File size: 13,381 Bytes
a639402 | 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 | import torch
import torch.nn as nn
import numpy as np
import pandas as pd
from pathlib import Path
import timm
class MultiViewSwinModel(nn.Module):
def __init__(self, model_name="swin_tiny_patch4_window7_224.ms_in22k",
pretrained=False, num_classes=3, attn_dim=1024, num_heads=4):
super(MultiViewSwinModel, self).__init__()
# Load backbone without classification head
self.backbone = timm.create_model(model_name, pretrained=pretrained, num_classes=0)
self.feature_dim = self.backbone.num_features # should be 1024 for swin_base
# Self-attention across views (3 views → sequence length 3)
self.attn = nn.MultiheadAttention(embed_dim=self.feature_dim, num_heads=num_heads, batch_first=True)
# Final classifier
self.classifier = nn.Linear(self.feature_dim, num_classes)
def extract_feat(self, x):
# Get unpooled features (B, H*W, C) and then pooled vector (B, C)
features = self.backbone.forward_features(x) # (B, H*W, C)
pooled = self.backbone.forward_head(features, pre_logits=True) # (B, C)
return pooled
def forward(self, img1, img2, img3):
# Per-view embeddings
f1 = self.extract_feat(img1) # [B, C]
f2 = self.extract_feat(img2)
f3 = self.extract_feat(img3)
# Stack views to form a sequence [B, 3, C]
views = torch.stack([f1, f2, f3], dim=1)
# Self-attention across views
attn_out, _ = self.attn(views, views, views) # [B, 3, C]
# Aggregate (mean pooling across 3 views)
fused = attn_out.mean(dim=1) # [B, C]
return self.classifier(fused)
class MultiViewSwinCrossAttn(nn.Module):
"""
Cross‑attention Swin for 3‑view breast‑MRI (pre / post / subtraction).
"""
def __init__(
self,
model_name: str = "swin_tiny_patch4_window7_224.ms_in22k",
pretrained: bool = False,
num_classes: int = 3,
num_xattn_layers: int = 2,
num_heads: int = 4,
dropout: float = 0.1,
):
super().__init__()
# 1) Swin backbone WITHOUT final linear head
self.backbone = timm.create_model(model_name, pretrained=pretrained, num_classes=0)
self.embed_dim = self.backbone.num_features # 1024 for swin‑base
# 2) Learnable CLS token (like ViT) and view‑type embeddings
self.cls_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
self.view_embed = nn.Parameter(torch.zeros(3, 1, self.embed_dim)) # pre/post/sub
# 3) Cross‑attention encoder (TransformerEncoder)
encoder_layer = nn.TransformerEncoderLayer(
d_model=self.embed_dim,
nhead=num_heads,
dim_feedforward=self.embed_dim * 4,
dropout=dropout,
activation="gelu",
batch_first=True,
norm_first=True,
)
self.xattn = nn.TransformerEncoder(encoder_layer, num_layers=num_xattn_layers)
# 4) Classification head
self.norm = nn.LayerNorm(self.embed_dim)
self.classifier = nn.Linear(self.embed_dim, num_classes)
# init
nn.init.trunc_normal_(self.cls_token, std=0.02)
nn.init.trunc_normal_(self.view_embed, std=0.02)
# ------------------------------------------------------------------ #
def _img_embed(self, x: torch.Tensor) -> torch.Tensor:
"""
Forward a single image through Swin and return pooled vector (B, C).
"""
feats = self.backbone.forward_features(x) # (B, H*W, C)
pooled = self.backbone.forward_head(feats, pre_logits=True) # (B, C)
return pooled
def forward(self, pre, post, sub):
"""
Args:
pre, post, sub: tensors (B, 3, 224, 224)
Returns:
logits (B, num_classes)
"""
# 1) Per‑view embeddings
z_pre = self._img_embed(pre) # (B, C)
z_post = self._img_embed(post)
z_sub = self._img_embed(sub)
# 2) Assemble sequence: [CLS] + three view tokens
B = z_pre.size(0)
cls = self.cls_token.expand(B, -1, -1) # (B, 1, C)
views = torch.stack([z_pre, z_post, z_sub], dim=1) # (B, 3, C)
# add view‑type embeddings (broadcast over batch)
views = views + self.view_embed.transpose(0, 1) # (B, 3, C)
tokens = torch.cat([cls, views], dim=1) # (B, 4, C)
# 3) Cross‑attention
tokens = self.xattn(tokens) # (B, 4, C)
# 4) CLS pooling → head
out = self.classifier(self.norm(tokens[:, 0])) # (B, num_classes)
return out
class CrossModalAttentionABMIL_Swin(nn.Module):
"""
Swin‑based Multiple‑Instance model with per‑slice cross‑modal attention
over (pre, post, sub) features, followed by ABMIL pooling.
Input : (B, 32, 3, 224, 224) # 3 modalities stacked in channel dim
Output : logits (B, num_classes), attention‑per‑slice (B, 32)
"""
def __init__(
self,
model_name: str = "swin_tiny_patch4_window7_224.ms_in22k",
pretrained: bool = False,
num_classes: int = 3,
hidden_dim: int = 256,
cross_attn_heads: int = 4,
dropout: float = 0.1,
):
super().__init__()
# ------------------------------------------------------------
# 1) Shared Swin backbone (no classifier head)
# ------------------------------------------------------------
self.backbone = timm.create_model(model_name, pretrained=pretrained, num_classes=0)
self.embed_dim = self.backbone.num_features # 768 for swin_tiny
# ------------------------------------------------------------
# 2) Per‑slice cross‑modal fusion
# Input: three embeddings (pre, post, sub) -> fused embedding
# ------------------------------------------------------------
encoder_layer = nn.TransformerEncoderLayer(
d_model=self.embed_dim,
nhead=cross_attn_heads,
dim_feedforward=self.embed_dim * 4,
dropout=dropout,
activation="gelu",
batch_first=True,
norm_first=True,
)
# One small transformer (2 layers) that attends across 3 tokens
self.slice_fuser = nn.TransformerEncoder(encoder_layer, num_layers=2)
# Token type embedding to differentiate modalities
self.mod_embed = nn.Parameter(torch.randn(3, 1, self.embed_dim))
# ------------------------------------------------------------
# 3) ABMIL Attention across 32 fused slices
# ------------------------------------------------------------
self.attn_V = nn.Linear(self.embed_dim, hidden_dim)
self.attn_U = nn.Linear(hidden_dim, 1)
# ------------------------------------------------------------
# 4) Classifier head
# ------------------------------------------------------------
self.norm = nn.LayerNorm(self.embed_dim)
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(self.embed_dim, num_classes)
# Init
nn.init.trunc_normal_(self.mod_embed, std=0.02)
nn.init.trunc_normal_(self.attn_V.weight, std=0.02)
nn.init.trunc_normal_(self.attn_U.weight, std=0.02)
# ------------------------------------------------------------
def _encode_slices(self, x: torch.Tensor) -> torch.Tensor:
"""
x : (B*N, 1, 224, 224) -> Swin pooled feats (B*N, embed_dim)
"""
feats = self.backbone.forward_features(x)
return self.backbone.forward_head(feats, pre_logits=True)
# ------------------------------------------------------------
def forward(self, volume: torch.Tensor):
"""
volume : (B, 32, 3, 224, 224)
"""
B, N, C_img, H, W = volume.shape # C_img = 3 modalities
assert C_img == 3, "Expect channel dim = 3 (pre, post, sub)"
# --------------------------------------------------------
# 1) Split modalities, flatten, and encode with Swin
# --------------------------------------------------------
# Each is (B*N, 1, H, W)
def expand3(x): return x.expand(-1, 3, -1, -1)
pre = expand3(volume[:, :, 0, :, :].contiguous().view(B * N, 1, H, W))
post = expand3(volume[:, :, 1, :, :].contiguous().view(B * N, 1, H, W))
sub = expand3(volume[:, :, 2, :, :].contiguous().view(B * N, 1, H, W))
feat_pre = self._encode_slices(pre).view(B, N, -1) # (B, N, C)
feat_post = self._encode_slices(post).view(B, N, -1)
feat_sub = self._encode_slices(sub).view(B, N, -1)
# --------------------------------------------------------
# 2) Cross‑modal attention fusion (per slice)
# --------------------------------------------------------
# Build token sequence [pre, post, sub] for each slice
# Shape before fuser: (B, N, 3, C) -> we fuse along dim=2
slice_tokens = torch.stack([feat_pre, feat_post, feat_sub], dim=2)
# Add modality embeddings
slice_tokens = slice_tokens + self.mod_embed.transpose(0, 1) # (B, N, 3, C)
slice_tokens = slice_tokens.view(B * N, 3, self.embed_dim) # (B*N, 3, C)
# Transformer encoder attends across the 3 tokens
fused = self.slice_fuser(slice_tokens)[:, 0] # take CLS‑like first token
fused = fused.view(B, N, self.embed_dim) # (B, 32, C)
# --------------------------------------------------------
# 3) ABMIL attention over 32 fused slices
# --------------------------------------------------------
A = torch.tanh(self.attn_V(fused)) # (B, N, hidden)
A = self.attn_U(A) # (B, N, 1)
A = torch.softmax(A, dim=1) # (B, N, 1)
patient_feat = (A * fused).sum(dim=1) # (B, C)
# --------------------------------------------------------
# 4) Head
# --------------------------------------------------------
logits = self.classifier(self.dropout(self.norm(patient_feat)))
return logits, A.squeeze(-1) # (B, num_classes), (B, 32)
class ABMIL_Swin(nn.Module):
"""
Attention‑based Multiple‑Instance Learning (ABMIL) model
for 3‑channel breast MRI slice triplets (pre / post / sub).
Input shape : (B, 32, 3, 224, 224)
Output : logits (B, num_classes) + attention weights (B, 32)
"""
def __init__(
self,
model_name: str = "swin_tiny_patch4_window7_224.ms_in22k",
pretrained: bool = False,
num_classes: int = 2,
hidden_dim: int = 256,
dropout: float = 0.1,
):
super().__init__()
# 1) Swin backbone WITHOUT final linear head
self.backbone = timm.create_model(model_name, pretrained=pretrained, num_classes=0)
self.embed_dim = self.backbone.num_features # e.g. 768 (tiny), 1024 (base)
# 2) Attention network (ABMIL)
self.attn_V = nn.Linear(self.embed_dim, hidden_dim)
self.attn_U = nn.Linear(hidden_dim, 1)
# 3) Patient‑level classifier
self.norm = nn.LayerNorm(self.embed_dim)
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(self.embed_dim, num_classes)
# Optional init
nn.init.trunc_normal_(self.attn_V.weight, std=0.02)
nn.init.trunc_normal_(self.attn_U.weight, std=0.02)
# ------------------------------------------------------------------ #
def _slice_embed(self, x: torch.Tensor) -> torch.Tensor:
"""
Forward one or more images through Swin and return pooled features.
Args:
x : (N, 3, 224, 224)
Returns:
pooled : (N, C)
"""
feats = self.backbone.forward_features(x) # (N, H*W, C)
pooled = self.backbone.forward_head(feats, pre_logits=True) # (N, C)
return pooled
# ------------------------------------------------------------------ #
def forward(self, volume: torch.Tensor):
"""
Args:
volume : (B, 32, 3, 224, 224)
Returns:
logits : (B, num_classes)
attn_w : (B, 32) -- attention per slice (sums to 1)
"""
B, N, C, H, W = volume.shape
x = volume.view(B * N, C, H, W) # flatten slices
# (B*N, 3, 224, 224) → (B*N, embed_dim) → reshape
slice_feats = self._slice_embed(x).view(B, N, -1) # (B, 32, embed_dim)
# ---------------- ABMIL attention ---------------- #
A = torch.tanh(self.attn_V(slice_feats)) # (B, 32, hidden)
A = self.attn_U(A) # (B, 32, 1)
A = torch.softmax(A, dim=1) # attention weights
# Weighted sum → patient embedding
patient_feat = (A * slice_feats).sum(dim=1) # (B, embed_dim)
# ---------------- Head ---------------- #
out = self.classifier(self.dropout(self.norm(patient_feat))) # (B, num_classes)
return out, A.squeeze(-1) # logits, attention weights |