本文档详细介绍了基于热力图的ICAS(颅内动脉粥样硬化性狭窄)检测与分类的完整算法流程。整个流程包括数据集构建、人脸检测、图像分割、特征提取和分类预测等多个阶段,采用多种机器学习和深度学习方法实现高精度的ICAS风险评估。
原始热力图数据
↓
数据集构建与标注
↓
YOLOv11人脸检测模型训练
↓
人脸区域检测与分割
↓
人脸热力数据集构建
↓
多种分类方法并行训练
├── 传统图像特征分类
├── 多模态特征融合分类
├── 深度学习分类
└── 对比学习分类
↓
模型性能评估与选择
↓
最终ICAS风险预测
- 数据来源: 2024年热成像数据
- 数据格式: JPG格式热力图图像
- 数据规模: 950张热力图
- 存储结构:
2025年合同完整数据/2024热成像数据/{patient_id}/{patient_id}1.jpg
使用 dataset/import_thermal_24h.py 脚本将热力图数据导入数据库:
# 数据导入流程
1. 扫描热力图文件目录
2. 提取患者ID信息
3. 记录文件路径和元数据
4. 存储到数据库files表使用 dataset/build_datasets.py 根据ICAS诊断结果构建分类数据集:
datasets/
├── thermal_24h/
│ ├── icas/ # ICAS阳性患者 (303张, 31.9%)
│ └── non_icas/ # ICAS阴性患者 (647张, 68.1%)
└── dataset_summary.json
- 标注工具: 专业图像标注工具
- 标注内容: 热力图中的人脸区域边界框
- 标注格式: YOLO格式 (class_id, x_center, y_center, width, height)
- 质量控制: 多人标注,交叉验证
# 训练配置
model_architecture: YOLOv11
input_size: 640x640
batch_size: 16
epochs: 100
optimizer: AdamW
learning_rate: 0.001
data_augmentation:
- rotation: ±15°
- scaling: 0.8-1.2
- brightness: ±20%
- contrast: ±20%# 检测流程伪代码
def detect_face_regions(thermal_image):
# 1. 图像预处理
image = preprocess_thermal_image(thermal_image)
# 2. YOLOv11推理
detections = yolo_model.predict(image)
# 3. 后处理
face_boxes = post_process_detections(detections)
# 4. 置信度筛选
valid_faces = filter_by_confidence(face_boxes, threshold=0.5)
return valid_faces# 分割流程
def segment_face_regions(image, face_boxes):
face_regions = []
for box in face_boxes:
# 1. 根据检测框裁剪人脸区域
face_region = crop_image(image, box)
# 2. 尺寸标准化
face_region = resize_image(face_region, target_size=(224, 224))
# 3. 质量检查
if quality_check(face_region):
face_regions.append(face_region)
return face_regions# 预处理流程
def preprocess_face_thermal_data():
processed_data = []
for image_path in thermal_images:
# 1. 加载原始热力图
thermal_img = load_thermal_image(image_path)
# 2. 人脸检测
face_boxes = detect_face_regions(thermal_img)
# 3. 人脸分割
face_regions = segment_face_regions(thermal_img, face_boxes)
# 4. 数据增强
augmented_faces = apply_augmentation(face_regions)
# 5. 质量筛选
valid_faces = quality_filter(augmented_faces)
processed_data.extend(valid_faces)
return processed_data# 数据集划分策略
train_ratio = 0.7 # 训练集70%
val_ratio = 0.15 # 验证集15%
test_ratio = 0.15 # 测试集15%
# 分层采样确保类别平衡
stratified_split(face_thermal_dataset, ratios=[0.7, 0.15, 0.15])基于 train_thermal_classifier1.py 实现:
# 特征提取流程
def extract_traditional_features(face_image):
features = []
# 1. 温度特征 (基于HSV色彩空间)
hsv = cv2.cvtColor(face_image, cv2.COLOR_BGR2HSV)
temp_features = extract_temperature_features(hsv)
# 2. 纹理特征 (LBP)
lbp_features = extract_lbp_features(face_image)
# 3. 形状特征 (轮廓分析)
shape_features = extract_shape_features(face_image)
features.extend([temp_features, lbp_features, shape_features])
return np.concatenate(features)# 多种分类器对比
classifiers = {
'LogisticRegression': LogisticRegression(),
'SVM': SVC(kernel='rbf'),
'RandomForest': RandomForestClassifier(),
'GradientBoosting': GradientBoostingClassifier(),
'XGBoost': XGBClassifier()
}基于 train_thermal_classifier2.py 实现:
# 多模态特征融合
def extract_multimodal_features(face_image, clinical_data):
# 1. 图像特征 (324维)
image_features = extract_enhanced_image_features(face_image)
# 2. 临床特征 (13维)
clinical_features = extract_clinical_features(clinical_data)
# 3. 特征融合
multimodal_features = np.concatenate([image_features, clinical_features])
return multimodal_features
def extract_enhanced_image_features(image):
# 温度特征 (~18维)
temp_features = extract_temperature_features(image)
# 分块LBP直方图 (288维)
block_lbp = extract_block_lbp_histogram(image, grid_size=(4,4))
# 全局LBP统计 (12维)
global_lbp = extract_global_lbp_statistics(image)
# 形状特征 (~6维)
shape_features = extract_shape_features(image)
return np.concatenate([temp_features, block_lbp, global_lbp, shape_features])def extract_clinical_features(patient_data):
# 基础生理指标 (8维)
basic_features = [
patient_data['age'], patient_data['gender_encoded'],
patient_data['height'], patient_data['weight'],
patient_data['bmi'], patient_data['waist'],
patient_data['hip'], patient_data['neck']
]
# 衍生特征 (3维)
derived_features = [
patient_data['waist'] / patient_data['hip'], # 腰臀比
patient_data['waist'] / patient_data['height'], # 腰身比
patient_data['neck'] / patient_data['height'] # 颈身比
]
# 分类特征 (2维)
categorical_features = [
categorize_bmi(patient_data['bmi']), # BMI分类
categorize_age(patient_data['age']) # 年龄分组
]
return np.concatenate([basic_features, derived_features, categorical_features])# 深度学习模型架构
class ThermalCNN(nn.Module):
def __init__(self, num_classes=2):
super(ThermalCNN, self).__init__()
# 特征提取层
self.features = nn.Sequential(
# Block 1
nn.Conv2d(3, 64, 3, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2),
# Block 2
nn.Conv2d(64, 128, 3, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2),
# Block 3
nn.Conv2d(128, 256, 3, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2),
)
# 分类层
self.classifier = nn.Sequential(
nn.AdaptiveAvgPool2d((7, 7)),
nn.Flatten(),
nn.Linear(256 * 7 * 7, 512),
nn.ReLU(inplace=True),
nn.Dropout(0.5),
nn.Linear(512, num_classes)
)
def forward(self, x):
x = self.features(x)
x = self.classifier(x)
return x# 训练配置
training_config = {
'optimizer': 'AdamW',
'learning_rate': 1e-4,
'batch_size': 32,
'epochs': 100,
'loss_function': 'CrossEntropyLoss',
'scheduler': 'CosineAnnealingLR',
'data_augmentation': True,
'early_stopping': True,
'patience': 10
}基于 train_thermal_classifier3.py 实现:
# 对比学习模型
class ContrastiveThermalClassifier(nn.Module):
def __init__(self, backbone='resnet50', projection_dim=128):
super().__init__()
# 骨干网络
self.backbone = self._build_backbone(backbone)
# 投影头
self.projection_head = nn.Sequential(
nn.Linear(2048, 512),
nn.ReLU(),
nn.Linear(512, projection_dim)
)
# 分类头
self.classification_head = nn.Linear(2048, 2)
def forward(self, x, mode='classification'):
features = self.backbone(x)
if mode == 'contrastive':
return self.projection_head(features)
else:
return self.classification_head(features)# 阶段1: 对比学习预训练
def contrastive_pretraining():
for epoch in range(contrastive_epochs):
for batch in contrastive_dataloader:
# 数据增强生成正负样本对
anchor, positive, negative = generate_triplets(batch)
# 特征提取
anchor_features = model(anchor, mode='contrastive')
positive_features = model(positive, mode='contrastive')
negative_features = model(negative, mode='contrastive')
# 对比损失
loss = contrastive_loss(anchor_features, positive_features, negative_features)
# 反向传播
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 阶段2: 分类微调
def classification_finetuning():
# 冻结骨干网络部分参数
freeze_backbone_layers(model.backbone, freeze_ratio=0.8)
for epoch in range(finetuning_epochs):
for batch in classification_dataloader:
images, labels = batch
# 分类预测
predictions = model(images, mode='classification')
# 分类损失
loss = classification_loss(predictions, labels)
# 反向传播
optimizer.zero_grad()
loss.backward()
optimizer.step()# 评估指标计算
def evaluate_model(y_true, y_pred, y_prob):
metrics = {
'accuracy': accuracy_score(y_true, y_pred),
'precision': precision_score(y_true, y_pred, average='weighted'),
'recall': recall_score(y_true, y_pred, average='weighted'),
'f1_score': f1_score(y_true, y_pred, average='weighted'),
'auc': roc_auc_score(y_true, y_prob[:, 1]),
'confusion_matrix': confusion_matrix(y_true, y_pred)
}
return metrics| 方法 | 最佳模型 | 测试准确率 | 测试F1分数 | 测试AUC | 训练时间 | 特征维度 |
|---|---|---|---|---|---|---|
| 多模态融合 | LightGBM | 74.83% | 53.85% | 75.46% | 28.51秒 | 42维 |
| 传统特征 | LogisticRegression | 51.75% | 37.84% | 54.14% | 28.12秒 | 26维 |
| 深度学习 | ThermalCNN | 待测试 | - | - | - | - |
| 对比学习 | ContrastiveNet | 待测试 | - | - | - | - |
# 模型选择决策树
def select_best_model(performance_metrics, requirements):
if requirements['accuracy_priority']:
return max(performance_metrics, key=lambda x: x['accuracy'])
elif requirements['speed_priority']:
return min(performance_metrics, key=lambda x: x['inference_time'])
elif requirements['interpretability_priority']:
return filter_interpretable_models(performance_metrics)
else:
# 综合评分
return calculate_weighted_score(performance_metrics)# 模型部署管道
class ICASPredictionPipeline:
def __init__(self, face_detector, feature_extractor, classifier):
self.face_detector = face_detector
self.feature_extractor = feature_extractor
self.classifier = classifier
def predict(self, thermal_image, clinical_data=None):
# 1. 人脸检测
face_regions = self.face_detector.detect(thermal_image)
if not face_regions:
return {'error': 'No face detected'}
# 2. 特征提取
if clinical_data is not None:
features = self.feature_extractor.extract_multimodal(
face_regions[0], clinical_data
)
else:
features = self.feature_extractor.extract_image_only(
face_regions[0]
)
# 3. 风险预测
risk_score = self.classifier.predict_proba(features)[0][1]
risk_level = self._categorize_risk(risk_score)
return {
'risk_score': risk_score * 100,
'risk_level': risk_level,
'confidence': self.classifier.predict_proba(features).max(),
'face_detected': True
}
def _categorize_risk(self, score):
if score < 0.3:
return 'low'
elif score < 0.7:
return 'medium'
else:
return 'high'# Flask API接口
@app.route('/api/predict/icas', methods=['POST'])
def predict_icas():
# 1. 接收上传的热力图
thermal_image = request.files['thermal_image']
clinical_data = request.json.get('clinical_data', None)
# 2. 图像预处理
image = preprocess_uploaded_image(thermal_image)
# 3. 模型预测
prediction = pipeline.predict(image, clinical_data)
# 4. 返回结果
return jsonify(prediction)# 图像质量检查
def quality_control_pipeline(image):
checks = {
'resolution_check': check_image_resolution(image),
'brightness_check': check_brightness_range(image),
'contrast_check': check_contrast_level(image),
'noise_check': check_noise_level(image),
'face_visibility_check': check_face_visibility(image)
}
quality_score = calculate_quality_score(checks)
return quality_score > 0.7 # 质量阈值# 在线学习与模型更新
class ModelUpdatePipeline:
def __init__(self):
self.feedback_buffer = []
self.update_threshold = 100 # 累积100个反馈后更新
def collect_feedback(self, prediction, ground_truth):
self.feedback_buffer.append({
'prediction': prediction,
'ground_truth': ground_truth,
'timestamp': datetime.now()
})
if len(self.feedback_buffer) >= self.update_threshold:
self.update_model()
def update_model(self):
# 增量学习更新模型
new_data = self.prepare_training_data(self.feedback_buffer)
self.model.partial_fit(new_data['X'], new_data['y'])
# 清空缓冲区
self.feedback_buffer = []- 创新点: 首次将热力图特征与临床特征深度融合
- 技术优势: 相比单一图像特征,准确率提升23.08%
- 应用价值: 更符合临床诊断的多维度评估需求
- 创新点: 4×4网格分块保留空间信息的LBP特征提取
- 技术优势: 288维分块特征提供丰富的局部纹理描述
- 应用价值: 捕捉人脸热力图的细微纹理变化
- 创新点: 针对医学热力图的对比学习策略
- 技术优势: 无监督预训练提升特征表达能力
- 应用价值: 减少对标注数据的依赖
本算法流程通过系统性的数据处理、模型训练和性能优化,构建了一个完整的ICAS热力图分析系统。多模态特征融合方法在当前数据集上表现最佳,为临床ICAS风险评估提供了有效的技术支持。
关键成果:
- 构建了950张热力图的标准化数据集
- 实现了74.83%的分类准确率
- 开发了多种可选的分类方法
- 建立了完整的部署应用流程
临床价值:
- 提高ICAS早期筛查效率
- 辅助医生临床决策
- 降低漏诊误诊风险
- 支持个性化医疗建议
文档版本: v1.0
创建日期: 2025年1月
维护团队: ICAS算法开发团队