train-modle / src /medical /medical_preprocessing.py
fokan's picture
Initial clean commit: Multi-Modal Knowledge Distillation Platform
ab4e093
"""
Medical Data Preprocessing for AI training
Optimized for medical images and text with memory constraints
"""
import logging
import numpy as np
from typing import Dict, Any, List, Optional, Tuple
import torch
import torch.nn.functional as F
from PIL import Image, ImageEnhance, ImageFilter
import cv2
import re
logger = logging.getLogger(__name__)
class MedicalPreprocessor:
"""
Medical data preprocessor with memory optimization
"""
def __init__(self, target_size: Tuple[int, int] = (512, 512),
normalize_images: bool = True):
"""
Initialize medical preprocessor
Args:
target_size: Target size for image resizing
normalize_images: Whether to normalize images
"""
self.target_size = target_size
self.normalize_images = normalize_images
# Medical text preprocessing patterns
self.medical_patterns = {
'measurements': r'\d+\.?\d*\s*(mm|cm|m|ml|l|kg|g|mg)',
'dates': r'\d{1,2}[/-]\d{1,2}[/-]\d{2,4}',
'times': r'\d{1,2}:\d{2}(?::\d{2})?',
'medical_codes': r'[A-Z]\d{2}\.?\d*',
'dosages': r'\d+\.?\d*\s*(mg|g|ml|units?)',
}
# Common medical abbreviations
self.medical_abbreviations = {
'pt': 'patient',
'pts': 'patients',
'dx': 'diagnosis',
'tx': 'treatment',
'hx': 'history',
'sx': 'symptoms',
'rx': 'prescription',
'w/': 'with',
'w/o': 'without',
'c/o': 'complains of',
'r/o': 'rule out',
's/p': 'status post',
'nkda': 'no known drug allergies',
'sob': 'shortness of breath',
'cp': 'chest pain',
'abd': 'abdomen',
'ext': 'extremities'
}
logger.info(f"Medical Preprocessor initialized with target size {target_size}")
def preprocess_medical_image(self, image: torch.Tensor,
modality: str = 'unknown',
enhance_contrast: bool = True) -> torch.Tensor:
"""
Preprocess medical image with modality-specific optimizations
Args:
image: Input image tensor
modality: Medical imaging modality (CT, MRI, X-ray, etc.)
enhance_contrast: Whether to enhance contrast
Returns:
Preprocessed image tensor
"""
try:
# Ensure image is float tensor
if image.dtype != torch.float32:
image = image.float()
# Handle different input shapes
if len(image.shape) == 2:
image = image.unsqueeze(0) # Add channel dimension
elif len(image.shape) == 4:
image = image.squeeze(0) # Remove batch dimension if present
# Resize to target size
if image.shape[-2:] != self.target_size:
image = F.interpolate(
image.unsqueeze(0),
size=self.target_size,
mode='bilinear',
align_corners=False
).squeeze(0)
# Apply modality-specific preprocessing
image = self._apply_modality_specific_processing(image, modality)
# Enhance contrast if requested
if enhance_contrast:
image = self._enhance_medical_image_contrast(image)
# Normalize if requested
if self.normalize_images:
image = self._normalize_medical_image(image)
# Ensure proper range [0, 1]
image = torch.clamp(image, 0.0, 1.0)
return image
except Exception as e:
logger.error(f"Error preprocessing medical image: {e}")
# Return dummy image on error
return torch.zeros(1, *self.target_size)
def _apply_modality_specific_processing(self, image: torch.Tensor,
modality: str) -> torch.Tensor:
"""Apply modality-specific image processing"""
modality_lower = modality.lower()
try:
if 'ct' in modality_lower:
# CT scan specific processing
image = self._process_ct_image(image)
elif 'mri' in modality_lower:
# MRI specific processing
image = self._process_mri_image(image)
elif 'xray' in modality_lower or 'x-ray' in modality_lower:
# X-ray specific processing
image = self._process_xray_image(image)
elif 'ultrasound' in modality_lower:
# Ultrasound specific processing
image = self._process_ultrasound_image(image)
return image
except Exception as e:
logger.warning(f"Error in modality-specific processing for {modality}: {e}")
return image
def _process_ct_image(self, image: torch.Tensor) -> torch.Tensor:
"""Process CT scan images"""
# CT images often need windowing adjustments
# Apply soft tissue window as default
image = torch.clamp(image, 0.0, 1.0)
# Enhance contrast for better tissue differentiation
image = self._apply_gamma_correction(image, gamma=0.8)
return image
def _process_mri_image(self, image: torch.Tensor) -> torch.Tensor:
"""Process MRI images"""
# MRI images often have good contrast already
# Apply mild enhancement
image = self._apply_gamma_correction(image, gamma=0.9)
return image
def _process_xray_image(self, image: torch.Tensor) -> torch.Tensor:
"""Process X-ray images"""
# X-rays often need contrast enhancement
image = self._enhance_medical_image_contrast(image, factor=1.2)
# Apply histogram equalization equivalent
image = self._apply_histogram_equalization(image)
return image
def _process_ultrasound_image(self, image: torch.Tensor) -> torch.Tensor:
"""Process ultrasound images"""
# Ultrasound images often need noise reduction
image = self._apply_noise_reduction(image)
return image
def _enhance_medical_image_contrast(self, image: torch.Tensor,
factor: float = 1.1) -> torch.Tensor:
"""Enhance contrast of medical images"""
try:
# Apply contrast enhancement
mean_val = torch.mean(image)
enhanced = (image - mean_val) * factor + mean_val
return torch.clamp(enhanced, 0.0, 1.0)
except Exception as e:
logger.warning(f"Error enhancing contrast: {e}")
return image
def _apply_gamma_correction(self, image: torch.Tensor,
gamma: float = 1.0) -> torch.Tensor:
"""Apply gamma correction to image"""
try:
return torch.pow(image, gamma)
except Exception as e:
logger.warning(f"Error applying gamma correction: {e}")
return image
def _apply_histogram_equalization(self, image: torch.Tensor) -> torch.Tensor:
"""Apply histogram equalization equivalent"""
try:
# Convert to numpy for processing
image_np = image.squeeze().numpy()
# Apply CLAHE (Contrast Limited Adaptive Histogram Equalization)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
# Convert to uint8 for CLAHE
image_uint8 = (image_np * 255).astype(np.uint8)
equalized = clahe.apply(image_uint8)
# Convert back to tensor
result = torch.from_numpy(equalized.astype(np.float32) / 255.0)
# Restore original shape
if len(image.shape) == 3:
result = result.unsqueeze(0)
return result
except Exception as e:
logger.warning(f"Error applying histogram equalization: {e}")
return image
def _apply_noise_reduction(self, image: torch.Tensor) -> torch.Tensor:
"""Apply noise reduction to image"""
try:
# Simple Gaussian blur for noise reduction
kernel_size = 3
sigma = 0.5
# Create Gaussian kernel
kernel = self._create_gaussian_kernel(kernel_size, sigma)
kernel = kernel.unsqueeze(0).unsqueeze(0) # Add batch and channel dims
# Apply convolution
if len(image.shape) == 3:
image_input = image.unsqueeze(0) # Add batch dimension
else:
image_input = image
filtered = F.conv2d(image_input, kernel, padding=kernel_size//2)
# Remove batch dimension if added
if len(image.shape) == 3:
filtered = filtered.squeeze(0)
return filtered
except Exception as e:
logger.warning(f"Error applying noise reduction: {e}")
return image
def _create_gaussian_kernel(self, kernel_size: int, sigma: float) -> torch.Tensor:
"""Create Gaussian kernel for filtering"""
coords = torch.arange(kernel_size, dtype=torch.float32)
coords -= kernel_size // 2
g = torch.exp(-(coords ** 2) / (2 * sigma ** 2))
g /= g.sum()
# Create 2D kernel
kernel = g[:, None] * g[None, :]
return kernel
def _normalize_medical_image(self, image: torch.Tensor) -> torch.Tensor:
"""Normalize medical image"""
try:
# Z-score normalization per image
mean_val = torch.mean(image)
std_val = torch.std(image)
if std_val > 0:
normalized = (image - mean_val) / std_val
# Scale to [0, 1] range
normalized = (normalized - normalized.min()) / (normalized.max() - normalized.min())
else:
normalized = image
return normalized
except Exception as e:
logger.warning(f"Error normalizing image: {e}")
return image
def preprocess_medical_text(self, text: str,
expand_abbreviations: bool = True,
remove_phi: bool = True) -> str:
"""
Preprocess medical text
Args:
text: Input medical text
expand_abbreviations: Whether to expand medical abbreviations
remove_phi: Whether to remove potential PHI (Protected Health Information)
Returns:
Preprocessed text
"""
try:
if not isinstance(text, str):
text = str(text)
# Convert to lowercase for processing
processed_text = text.lower()
# Remove potential PHI if requested
if remove_phi:
processed_text = self._remove_phi(processed_text)
# Expand medical abbreviations
if expand_abbreviations:
processed_text = self._expand_medical_abbreviations(processed_text)
# Clean up text
processed_text = self._clean_medical_text(processed_text)
# Limit length to prevent memory issues
max_length = 2048
if len(processed_text) > max_length:
processed_text = processed_text[:max_length] + "..."
return processed_text
except Exception as e:
logger.error(f"Error preprocessing medical text: {e}")
return text # Return original text on error
def _remove_phi(self, text: str) -> str:
"""Remove potential Protected Health Information"""
# Remove dates
text = re.sub(self.medical_patterns['dates'], '[DATE]', text)
# Remove times
text = re.sub(self.medical_patterns['times'], '[TIME]', text)
# Remove phone numbers
text = re.sub(r'\b\d{3}[-.]?\d{3}[-.]?\d{4}\b', '[PHONE]', text)
# Remove email addresses
text = re.sub(r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b', '[EMAIL]', text)
# Remove potential names (very basic - would need more sophisticated NER in practice)
text = re.sub(r'\b[A-Z][a-z]+ [A-Z][a-z]+\b', '[NAME]', text)
return text
def _expand_medical_abbreviations(self, text: str) -> str:
"""Expand common medical abbreviations"""
for abbrev, expansion in self.medical_abbreviations.items():
# Use word boundaries to avoid partial matches
pattern = r'\b' + re.escape(abbrev) + r'\b'
text = re.sub(pattern, expansion, text, flags=re.IGNORECASE)
return text
def _clean_medical_text(self, text: str) -> str:
"""Clean and normalize medical text"""
# Remove extra whitespace
text = re.sub(r'\s+', ' ', text)
# Remove special characters but keep medical-relevant ones
text = re.sub(r'[^\w\s\-\.\,\:\;\(\)\/\%]', '', text)
# Strip leading/trailing whitespace
text = text.strip()
return text
def batch_preprocess_medical_data(self, batch: Dict[str, Any]) -> Dict[str, Any]:
"""Preprocess a batch of medical data"""
processed_batch = {}
try:
# Process images if present
if 'images' in batch and batch['images'] is not None:
images = batch['images']
processed_images = []
for i, image in enumerate(images):
# Get modality if available
modality = 'unknown'
if 'modalities' in batch and i < len(batch['modalities']):
modality = batch['modalities'][i]
processed_image = self.preprocess_medical_image(image, modality)
processed_images.append(processed_image)
processed_batch['images'] = torch.stack(processed_images)
# Process texts if present
if 'texts' in batch:
texts = batch['texts']
processed_texts = []
for text in texts:
processed_text = self.preprocess_medical_text(text)
processed_texts.append(processed_text)
processed_batch['texts'] = processed_texts
# Copy other fields
for key, value in batch.items():
if key not in ['images', 'texts']:
processed_batch[key] = value
return processed_batch
except Exception as e:
logger.error(f"Error in batch preprocessing: {e}")
return batch # Return original batch on error