File size: 11,871 Bytes
041602e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
"""
MultiMAE 3D Utility Functions
- Patchify / Unpatchify
- Patch shuffling for masking
- 3D sinusoidal positional embeddings
- Dirichlet masking with missing modality support
"""

from typing import Union, Tuple, Dict, List

import torch
import torch.nn as nn
from torch.distributions import Dirichlet
from einops import rearrange


def to_3tuple(x):
    if isinstance(x, (list, tuple)):
        assert len(x) == 3
        return tuple(x)
    return (x, x, x)


def calc_patchified_dim(
    img_size: Union[int, Tuple[int, int, int]],
    patch_size: Union[int, Tuple[int, int, int]],
) -> Tuple[int, int, int]:
    img_size = to_3tuple(img_size)
    patch_size = to_3tuple(patch_size)
    return tuple(img_size[i] // patch_size[i] for i in range(3))


def patchify(
    image: torch.Tensor,
    patch_size: Union[int, Tuple[int, int, int]],
) -> torch.Tensor:
    """
    Convert image to patches.
    image: [B, C, D, H, W]
    returns: [B, num_patches, C, pd, ph, pw]
    """
    patch_size = to_3tuple(patch_size)
    img_size = image.shape[-3:]
    patchified_dim = calc_patchified_dim(img_size, patch_size)
    patches = rearrange(
        image,
        "b c (nd pd) (nh ph) (nw pw) -> b (nd nh nw) c pd ph pw",
        pd=patch_size[0],
        ph=patch_size[1],
        pw=patch_size[2],
        nd=patchified_dim[0],
        nh=patchified_dim[1],
        nw=patchified_dim[2],
    )
    return patches


def unpatchify(
    patches: torch.Tensor,
    img_size: Union[int, Tuple[int, int, int]],
    patch_size: Union[int, Tuple[int, int, int]],
) -> torch.Tensor:
    """
    Convert patches back to image.
    patches: [B, num_patches, C, pd, ph, pw]
    returns: [B, C, D, H, W]
    """
    patch_size = to_3tuple(patch_size)
    img_size = to_3tuple(img_size)
    patchified_dim = calc_patchified_dim(img_size, patch_size)
    image = rearrange(
        patches,
        "b (nd nh nw) c pd ph pw -> b c (nd pd) (nh ph) (nw pw)",
        pd=patch_size[0],
        ph=patch_size[1],
        pw=patch_size[2],
        nd=patchified_dim[0],
        nh=patchified_dim[1],
        nw=patchified_dim[2],
    )
    return image


def shuffle_patches(
    patches: torch.Tensor,
    permutations: torch.Tensor = None,
    mask_probs: torch.Tensor = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Shuffle patches along the patch dimension.

    When mask_probs is None: uniform random shuffle.
    When mask_probs is provided: Gumbel-top-k weighted shuffle.
      Patches with higher mask_probs end up at higher indices (masked),
      patches with lower mask_probs end up at lower indices (visible).

    Args:
        patches: [B, N, ...]
        permutations: optional pre-computed permutation indices [B, N]
        mask_probs: optional [N] per-patch masking probability (sums to 1)

    Returns:
        (shuffled_patches, perm_indices)
    """
    batch_size, num_patches = patches.shape[0], patches.shape[1]
    if permutations is not None:
        perm_idx = permutations
    else:
        rand = torch.rand(batch_size, num_patches, device=patches.device)

        if mask_probs is not None:
            # Gumbel-top-k trick for weighted sampling without replacement.
            # key_i = log(p_i) + Gumbel(0,1)_i
            # Top-k of keys = sample from Multinomial(p, k)
            # After ascending argsort: low keys → visible, high keys → masked.
            mask_probs = mask_probs.to(patches.device)
            gumbel = -torch.log(-torch.log(rand.clamp(1e-20, 1.0 - 1e-20)))
            log_probs = torch.log(mask_probs.clamp(min=1e-20))  # [N]
            keys = gumbel + log_probs.unsqueeze(0)  # [B, N]
            perm_idx = torch.argsort(keys, dim=1)
        else:
            perm_idx = torch.argsort(rand, dim=1)

    shuffled = patches[torch.arange(batch_size, device=patches.device)[:, None], perm_idx]
    return shuffled, perm_idx


def unshuffle_patches(
    patches: torch.Tensor,
    perm_idx: torch.Tensor,
) -> torch.Tensor:
    """
    Inverse of shuffle_patches.
    """
    batch_size = patches.shape[0]
    inv_idx = torch.argsort(perm_idx, dim=1)
    return patches[torch.arange(batch_size, device=patches.device)[:, None], inv_idx]


def build_3d_sincos_position_embedding(
    grid_size: Tuple[int, int, int],
    embed_dim: int,
    temperature: float = 10000.0,
) -> nn.Parameter:
    """
    Build 3D sinusoidal positional embedding.
    returns: [1, num_patches, embed_dim] (frozen parameter)
    """
    grid_size = to_3tuple(grid_size)
    h, w, d = grid_size

    assert embed_dim % 6 == 0, \
        f"embed_dim ({embed_dim}) must be divisible by 6 for 3D sincos pos embed"

    pos_dim = embed_dim // 6
    omega = torch.arange(pos_dim, dtype=torch.float32) / pos_dim
    omega = 1.0 / (temperature ** omega)

    grid_h = torch.arange(h, dtype=torch.float32)
    grid_w = torch.arange(w, dtype=torch.float32)
    grid_d = torch.arange(d, dtype=torch.float32)

    out_h = torch.einsum("m,d->md", grid_h.flatten(), omega)
    out_w = torch.einsum("m,d->md", grid_w.flatten(), omega)
    out_d = torch.einsum("m,d->md", grid_d.flatten(), omega)

    # Expand to full grid: [H*W*D, pos_dim] for each axis
    # Use meshgrid ordering to get correct spatial layout
    grid_h_idx, grid_w_idx, grid_d_idx = torch.meshgrid(
        torch.arange(h), torch.arange(w), torch.arange(d), indexing="ij"
    )
    grid_h_flat = grid_h_idx.flatten()  # [H*W*D]
    grid_w_flat = grid_w_idx.flatten()
    grid_d_flat = grid_d_idx.flatten()

    pos_emb = torch.cat([
        torch.sin(out_h[grid_h_flat]),
        torch.cos(out_h[grid_h_flat]),
        torch.sin(out_w[grid_w_flat]),
        torch.cos(out_w[grid_w_flat]),
        torch.sin(out_d[grid_d_flat]),
        torch.cos(out_d[grid_d_flat]),
    ], dim=1)[None, :, :]  # [1, num_patches, embed_dim]

    pos_emb = nn.Parameter(pos_emb)
    pos_emb.requires_grad = False
    return pos_emb


def generate_dirichlet_mask_ratios(
    num_modalities: int,
    alpha: float,
    overall_mask_ratio: float,
) -> torch.Tensor:
    """
    Sample per-modality mask ratios from a Dirichlet distribution.
    The total visible budget is distributed among modalities.

    Returns: [num_modalities] tensor of per-modality mask ratios
    """
    dirichlet = Dirichlet(torch.tensor([float(alpha)] * num_modalities))
    visible_ratio = 1.0 - overall_mask_ratio
    total_visible = visible_ratio * num_modalities
    visible_per_mod = total_visible * dirichlet.sample()
    # Clamp to [0, 1]
    mask_ratios = (1.0 - visible_per_mod).clamp(0.0, 1.0)
    return mask_ratios


def compute_mask_ratios(
    modality_names: List[str],
    observed: torch.Tensor,
    mask_ratio: float = 0.75,
    use_dirichlet: bool = True,
    dirichlet_alpha: float = 1.0,
) -> Dict[str, float]:
    """
    Compute per-modality mask ratios, respecting observed mask.
    Missing modalities (observed=0) get mask_ratio=1.0.
    Present modalities get Dirichlet or uniform masking.

    Args:
        modality_names: list of modality names, e.g. ['T1', 'T2', 'Flair', 'PET']
        observed: [M] bool/float tensor, 1.0=present, 0.0=missing
            NOTE: This is per-sample, called once per sample in the batch.
            For simplicity, we use the same mask ratio for the whole batch
            (based on which modalities are present in the majority of the batch).
        mask_ratio: overall target mask ratio for present modalities
        use_dirichlet: whether to use Dirichlet distribution
        dirichlet_alpha: Dirichlet concentration parameter

    Returns:
        dict mapping modality_name -> mask_ratio (float)
    """
    ratios = {}
    present_mods = [name for i, name in enumerate(modality_names) if observed[i] > 0.5]
    missing_mods = [name for i, name in enumerate(modality_names) if observed[i] <= 0.5]

    # Missing modalities: fully masked
    for name in missing_mods:
        ratios[name] = 1.0

    # Present modalities: Dirichlet or uniform
    if len(present_mods) > 0:
        if use_dirichlet and len(present_mods) > 1:
            # Dirichlet masking among present modalities
            dir_ratios = generate_dirichlet_mask_ratios(
                num_modalities=len(present_mods),
                alpha=dirichlet_alpha,
                overall_mask_ratio=mask_ratio,
            )
            for i, name in enumerate(present_mods):
                ratios[name] = dir_ratios[i].item()
        else:
            # Uniform masking
            for name in present_mods:
                ratios[name] = mask_ratio

    return ratios


def mask_data(
    batch: Dict[str, torch.Tensor],
    modality_names: List[str],
    observed: torch.Tensor,
    mask_ratio: float = 0.75,
    patch_size: Union[int, Tuple[int, int, int]] = 16,
    use_dirichlet: bool = True,
    dirichlet_alpha: float = 1.0,
    patch_mask_probs: torch.Tensor = None,
) -> Tuple[
    Dict[str, torch.Tensor],
    Dict[str, torch.Tensor],
    Dict[str, torch.Tensor],
    Dict[str, float],
]:
    """
    Core masking function for MultiMAE pretraining.

    For each modality:
    - Patchify the image
    - Shuffle patches (optionally weighted by anatomy importance)
    - Split into selected (visible) and masked based on mask_ratio
    - Missing modalities (observed=0) get 100% masking

    Args:
        batch: dict mapping modality name -> [B, 1, D, H, W] tensor
        modality_names: ordered list of modality names
        observed: [B, M] tensor indicating which modalities are present
        mask_ratio: target mask ratio for present modalities
        patch_size: patch size for patchification
        use_dirichlet: whether to use Dirichlet distribution
        dirichlet_alpha: Dirichlet concentration parameter
        patch_mask_probs: optional [N_patches] per-patch masking probability
            from anatomy-aware masking. When provided, uses Gumbel-top-k
            weighted sampling instead of uniform random shuffling.
            Higher probability = more likely to be masked.

    Returns:
        selected_patches: {modality: [B, num_selected, C, pd, ph, pw]}
        masked_patches: {modality: [B, num_masked, C, pd, ph, pw]}
        perm_indices: {modality: [B, num_patches]}
        mask_ratios: {modality: float}
    """
    patch_size = to_3tuple(patch_size)
    batch_size = observed.shape[0]

    # Union strategy: if ANY sample in the batch has a modality, it gets
    # partial masking. Samples where this modality is missing contribute
    # zero-valued patches (harmless in encoder, excluded from loss).
    # This ensures no information is wasted when modalities are present
    # in a minority of samples.
    batch_observed = (observed.max(dim=0).values > 0.5).float()  # [M]
    mask_ratios = compute_mask_ratios(
        modality_names=modality_names,
        observed=batch_observed,
        mask_ratio=mask_ratio,
        use_dirichlet=use_dirichlet,
        dirichlet_alpha=dirichlet_alpha,
    )

    selected_patches = {}
    masked_patches = {}
    perm_indices = {}

    for mod_name in modality_names:
        # Patchify: [B, 1, D, H, W] -> [B, num_patches, 1, pd, ph, pw]
        patches = patchify(batch[mod_name], patch_size)
        num_patches = patches.shape[1]

        # Shuffle patches (weighted by anatomy importance if provided)
        shuffled, perm_idx = shuffle_patches(patches, mask_probs=patch_mask_probs)
        perm_indices[mod_name] = perm_idx

        # Split into selected and masked
        mod_mask_ratio = mask_ratios[mod_name]
        num_selected = int((1.0 - mod_mask_ratio) * num_patches)
        # Ensure at least 0 selected (for fully masked modalities)
        num_selected = max(0, num_selected)

        selected_patches[mod_name] = shuffled[:, :num_selected]
        masked_patches[mod_name] = shuffled[:, num_selected:]

    return selected_patches, masked_patches, perm_indices, mask_ratios