bone_2026/src/models/region_detector.py

"""
Region Detector - определяет анатомическую область исследования
"""
import torch
import torch.nn as nn
import torchvision.models as models
from typing import Dict, List, Tuple, Optional
import numpy as np


class RegionDetector(nn.Module):
    """
    Определение анатомической области исследования.
    Классы: spine, hip_left, hip_right, hip (unknown side)
    """
    
    REGIONS = ['spine', 'hip_left', 'hip_right', 'hip']
    
    def __init__(self,
                 backbone: str = 'efficientnet_b0',
                 num_classes: int = 4,
                 pretrained: bool = True,
                 dropout: float = 0.3):
        super().__init__()
        
        self.backbone_name = backbone
        self.num_classes = num_classes
        
        # Загрузка backbone
        if backbone == 'efficientnet_b0':
            self.backbone = models.efficientnet_b0(
                weights='IMAGENET1K_V1' if pretrained else None
            )
            feature_dim = 1280
            self.backbone.classifier = nn.Identity()
        elif backbone == 'resnet18':
            self.backbone = models.resnet18(
                weights='IMAGENET1K_V1' if pretrained else None
            )
            feature_dim = 512
            self.backbone.fc = nn.Identity()
        else:
            raise ValueError(f"Unknown backbone: {backbone}")
        
        # Классификатор
        self.classifier = nn.Sequential(
            nn.Dropout(dropout),
            nn.Linear(feature_dim, 256),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(256, num_classes)
        )
        
        self.feature_dim = feature_dim
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        features = self.backbone(x)
        return self.classifier(features)
    
    def predict(self, x: torch.Tensor) -> Dict:
        """
        Предсказание региона.
        
        Args:
            x: Input tensor [B, 3, H, W]
            
        Returns:
            Dict с keys: predicted_region, probabilities, confidence
        """
        self.eval()
        with torch.no_grad():
            logits = self.forward(x)
            probs = torch.softmax(logits, dim=1)
            conf, pred = probs.max(dim=1)
            
            results = []
            for i in range(x.shape[0]):
                region_idx = pred[i].item()
                results.append({
                    'predicted_region': self.REGIONS[region_idx],
                    'region_index': region_idx,
                    'confidence': conf[i].item(),
                    'probabilities': {
                        self.REGIONS[j]: probs[i, j].item()
                        for j in range(self.num_classes)
                    }
                })
        
        return results[0] if len(results) == 1 else results
    
    def get_info(self) -> Dict:
        return {
            'name': 'RegionDetector',
            'backbone': self.backbone_name,
            'num_classes': self.num_classes,
            'regions': self.REGIONS,
            'feature_dim': self.feature_dim
        }


class RegionDetectorWithFeatures(RegionDetector):
    """Region detector с извлечением признаков"""
    
    def extract_features(self, x: torch.Tensor) -> torch.Tensor:
        return self.backbone(x)


def create_region_detector(
    backbone: str = 'efficientnet_b0',
    num_classes: int = 4,
    pretrained: bool = True,
    device: str = 'cpu'
) -> RegionDetector:
    """Создание модели"""
    model = RegionDetector(
        backbone=backbone,
        num_classes=num_classes,
        pretrained=pretrained
    )
    model.to(device)
    return model


# Функция для определения региона из изображения (rule-based fallback)
def determine_region_from_image(image: np.ndarray) -> str:
    """
    Определение региона из изображения без модели.
    Использует форму яркой области.
    
    Args:
        image: numpy array [H, W] или [H, W, 3]
        
    Returns:
        str: 'spine', 'hip_left', 'hip_right', 'hip'
    """
    if len(image.shape) == 3:
        image = np.mean(image, axis=2)
    
    # Нормализация
    image = (image - image.min()) / (image.max() - image.min() + 1e-8)
    
    h, w = image.shape
    
    # Порог для яркой области
    threshold = np.percentile(image, 95)
    binary = image > threshold
    
    if binary.sum() == 0:
        return 'unknown'
    
    # Aspect ratio яркой области
    from scipy import ndimage
    rows = np.any(binary, axis=1)
    cols = np.any(binary, axis=0)
    
    if not (rows.any() and cols.any()):
        return 'unknown'
    
    try:
        rmin, rmax = np.where(rows)[0][[0, -1]]
        cmin, cmax = np.where(cols)[0][[0, -1]]
        
        bbox_h = rmax - rmin
        bbox_w = cmax - cmin
        bbox_aspect = bbox_h / (bbox_w + 1e-6)
        
        # Симметрия
        h_mid, w_mid = h // 2, w // 2
        left_half = image[:, :w_mid]
        right_half = np.fliplr(image[:, w_mid:])
        min_w = min(left_half.shape[1], right_half.shape[1])
        symmetry = 1 - np.abs(left_half[:, :min_w] - right_half[:, :min_w]).mean() / (image.std() + 1e-6)
        
        # Left/right brightness ratio
        left_bright = binary[:, :w//2].sum()
        right_bright = binary[:, w//2:].sum()
        left_right_ratio = left_bright / (right_bright + 1e-6)
        
        # Классификация
        if bbox_aspect < 1.5:
            return 'spine'
        elif bbox_aspect < 1.8:
            return 'spine' if symmetry > 0.35 else 'hip'
        else:
            if left_right_ratio > 1.3:
                return 'hip_right'
            elif left_right_ratio < 0.7:
                return 'hip_left'
            else:
                return 'hip'
    except:
        return 'unknown'