File size: 9,672 Bytes
80df7e4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 | """Pure-torch CT preprocessing: HU scaling, centre cropping and grid patching.
This module is vendored into the Hugging Face Hub repository by `scripts/export_hf.py`,
so it must import nothing beyond `torch` and the standard library.
The functions here reproduce the MONAI pipeline used for pretraining and evaluation
(`ScaleIntensityRanged` -> `LargestMultipleCenterCropd` -> `GridPatchd`) exactly. Two details
are load-bearing and are covered by `tests/test_windowing_parity.py`:
* Patch order is C-order with depth varying fastest, i.e. crop `n` is at grid position
`n = (h * n_w + w) * n_d + d`. This matches both MONAI's `GridPatch` and the coordinates
`RotaryPositionEmbedding` builds, so getting it wrong shifts every embedding silently.
* The centre-crop offset is `start = (S // 2) - (R // 2)`, which is not the same as
`(S - R) // 2` when the crop is odd and the volume even.
"""
from __future__ import annotations
import warnings
from typing import Optional, Sequence, Tuple
import torch
import torch.nn.functional as F
DEFAULT_CROP_SIZE: Tuple[int, int, int] = (128, 128, 64) # (H, W, D)
DEFAULT_HU_RANGE: Tuple[float, float] = (-1000.0, 1000.0)
__all__ = [
'DEFAULT_CROP_SIZE',
'DEFAULT_HU_RANGE',
'scale_intensity_range',
'largest_multiple_crop_size',
'largest_multiple_center_crop',
'grid_patch',
'window_scan',
]
def scale_intensity_range(
x: torch.Tensor,
a_min: float = DEFAULT_HU_RANGE[0],
a_max: float = DEFAULT_HU_RANGE[1],
b_min: float = 0.0,
b_max: float = 1.0,
clip: bool = True,
) -> torch.Tensor:
"""Map intensities from `[a_min, a_max]` onto `[b_min, b_max]`.
Mirrors `monai.transforms.ScaleIntensityRange`. SPECTRE was pretrained with the default
HU window of [-1000, 1000] -> [0, 1] with clipping.
"""
if a_max == a_min:
raise ValueError(f"a_min and a_max must differ, got a_min == a_max == {a_min}.")
x = (x - a_min) / (a_max - a_min)
x = x * (b_max - b_min) + b_min
if clip:
x = torch.clamp(x, min(b_min, b_max), max(b_min, b_max))
return x
def largest_multiple_crop_size(
spatial_shape: Sequence[int],
crop_size: Sequence[int] = DEFAULT_CROP_SIZE,
) -> Tuple[int, ...]:
"""Largest per-axis size that is a whole multiple of `crop_size`.
An axis shorter than its crop keeps its original size (there is no non-zero multiple to
take), matching `spectre.transforms.LargestMultipleCenterCropd`.
"""
if len(spatial_shape) != len(crop_size):
raise ValueError(
f"spatial_shape {tuple(spatial_shape)} and crop_size {tuple(crop_size)} must have "
f"the same number of dimensions."
)
return tuple(
(s // c) * c if s >= c else s
for s, c in zip(spatial_shape, crop_size)
)
def largest_multiple_center_crop(
x: torch.Tensor,
crop_size: Sequence[int] = DEFAULT_CROP_SIZE,
) -> torch.Tensor:
"""Centre-crop a channel-first `(C, H, W, D)` volume to a whole multiple of `crop_size`."""
if x.ndim != 4:
raise ValueError(f"Expected a channel-first (C, H, W, D) volume, got shape {tuple(x.shape)}.")
spatial_shape = tuple(x.shape[1:])
roi_size = largest_multiple_crop_size(spatial_shape, crop_size)
# MONAI's CenterSpatialCrop offset: roi_center - (roi_size // 2), clamped at 0. This is not
# the same as (S - R) // 2 for odd R and even S, so keep the formula verbatim.
slices = []
for size, roi in zip(spatial_shape, roi_size):
start = max(size // 2 - roi // 2, 0)
slices.append(slice(start, start + roi))
return x[(slice(None), *slices)]
def grid_patch(
x: torch.Tensor,
crop_size: Sequence[int] = DEFAULT_CROP_SIZE,
) -> Tuple[torch.Tensor, Tuple[int, int, int]]:
"""Tile a `(C, H, W, D)` volume into non-overlapping crops.
Every spatial axis must already be a whole multiple of `crop_size` (see
`largest_multiple_center_crop`).
Returns:
The crops `(N, C, cH, cW, cD)` and the grid `(n_h, n_w, n_d)` with `N == n_h*n_w*n_d`.
Crop `n` sits at grid position `n = (h * n_w + w) * n_d + d`.
"""
if x.ndim != 4:
raise ValueError(f"Expected a channel-first (C, H, W, D) volume, got shape {tuple(x.shape)}.")
channels = x.shape[0]
spatial_shape = tuple(x.shape[1:])
c_h, c_w, c_d = (int(c) for c in crop_size)
for axis, (size, crop) in enumerate(zip(spatial_shape, (c_h, c_w, c_d))):
if crop <= 0:
raise ValueError(f"crop_size must be positive, got {tuple(crop_size)}.")
if size % crop != 0:
raise ValueError(
f"Axis {'HWD'[axis]} has size {size}, which is not a whole multiple of the crop "
f"size {crop}. Call largest_multiple_center_crop() first, or use window_scan() "
f"which does this for you."
)
n_h, n_w, n_d = (s // c for s, c in zip(spatial_shape, (c_h, c_w, c_d)))
# Depth varies fastest, matching MONAI's GridPatch and the RoPE coordinate order.
crops = x.contiguous().view(channels, n_h, c_h, n_w, c_w, n_d, c_d)
crops = crops.permute(1, 3, 5, 0, 2, 4, 6)
crops = crops.reshape(n_h * n_w * n_d, channels, c_h, c_w, c_d)
return crops, (n_h, n_w, n_d)
def _pad_short_axes(
x: torch.Tensor,
crop_size: Sequence[int],
pad_value: float,
) -> torch.Tensor:
"""Symmetrically pad any spatial axis shorter than its crop up to exactly one crop."""
spatial_shape = tuple(x.shape[1:])
short = [
(axis, size, int(crop))
for axis, (size, crop) in enumerate(zip(spatial_shape, crop_size))
if size < crop
]
if not short:
return x
# F.pad takes the last dimension first, so build the spec back to front: D, W, H.
pad_spec: list = []
for axis in reversed(range(3)):
size, crop = spatial_shape[axis], int(crop_size[axis])
if size < crop:
total = crop - size
pad_spec.extend([total // 2, total - total // 2])
else:
pad_spec.extend([0, 0])
detail = ', '.join(
f"{'HWD'[axis]}={size} < {crop}" for axis, size, crop in short
)
warnings.warn(
f"Scan is smaller than one crop along {detail}. Padding with {pad_value} to reach the "
f"crop size {tuple(int(c) for c in crop_size)}. SPECTRE was not pretrained on padded "
f"volumes, so the resulting embedding is out of distribution. Pass pad_short_axes=False "
f"to raise instead.",
RuntimeWarning,
stacklevel=3,
)
return F.pad(x, pad_spec, mode='constant', value=pad_value)
def window_scan(
x: torch.Tensor,
crop_size: Sequence[int] = DEFAULT_CROP_SIZE,
*,
scale_intensity: bool = True,
hu_range: Tuple[float, float] = DEFAULT_HU_RANGE,
pad_short_axes: bool = True,
pad_value: Optional[float] = None,
) -> Tuple[torch.Tensor, Tuple[int, int, int]]:
"""Turn a whole CT volume into the crops SPECTRE's backbone consumes.
Applies, in the order used for pretraining: HU scaling -> centre crop to a whole multiple of
`crop_size` -> grid patching.
Args:
x: A `(H, W, D)` or `(C, H, W, D)` volume. In raw Hounsfield Units unless
`scale_intensity=False`.
crop_size: Spatial size `(H, W, D)` of one crop.
scale_intensity: Map `hu_range` onto [0, 1] with clipping. Turn off if `x` is already
normalised.
hu_range: The `(a_min, a_max)` HU window. SPECTRE was pretrained with (-1000, 1000).
pad_short_axes: Pad axes shorter than one crop instead of raising.
pad_value: Value to pad with. Defaults to the scaled air value (0.0) when
`scale_intensity` is on, and to `x.min()` otherwise.
Returns:
The crops `(N, C, cH, cW, cD)` and the grid `(n_h, n_w, n_d)`.
"""
if x.ndim == 3:
x = x.unsqueeze(0)
elif x.ndim != 4:
raise ValueError(
f"Expected a (H, W, D) or (C, H, W, D) volume, got shape {tuple(x.shape)}."
)
if len(crop_size) != 3:
raise ValueError(f"crop_size must have 3 elements (H, W, D), got {tuple(crop_size)}.")
if scale_intensity:
if float(x.min()) > -100.0:
warnings.warn(
f"Input does not look like Hounsfield Units (min={float(x.min()):.1f}). It will "
f"be rescaled from {hu_range} to [0, 1], which double-scales already-normalised "
f"data. Pass scale_intensity=False if the volume is already normalised.",
RuntimeWarning,
stacklevel=2,
)
x = scale_intensity_range(x, a_min=hu_range[0], a_max=hu_range[1])
if pad_value is None:
pad_value = 0.0 if scale_intensity else float(x.min())
too_short = [
f"{'HWD'[axis]}={int(size)} < {int(crop)}"
for axis, (size, crop) in enumerate(zip(x.shape[1:], crop_size))
if size < crop
]
if too_short:
if not pad_short_axes:
raise ValueError(
f"Scan is smaller than one crop along {', '.join(too_short)}. Pass "
f"pad_short_axes=True to pad with air, or supply a scan covering at least "
f"{tuple(int(c) for c in crop_size)} voxels."
)
x = _pad_short_axes(x, crop_size, pad_value)
x = largest_multiple_center_crop(x, crop_size)
return grid_patch(x, crop_size)
|