1
|
|
|
import torch |
2
|
|
|
from torchio.data.image import ScalarImage |
3
|
|
|
from ....data.subject import Subject |
4
|
|
|
from ...intensity_transform import IntensityTransform |
5
|
|
|
from typing import Optional |
6
|
|
|
from math import ceil |
7
|
|
|
|
8
|
|
|
|
9
|
|
|
class Projection(IntensityTransform): |
10
|
|
|
"""Project intensities along a given axis, possibly with sliding slabs. |
11
|
|
|
|
12
|
|
|
Args: |
13
|
|
|
axis: Possible inputs are ``'Left'``, ``'Right'``, ``'Anterior'``, |
14
|
|
|
``'Posterior'``, ``'Inferior'``, ``'Superior'``. Lower-case |
15
|
|
|
versions and first letters are also valid, as only the first |
16
|
|
|
letter will be used. |
17
|
|
|
slab_thickness: Thickness of slab projections. In other words, the |
18
|
|
|
number of voxels in the ``axis`` dimension to project across. |
19
|
|
|
If ``None``, the projection will be done across the entire span of |
20
|
|
|
the ``axis`` dimension (i.e. ``axis`` dimension will be reduced to |
21
|
|
|
1). |
22
|
|
|
stride: Number of voxels to stride along the ``axis`` dimension between |
23
|
|
|
slab projections. |
24
|
|
|
projection_type: Type of intensity projection. Possible inputs are |
25
|
|
|
``'max'`` (the default), ``'min'``, ``'mean'``, ``'median'``, or |
26
|
|
|
``'quantile'``. If ``'quantile'`` is used, ``q`` must also be |
27
|
|
|
supplied. |
28
|
|
|
q: Quantile to use for intensity projections. This argument is required |
29
|
|
|
if ``projection_type`` is ``'quantile'`` and is silently ignored |
30
|
|
|
otherwise. |
31
|
|
|
full_slabs_only: Boolean. Should projections be done only for slabs |
32
|
|
|
that are ``slab_thickness`` thick? Default is ``True``. |
33
|
|
|
If ``False``, some slabs may not be ``slab_thickness`` thick |
34
|
|
|
depending on the size of the image, slab thickness, and stride. |
35
|
|
|
|
36
|
|
|
Example: |
37
|
|
|
>>> import torchio as tio |
38
|
|
|
>>> sub = tio.datasets.Colin27() |
39
|
|
|
>>> axial_mips = tio.Projection("S", slab_thickness=20) |
40
|
|
|
>>> sub_t = axial_mips(sub) |
41
|
|
|
>>> sub_t.t1.plot() |
42
|
|
|
""" |
43
|
|
|
def __init__( |
44
|
|
|
self, |
45
|
|
|
axis: str, |
46
|
|
|
slab_thickness: Optional[int] = None, |
47
|
|
|
stride: Optional[int] = 1, |
48
|
|
|
projection_type: Optional[str] = 'max', |
49
|
|
|
q: Optional[float] = None, |
50
|
|
|
full_slabs_only: Optional[bool] = True, |
51
|
|
|
**kwargs |
52
|
|
|
): |
53
|
|
|
super().__init__(**kwargs) |
54
|
|
|
self.args_names = ( |
55
|
|
|
'axis', 'slab_thickness', 'stride', |
56
|
|
|
'projection_type', 'q', 'full_slabs_only' |
57
|
|
|
) |
58
|
|
|
self.axis = axis |
59
|
|
|
self.slab_thickness = slab_thickness |
60
|
|
|
self.stride = stride |
61
|
|
|
self.projection_type = projection_type |
62
|
|
|
self.q = q |
63
|
|
|
self.full_slabs_only = full_slabs_only |
64
|
|
|
self.projection_fun = self.get_projection_function() |
65
|
|
|
|
66
|
|
|
def get_projection_function(self): |
67
|
|
|
if self.projection_type == 'max': |
68
|
|
|
projection_fun = torch.amax |
69
|
|
|
elif self.projection_type == 'min': |
70
|
|
|
projection_fun = torch.amin |
71
|
|
|
elif self.projection_type == 'mean': |
72
|
|
|
projection_fun = torch.mean |
73
|
|
|
elif self.projection_type == 'median': |
74
|
|
|
projection_fun = torch.median |
75
|
|
|
elif self.projection_type == 'quantile': |
76
|
|
|
projection_fun = torch.quantile |
77
|
|
|
self.validate_quantile() |
78
|
|
|
else: |
79
|
|
|
message = ( |
80
|
|
|
'`projection_type` must be one of "max", "min", "mean",' |
81
|
|
|
' "median", or "quantile".' |
82
|
|
|
) |
83
|
|
|
raise ValueError(message) |
84
|
|
|
return projection_fun |
85
|
|
|
|
86
|
|
|
def validate_quantile(self): |
87
|
|
|
message = ( |
88
|
|
|
'For `projection_type="quantile"`, `q` must be a scalar value' |
89
|
|
|
f'in the range [0, 1], not {self.q}.' |
90
|
|
|
) |
91
|
|
|
if self.q is None: |
92
|
|
|
raise ValueError(message) |
93
|
|
|
elif 0 <= self.q <= 1: |
94
|
|
|
pass |
95
|
|
|
else: |
96
|
|
|
raise ValueError(message) |
97
|
|
|
|
98
|
|
|
def apply_transform(self, subject: Subject) -> Subject: |
99
|
|
|
for image in self.get_images(subject): |
100
|
|
|
self.apply_projection(image) |
101
|
|
|
return subject |
102
|
|
|
|
103
|
|
|
def apply_projection(self, image: ScalarImage) -> None: |
104
|
|
|
self.axis_index = image.axis_name_to_index(self.axis) |
105
|
|
|
self.axis_span = image.shape[self.axis_index] |
106
|
|
|
if self.slab_thickness is None: |
107
|
|
|
self.slab_thickness = self.axis_span |
108
|
|
|
elif self.slab_thickness > self.axis_span: |
109
|
|
|
self.slab_thickness = self.axis_span |
110
|
|
|
image.set_data(self.projection(image.data)) |
111
|
|
|
|
112
|
|
|
def projection(self, tensor: torch.Tensor) -> torch.Tensor: |
113
|
|
|
if self.projection_type in ['mean', 'quantile']: |
114
|
|
|
tensor = tensor.to(torch.float) |
115
|
|
|
|
116
|
|
|
if self.full_slabs_only: |
117
|
|
|
start_index = 0 |
118
|
|
|
num_slabs = 0 |
119
|
|
|
while start_index + self.slab_thickness <= self.axis_span: |
120
|
|
|
num_slabs += 1 |
121
|
|
|
start_index += self.stride |
122
|
|
|
else: |
123
|
|
|
num_slabs = ceil(self.axis_span / self.stride) |
124
|
|
|
|
125
|
|
|
slabs = [] |
126
|
|
|
start_index = 0 |
127
|
|
|
end_index = start_index + self.slab_thickness |
128
|
|
|
|
129
|
|
|
for _ in range(num_slabs): |
130
|
|
|
slab_indices = torch.tensor(list(range(start_index, end_index))) |
131
|
|
|
slab = tensor.index_select(self.axis_index, slab_indices) |
132
|
|
|
if self.projection_type == 'median': |
133
|
|
|
projected, _ = self.projection_fun( |
134
|
|
|
slab, dim=self.axis_index, keepdim=True) |
135
|
|
|
elif self.projection_type == 'quantile': |
136
|
|
|
projected = self.projection_fun( |
137
|
|
|
slab, q=self.q, dim=self.axis_index, keepdim=True) |
138
|
|
|
else: |
139
|
|
|
projected = self.projection_fun( |
140
|
|
|
slab, dim=self.axis_index, keepdim=True) |
141
|
|
|
slabs.append(projected) |
142
|
|
|
start_index += self.stride |
143
|
|
|
end_index = start_index + self.slab_thickness |
144
|
|
|
if end_index > self.axis_span: |
145
|
|
|
end_index = self.axis_span |
146
|
|
|
|
147
|
|
|
return torch.cat(slabs, dim=self.axis_index) |
148
|
|
|
|