Spaces:
Running
Running
File size: 14,018 Bytes
9bfed75 433e26f 9bfed75 ded6c17 9bfed75 ded6c17 9bfed75 433e26f 9bfed75 433e26f 9bfed75 433e26f 9bfed75 433e26f 9bfed75 433e26f 9bfed75 433e26f 9bfed75 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 | """Clinical safety validation for responsible deployment.
Implements safety checks for surgical outcome predictions:
1. Identity preservation: verify output preserves patient identity
2. Anatomical plausibility: check landmark displacements are realistic
3. Out-of-distribution detection: flag unusual inputs
4. Watermarking: mark AI-generated images
5. Consent metadata: embed provenance information
Usage:
from landmarkdiff.safety import SafetyValidator
validator = SafetyValidator()
result = validator.validate(
input_image=image,
output_image=generated,
landmarks_original=face.landmarks,
landmarks_manipulated=manip.landmarks,
procedure="rhinoplasty",
)
if not result.passed:
print(f"Safety check failed: {result.failures}")
"""
from __future__ import annotations
from dataclasses import dataclass, field
import cv2
import numpy as np
@dataclass
class SafetyResult:
"""Result of safety validation checks."""
passed: bool = True
failures: list[str] = field(default_factory=list)
warnings: list[str] = field(default_factory=list)
checks: dict[str, bool] = field(default_factory=dict)
details: dict[str, object] = field(default_factory=dict)
def __repr__(self) -> str:
return (
f"SafetyResult(passed={self.passed}, "
f"failures={self.failures}, "
f"warnings={self.warnings}, "
f"checks={self.checks}, "
f"details={self.details})"
)
def __eq__(self, other: object) -> bool:
if not isinstance(other, SafetyResult):
return NotImplemented
return (
self.passed == other.passed
and self.failures == other.failures
and self.warnings == other.warnings
and self.checks == other.checks
and self.details == other.details
)
def add_failure(self, name: str, message: str) -> None:
self.passed = False
self.failures.append(message)
self.checks[name] = False
def add_warning(self, name: str, message: str) -> None:
self.warnings.append(message)
def add_pass(self, name: str) -> None:
self.checks[name] = True
def summary(self) -> str:
lines = [f"Safety: {'PASS' if self.passed else 'FAIL'}"]
for name, ok in self.checks.items():
lines.append(f" [{'OK' if ok else 'FAIL'}] {name}")
for w in self.warnings:
lines.append(f" [WARN] {w}")
return "\n".join(lines)
class SafetyValidator:
"""Clinical safety validation for surgical predictions."""
def __init__(
self,
identity_threshold: float = 0.5,
max_displacement_fraction: float = 0.05,
min_face_confidence: float = 0.5,
max_yaw_degrees: float = 45.0,
watermark_enabled: bool = True,
watermark_text: str = "AI-GENERATED PREDICTION",
):
self.identity_threshold = identity_threshold
self.max_displacement_fraction = max_displacement_fraction
self.min_face_confidence = min_face_confidence
self.max_yaw_degrees = max_yaw_degrees
self.watermark_enabled = watermark_enabled
self.watermark_text = watermark_text
def validate(
self,
input_image: np.ndarray,
output_image: np.ndarray,
landmarks_original: np.ndarray | None = None,
landmarks_manipulated: np.ndarray | None = None,
procedure: str | None = None,
face_confidence: float = 1.0,
) -> SafetyResult:
"""Run all safety checks on a prediction.
Args:
input_image: Original patient image (BGR, uint8).
output_image: Generated prediction (BGR, uint8).
landmarks_original: Original landmarks (N, 2-3), normalized [0, 1].
landmarks_manipulated: Manipulated landmarks (N, 2-3), normalized [0, 1].
procedure: Surgical procedure name.
face_confidence: MediaPipe face detection confidence.
Returns:
SafetyResult with all check results.
"""
result = SafetyResult()
# 1. Face detection confidence
self._check_face_confidence(result, face_confidence)
# 2. Identity preservation
self._check_identity(result, input_image, output_image)
# 3. Anatomical plausibility
if landmarks_original is not None and landmarks_manipulated is not None:
self._check_anatomical_plausibility(
result, landmarks_original, landmarks_manipulated, procedure
)
# 4. Output quality
self._check_output_quality(result, output_image)
# 5. OOD detection (basic)
self._check_ood(result, input_image)
return result
def _check_face_confidence(self, result: SafetyResult, confidence: float) -> None:
"""Check face detection confidence."""
if confidence < self.min_face_confidence:
result.add_failure(
"face_confidence",
f"Face detection confidence {confidence:.2f} below threshold "
f"{self.min_face_confidence}",
)
else:
result.add_pass("face_confidence")
result.details["face_confidence"] = confidence
def _check_identity(
self,
result: SafetyResult,
input_image: np.ndarray,
output_image: np.ndarray,
) -> None:
"""Check identity preservation using ArcFace similarity."""
try:
from landmarkdiff.evaluation import compute_identity_similarity
sim = compute_identity_similarity(output_image, input_image)
result.details["identity_similarity"] = float(sim)
if sim < self.identity_threshold:
result.add_failure(
"identity",
f"Identity similarity {sim:.3f} below threshold {self.identity_threshold}",
)
else:
result.add_pass("identity")
except Exception as e:
result.add_warning("identity", f"Identity check failed: {e}")
def _check_anatomical_plausibility(
self,
result: SafetyResult,
landmarks_orig: np.ndarray,
landmarks_manip: np.ndarray,
procedure: str | None,
) -> None:
"""Check that landmark displacements are anatomically plausible."""
if len(landmarks_orig) != len(landmarks_manip):
result.add_failure(
"anatomical",
f"Landmark count mismatch: {len(landmarks_orig)} vs {len(landmarks_manip)}",
)
return
# Compute displacement magnitudes
n = min(len(landmarks_orig), len(landmarks_manip))
orig = landmarks_orig[:n, :2] # (N, 2), normalized [0, 1]
manip = landmarks_manip[:n, :2]
displacements = np.linalg.norm(manip - orig, axis=1)
max_disp = float(displacements.max())
mean_disp = float(displacements.mean())
result.details["max_displacement"] = max_disp
result.details["mean_displacement"] = mean_disp
# Check maximum displacement
if max_disp > self.max_displacement_fraction:
result.add_failure(
"anatomical_magnitude",
f"Maximum displacement {max_disp:.4f} exceeds threshold "
f"{self.max_displacement_fraction}",
)
else:
result.add_pass("anatomical_magnitude")
# Check procedure-specific regions
if procedure:
self._check_procedure_regions(result, orig, manip, displacements, procedure)
def _check_procedure_regions(
self,
result: SafetyResult,
orig: np.ndarray,
manip: np.ndarray,
displacements: np.ndarray,
procedure: str,
) -> None:
"""Verify displacement is concentrated in expected anatomical regions."""
from landmarkdiff.landmarks import LANDMARK_REGIONS
# Expected regions by procedure
expected_regions = {
"rhinoplasty": ["nose"],
"blepharoplasty": ["eye_left", "eye_right"],
"rhytidectomy": ["jawline"],
"orthognathic": ["jawline", "lips"],
}
expected = expected_regions.get(procedure, [])
if not expected:
result.add_pass("procedure_region")
return
# Get expected region indices
expected_indices = set()
for region in expected:
if region in LANDMARK_REGIONS:
expected_indices.update(LANDMARK_REGIONS[region])
if not expected_indices:
result.add_pass("procedure_region")
return
# Check: is most displacement in expected regions?
n = min(len(displacements), len(orig))
expected_mask = np.array([i in expected_indices for i in range(n)])
if expected_mask.sum() > 0 and (~expected_mask).sum() > 0:
expected_disp = displacements[expected_mask].mean()
unexpected_disp = displacements[~expected_mask].mean()
result.details["expected_region_disp"] = float(expected_disp)
result.details["unexpected_region_disp"] = float(unexpected_disp)
# Expected regions should have more displacement
if unexpected_disp > expected_disp * 2 and unexpected_disp > 0.005:
result.add_warning(
"procedure_region",
f"{procedure}: unexpected regions displaced more than expected "
f"({unexpected_disp:.4f} vs {expected_disp:.4f})",
)
else:
result.add_pass("procedure_region")
else:
result.add_pass("procedure_region")
def _check_output_quality(self, result: SafetyResult, output: np.ndarray) -> None:
"""Check output image quality (not blank, not corrupted)."""
if output is None or output.size == 0:
result.add_failure("output_quality", "Output image is empty")
return
# Check for blank/black images
mean_val = output.mean()
if mean_val < 5:
result.add_failure("output_quality", f"Output is nearly black (mean={mean_val:.1f})")
return
if mean_val > 250:
result.add_failure("output_quality", f"Output is nearly white (mean={mean_val:.1f})")
return
# Check for artifacts (extreme variance)
std_val = output.std()
if std_val < 10:
result.add_warning(
"output_quality",
f"Output has very low variance (std={std_val:.1f}), may be uniform",
)
result.add_pass("output_quality")
result.details["output_mean"] = float(mean_val)
result.details["output_std"] = float(std_val)
def _check_ood(self, result: SafetyResult, image: np.ndarray) -> None:
"""Basic out-of-distribution detection.
Checks image properties against expected ranges for face photos.
"""
h, w = image.shape[:2]
# Resolution check
if min(h, w) < 128:
result.add_warning("ood", f"Image resolution too low: {w}x{h}")
# Aspect ratio (faces should be roughly square after preprocessing)
aspect = max(h, w) / max(min(h, w), 1)
if aspect > 3.0:
result.add_warning("ood", f"Unusual aspect ratio: {aspect:.1f}")
# Color distribution (face photos should have some skin tones)
if len(image.shape) == 3 and image.shape[2] == 3:
mean_b, mean_g, mean_r = image.mean(axis=(0, 1))
# Face images typically have red channel > blue channel
if mean_b > mean_r * 1.5:
result.add_warning("ood", "Image appears very blue (not typical face photo)")
result.add_pass("ood_basic")
def apply_watermark(
self,
image: np.ndarray,
text: str | None = None,
opacity: float = 0.3,
) -> np.ndarray:
"""Apply a text watermark to the output image.
Places semi-transparent text at the bottom of the image to indicate
it is AI-generated.
"""
if not self.watermark_enabled:
return image
text = text or self.watermark_text
result = image.copy()
h, w = result.shape[:2]
# Create text overlay
font = cv2.FONT_HERSHEY_SIMPLEX
font_scale = max(0.3, w / 1500)
thickness = max(1, int(w / 500))
text_size = cv2.getTextSize(text, font, font_scale, thickness)[0]
x = (w - text_size[0]) // 2
y = h - 10
# Semi-transparent background bar
bar_y1 = y - text_size[1] - 10
bar_y2 = h
overlay = result.copy()
cv2.rectangle(overlay, (0, bar_y1), (w, bar_y2), (0, 0, 0), -1)
cv2.addWeighted(overlay, opacity, result, 1 - opacity, 0, result)
# White text
cv2.putText(result, text, (x, y), font, font_scale, (255, 255, 255), thickness, cv2.LINE_AA)
return result
def embed_metadata(
self,
image_path: str,
procedure: str,
intensity: float,
model_version: str = "0.3.0",
) -> None:
"""Embed provenance metadata in the output image.
Writes EXIF/PNG metadata with generation parameters for traceability.
"""
import json
from pathlib import Path
meta = {
"generator": "LandmarkDiff",
"version": model_version,
"procedure": procedure,
"intensity": intensity,
"disclaimer": "AI-generated surgical prediction for visualization only. "
"Not a guarantee of surgical outcome.",
}
# Save as sidecar JSON (PNG doesn't have easy EXIF support)
meta_path = Path(image_path).with_suffix(".meta.json")
with open(meta_path, "w") as f:
json.dump(meta, f, indent=2)
|