannelFirstd(keys, meta_keys=None, meta_key_postfix='meta_dict', strict_check=True, allow_missing_keys=False, channel_dim=None)
随机旋转
class monai.transforms.RandRotated(keys, range_x=0.0, range_y=0.0, range_z=0.0, prob=0.1,
keep_size=True, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER,
align_corners=False, dtype=<class 'numpy.float32'>, allow_missing_keys=False)
例如
RandRotated(keys=['image', 'seg'], range_x=10, range_y=10, range_z=10, allow_missing_keys=True), # (-10, 10)
需要注意的是，这个默认的只有float能旋转，而且返回是float，所以可以考虑跟一个转类型的
翻转
class monai.transforms.RandFlipd(keys, prob=0.1, spatial_axis=None, allow_missing_keys=False)
需要注意，这个翻转的话，是指定的轴一起翻转
所以可以考虑多写几个
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=0, allow_missing_keys=True),
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=1, allow_missing_keys=True),
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=2, allow_missing_keys=True)
pad
class monai.transforms.Padd(keys, padder, mode=PytorchPadMode.CONSTANT, allow_missing_keys=False)
这个似乎必须带有通道维度，并且的pad形状要跟带通道后的形状一样
clip
这个确实没找到，但是找到了一种代替的方法
# clip(-325, 325)
transforms.ThresholdIntensityd(keys=['image'], threshold=-325, above=True, cval=-325), # max(img, -325)
transforms.ThresholdIntensityd(keys=['image'], threshold=325, above=False, cval=325), # min(img, 325)
z-score
class monai.transforms.NormalizeIntensityd(keys, subtrahend=None, divisor=None, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>, allow_missing_keys=False)
例如
transforms.NormalizeIntensityd(keys=['image']), # z-score
分位点
from monai.transforms import MapTransform
class Percentile(MapTransform):
def __init__(
self,
keys: KeysCollection,
lower_percentile: float = 0.,
upper_percentile: float = 100.,
allow_missing_keys: bool = False,
) -> None:
"""
Args:
keys: keys of the corresponding items to be transformed.
See also: :py:class:`monai.transforms.compose.MapTransform`
lower_percentile: lower percentile(0-100)
upper_percentile: upper percentile(0-100)
allow_missing_keys: don't raise exception if key is missing.
"""
MapTransform.__init__(self, keys, allow_missing_keys)
self.lower_percentile = lower_percentile / 100.
self.upper_percentile = upper_percentile / 100.
def __call__(self, data):
# the first dim of data should be the channel(CHW[D])
d = dict(data)
for key in self.key_iterator(d):
images = data[key]
lower = torch.quantile(images, self.lower_percentile)
upper = torch.quantile(images, self.upper_percentile)
images = torch.clip(images, lower, upper)
d[key] = images
return d
完整的例子
#!/usr/bin/env python
# _*_ coding:utf-8 _*_
import numpy as np
import torch
from monai.config import KeysCollection
from monai.data import NibabelWriter
from monai.transforms import MapTransform, Compose, LoadImaged, CastToTyped, EnsureChannelFirstd, RandSpatialCropd, \
RandRotated, RandScaleIntensityd, RandShiftIntensityd, RandFlipd, Pad, Padd
data = {
'flair': '/mnt/data/datasets/BraTS_2018/MICCAI_BraTS_2018_Data_Training/HGG/Brats18_2013_2_1/Brats18_2013_2_1_flair.nii.gz',
'seg': '/mnt/data/datasets/BraTS_2018/MICCAI_BraTS_2018_Data_Training/HGG/Brats18_2013_2_1/Brats18_2013_2_1_seg.nii.gz',
't1': '/mnt/data/datasets/BraTS_2018/MICCAI_BraTS_2018_Data_Training/HGG/Brats18_2013_2_1/Brats18_2013_2_1_t1.nii.gz',
't1ce': '/mnt/data/datasets/BraTS_2018/MICCAI_BraTS_2018_Data_Training/HGG/Brats18_2013_2_1/Brats18_2013_2_1_t1ce.nii.gz',
't2': '/mnt/data/datasets/BraTS_2018/MICCAI_BraTS_2018_Data_Training/HGG/Brats18_2013_2_1/Brats18_2013_2_1_t2.nii.gz'
}
class StackImagesd(MapTransform):
"""
stack images
add the result in the dict with the key 'image'
"""
def __call__(self, data):
d = dict(data)
result = []
for key in self.key_iterator(d):
result.append(d[key])
d['image'] = torch.stack(result, dim=0) # (H, W, D)->(4, H, W, D)
return d
class PercentileAndZScored(MapTransform):
def __init__(
self,
keys: KeysCollection,
lower_percentile: float = 0.,
upper_percentile: float = 100.,
allow_missing_keys: bool = False,
) -> None:
"""
Args:
keys: keys of the corresponding items to be transformed.
See also: :py:class:`monai.transforms.compose.MapTransform`
lower_percentile: lower percentile(0-100)
upper_percentile: upper percentile(0-100)
allow_missing_keys: don't raise exception if key is missing.
"""
MapTransform.__init__(self, keys, allow_missing_keys)
self.lower_percentile = lower_percentile / 100.
self.upper_percentile = upper_percentile / 100.
def __call__(self, data):
# the first dim of data should be the channel(CHW[D])
d = dict(data)
for key in self.key_iterator(d):
images = data[key]
C = images.size(0)
mask = images.sum(0) > 0 # brain
for k in range(C):
x = images[k, ...]
y = x[mask]
lower = torch.quantile(y, self.lower_percentile)
upper = torch.quantile(y, self.upper_percentile)
x[mask & (x < lower)] = lower
x[mask & (x > upper)] = upper
# z-score across the brain
y = x[mask]
x -= y.mean()
x /= y.std()
images[k, ...] = x
d[key] = images
return d
if __name__ == '__main__':
transform = Compose([
LoadImaged(['flair', 't1', 't1ce', 't2', 'seg'], image_only=True, allow_missing_keys=True),
CastToTyped(keys=['seg'], dtype=torch.long, allow_missing_keys=True),
EnsureChannelFirstd(keys=['seg'], allow_missing_keys=True),
StackImagesd(keys=['flair', 't1', 't1ce', 't2']), # add ['image']
PercentileAndZScored(keys=['image'], lower_percentile=0.2, upper_percentile=99.8),
RandSpatialCropd(keys=['image', 'seg'], roi_size=(128, 128, 128), random_size=False, allow_missing_keys=True),
RandRotated(keys=['image', 'seg'], range_x=10, range_y=10, range_z=10, allow_missing_keys=True), # (-10, 10)
CastToTyped(keys=['seg'], dtype=torch.long, allow_missing_keys=True),
RandScaleIntensityd(keys=['image'], factors=0.1), # (-0.1, 0.1), img * (1 + scale)
RandShiftIntensityd(keys=['image'], offsets=0.1), # (-0.1, 0.1), img + offset
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=0, allow_missing_keys=True),
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=1, allow_missing_keys=True),
RandFlipd(keys=["image", "seg"], prob=0.5, spatial_axis=2, allow_missing_keys=True),
CastToTyped(keys=['seg'], dtype=torch.long, allow_missing_keys=True),
Padd(keys=["image", "seg"], padder=Pad([(0, 0), (0, 0), (0, 0), (0, 5)]), allow_missing_keys=True),
])
result = transform(data)
print(result['seg'].shape)
保存
nii
默认是会重采样的，重采样的依据是metadata的spatial_shape，affine，original_affine
writer = NibabelWriter(output_dtype=torch.uint8)
writer.set_data_array(mask.squeeze(0), channel_dim=0) # (C, H, W, D)
writer.set_metadata(mask.meta, resample=True, mode='nearest')
writer.write('faQ.nii.gz', verbose=True)
不重采样
下面这段，会把affine直接写入数据中，而忽略原来的affine
writer = NibabelWriter(output_dtype=torch.uint8)
writer.set_data_array(mask, channel_dim=0)# (C, H, W, D)
writer.set_metadata({
'spatial_shape': result['image'].meta['spatial_shape'],
'affine': result['image'].meta['original_affine'],
'original_affine': result['image'].meta['original_affine']
}, resample=False, mode='nearest')
writer.write('faQ.nii.gz')
Loss
都在monai.losses里
常见有：dice，focal loss， GeneralizedDiceLoss
metric
需要注意的是，很多metric要求传进去one_hot形式
具体可以看下面的例子
import argparse
import os
from glob import glob
import torch
from monai import transforms
from monai.metrics import DiceMetric
from monai.networks import one_hot
from monai.utils import MetricReduction
from tqdm.contrib import tzip
metric = DiceMetric(include_background=False)
loader = transforms.LoadImage(image_only=True, dtype=torch.uint8)
predict_paths = get_data(args.predict_path)
gt_paths = get_data(args.gt_path)
CLASS_NUMBER = 5
for predict_path, gt_path in tzip(predict_paths, gt_paths):
predict = loader(predict_path).unsqueeze(0)
gt = loader(gt_path).unsqueeze(0)
predict = one_hot(predict, CLASS_NUMBER, dtype=torch.uint8, dim=0).unsqueeze(0)
gt = one_hot(gt, CLASS_NUMBER, dtype=torch.uint8, dim=0).unsqueeze(0)
dice = metric(predict, gt) # (B, C) =(B,5-1) (include_background=False)
print('mean_dice: liver={:.2f}, kidney={:.2f}, spleen={:.2f}, pancreas={:.2f}'.format(
*metric.aggregate(MetricReduction.MEAN_BATCH)))
print('mean_dice: {:.2f}'.format(metric.aggregate(MetricReduction.MEAN).item()))
类似资料：