Spaces:
Sleeping
Sleeping
File size: 10,850 Bytes
8960670 | 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 | #!/usr/bin/env python3
"""
3D GradCAM for OncoVision-X Lung Nodule Classification.
Generates gradient-weighted class activation maps to visualize
which regions of the input the model focuses on for its predictions.
Usage:
from src.explainability.gradcam import GradCAM3D
gradcam = GradCAM3D(model, target_layer='nodule_stream')
heatmap = gradcam(nodule_patch, context_patch)
"""
import numpy as np
import torch
import torch.nn.functional as F
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from pathlib import Path
class GradCAM3D:
"""3D Gradient-weighted Class Activation Mapping.
Generates heatmaps showing which spatial regions of the input
volume contribute most to the model's prediction.
Supports two target streams:
- 'nodule_stream': Visualize nodule patch focus (64Β³)
- 'context_stream': Visualize context patch focus (48Β³)
"""
def __init__(self, model, target_stream='nodule_stream'):
"""
Args:
model: DCANet model (unwrapped from DataParallel)
target_stream: 'nodule_stream' or 'context_stream'
"""
self.model = model
self.model.eval()
self.target_stream = target_stream
# Storage for hooks
self.activations = None
self.gradients = None
# Register hooks on the target stream's last conv layer
if target_stream == 'nodule_stream':
# Hook into the backbone's final feature extraction
target_layer = self._get_nodule_target()
elif target_stream == 'context_stream':
target_layer = self._get_context_target()
else:
raise ValueError(f"Unknown target_stream: {target_stream}")
target_layer.register_forward_hook(self._save_activation)
target_layer.register_full_backward_hook(self._save_gradient)
def _get_nodule_target(self):
"""Get the last convolutional layer of the nodule stream backbone."""
backbone = self.model.nodule_stream.backbone
# For EfficientNet, the last conv features are in the final block
# Use the conv_head or the last conv layer
if hasattr(backbone, 'conv_head'):
return backbone.conv_head
elif hasattr(backbone, 'blocks'):
# Last block's last conv
return backbone.blocks[-1]
else:
# Fallback: last named module with Conv2d
last_conv = None
for module in backbone.modules():
if isinstance(module, torch.nn.Conv2d):
last_conv = module
if last_conv is None:
raise RuntimeError("Could not find target layer in nodule backbone")
return last_conv
def _get_context_target(self):
"""Get the last conv block of the context stream."""
return self.model.context_stream.block3
def _save_activation(self, module, input, output):
"""Forward hook to save activations."""
self.activations = output.detach()
def _save_gradient(self, module, grad_input, grad_output):
"""Backward hook to save gradients."""
self.gradients = grad_output[0].detach()
def generate(self, nodule_patch, context_patch, device='cpu'):
"""Generate GradCAM heatmap.
Args:
nodule_patch: numpy array (64, 64, 64) or (1, 1, 64, 64, 64)
context_patch: numpy array (48, 48, 48) or (1, 1, 48, 48, 48)
device: torch device
Returns:
heatmap: numpy array, same spatial dims as target stream input
probability: model prediction probability
"""
# Ensure tensor format
if isinstance(nodule_patch, np.ndarray):
if nodule_patch.ndim == 3:
nodule_patch = nodule_patch[np.newaxis, np.newaxis, ...]
nodule_patch = torch.from_numpy(nodule_patch.astype(np.float32))
if isinstance(context_patch, np.ndarray):
if context_patch.ndim == 3:
context_patch = context_patch[np.newaxis, np.newaxis, ...]
context_patch = torch.from_numpy(context_patch.astype(np.float32))
nodule_patch = nodule_patch.to(device).requires_grad_(True)
context_patch = context_patch.to(device).requires_grad_(True)
# Forward pass
self.model.zero_grad()
logits = self.model(nodule_patch, context_patch)
prob = torch.sigmoid(logits.squeeze()).item()
# Backward pass β gradient w.r.t. the class score
logits.squeeze().backward()
# Compute GradCAM
if self.gradients is None or self.activations is None:
raise RuntimeError("Hooks did not capture gradients/activations")
gradients = self.gradients
activations = self.activations
# For nodule stream: activations are 2D per-slice (B*D, C, H, W)
# For context stream: activations are 3D (B, C, D, H, W)
if self.target_stream == 'nodule_stream':
# Global average pooling of gradients over spatial dims
weights = gradients.mean(dim=(-2, -1), keepdim=True) # (N, C, 1, 1)
cam = (weights * activations).sum(dim=1) # (N, H, W)
cam = F.relu(cam)
# Reshape to 3D: N slices β (D, H, W)
D = nodule_patch.shape[2] # depth
num_per_slice = cam.shape[0] // nodule_patch.shape[0]
cam = cam[:num_per_slice] # Take first batch item
# Resize each slice to original spatial size
target_h, target_w = nodule_patch.shape[3], nodule_patch.shape[4]
cam_resized = F.interpolate(
cam.unsqueeze(1), # (D, 1, h, w)
size=(target_h, target_w),
mode='bilinear', align_corners=False
).squeeze(1) # (D, H, W)
heatmap = cam_resized.cpu().numpy()
else:
# Context stream: 3D activations
weights = gradients.mean(dim=(-3, -2, -1), keepdim=True)
cam = (weights * activations).sum(dim=1, keepdim=True)
cam = F.relu(cam)
target_size = context_patch.shape[2:]
cam = F.interpolate(cam, size=target_size, mode='trilinear', align_corners=False)
heatmap = cam.squeeze().cpu().numpy()
# Normalize to [0, 1]
if heatmap.max() > 0:
heatmap = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-8)
return heatmap, prob
def __call__(self, nodule_patch, context_patch, device='cpu'):
return self.generate(nodule_patch, context_patch, device)
def plot_gradcam_slices(scan_patch, heatmap, probability, output_path,
num_slices=8, title="GradCAM Visualization"):
"""Plot GradCAM overlay on selected slices.
Args:
scan_patch: 3D numpy array (D, H, W) β original input
heatmap: 3D numpy array (D, H, W) β GradCAM heatmap
probability: float β model prediction
output_path: str β save path
num_slices: int β number of slices to display
title: str β plot title
"""
D = scan_patch.shape[0]
slice_indices = np.linspace(D // 8, D - D // 8, num_slices, dtype=int)
fig, axes = plt.subplots(2, num_slices, figsize=(3 * num_slices, 7))
label = "MALIGNANT" if probability > 0.5 else "BENIGN"
color = 'red' if probability > 0.5 else 'green'
fig.suptitle(
f"{title}\nPrediction: {label} ({probability:.1%})",
fontsize=14, fontweight='bold', color=color
)
for i, idx in enumerate(slice_indices):
# Original slice
axes[0, i].imshow(scan_patch[idx], cmap='gray', vmin=-1, vmax=1)
axes[0, i].set_title(f"Slice {idx}", fontsize=9)
axes[0, i].axis('off')
# Overlay
axes[1, i].imshow(scan_patch[idx], cmap='gray', vmin=-1, vmax=1)
overlay = axes[1, i].imshow(
heatmap[idx], cmap='jet', alpha=0.5, vmin=0, vmax=1
)
axes[1, i].set_title(f"GradCAM", fontsize=9)
axes[1, i].axis('off')
# Colorbar
cbar_ax = fig.add_axes([0.92, 0.15, 0.02, 0.35])
fig.colorbar(overlay, cax=cbar_ax, label='Attention')
plt.tight_layout(rect=[0, 0, 0.9, 0.92])
Path(output_path).parent.mkdir(parents=True, exist_ok=True)
plt.savefig(output_path, dpi=150, bbox_inches='tight')
plt.close()
return output_path
def generate_gradcam_report(model, dataloader, device, output_dir,
num_samples=10, stream='nodule_stream'):
"""Generate GradCAM visualizations for multiple samples.
Args:
model: Trained DCANet model
dataloader: Test DataLoader
device: torch device
output_dir: Directory to save GradCAM plots
num_samples: Number of samples to visualize
stream: Which stream to visualize
Returns:
List of output paths
"""
import torch.nn as nn
# Unwrap DataParallel if needed
if isinstance(model, nn.DataParallel):
raw_model = model.module
else:
raw_model = model
gradcam = GradCAM3D(raw_model, target_stream=stream)
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
outputs = []
sample_count = 0
for nodule, context, labels in dataloader:
for i in range(nodule.shape[0]):
if sample_count >= num_samples:
return outputs
nod = nodule[i:i+1]
ctx = context[i:i+1]
label = labels[i].item()
try:
heatmap, prob = gradcam.generate(nod, ctx, device)
except Exception as e:
print(f" GradCAM failed for sample {sample_count}: {e}")
continue
# Get the original scan patch for overlay
scan_slice = nod.squeeze().numpy()
# Crop heatmap to match scan if needed
if heatmap.shape != scan_slice.shape:
min_d = min(heatmap.shape[0], scan_slice.shape[0])
heatmap = heatmap[:min_d]
scan_slice = scan_slice[:min_d]
gt_str = "pos" if label == 1 else "neg"
pred_str = "malignant" if prob > 0.5 else "benign"
correct = (label == 1 and prob > 0.5) or (label == 0 and prob <= 0.5)
out_path = output_dir / f"gradcam_{sample_count:03d}_{gt_str}_pred_{pred_str}.png"
plot_gradcam_slices(
scan_slice, heatmap, prob, str(out_path),
title=f"Sample {sample_count} β GT: {'Cancer' if label==1 else 'Benign'} | "
f"{'Correct' if correct else 'WRONG'}"
)
outputs.append(str(out_path))
sample_count += 1
return outputs
|