YAML
---
name: medical-motion-detection-registration
description: Builds simple, explainable 2D motion detection pipelines for medical radiation safety using fixed-reference image registration, displacement fields, warped residuals, ROI/head masks, and temporal stage logic. Use when developing deformation-based motion detection systems for segmentation control and radiation stop/recovery decisions.
---

Medical Motion Detection Registration

Image Registration:

Python
def register_images(reference, current):
    # Lucas-Kanade optical flow for dense registration
    flow = cv2.calcOpticalFlowPyrLK(
        reference.astype(np.uint8), 
        current.astype(np.uint8),
        corners, None, **lk_params
    )
    
    # Convert sparse flow to dense displacement field
    displacement_field = interpolate_flow(flow, reference.shape)
    return displacement_field

def apply_displacement(image, displacement_field):
    h, w = image.shape
    map_x = np.arange(w) + displacement_field[:, :, 0]
    map_y = np.arange(h) + displacement_field[:, :, 1]
    warped = cv2.remap(image, map_x, map_y, cv2.INTER_LINEAR)
    return warped

ROI Mask Generation:

Python
def create_head_mask(image, method='threshold'):
    if method == 'threshold':
        # Simple intensity thresholding
        mask = image > np.percentile(image, 20)
    elif method == 'contour':
        # Find largest contour (assumes head is largest object)
        contours, _ = cv2.findContours(
            (image > threshold).astype(np.uint8), 
            cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
        )
        largest_contour = max(contours, key=cv2.contourArea)
        mask = cv2.fillPoly(np.zeros_like(image), [largest_contour], 1)
    
    # Morphological cleanup
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
    mask = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_CLOSE, kernel)
    return mask.astype(bool)

Temporal Stage Logic:

Python
class MotionDetector:
    def __init__(self, motion_threshold=5.0, hysteresis_factor=0.8):
        self.threshold = motion_threshold
        self.resume_threshold = motion_threshold * hysteresis_factor
        self.motion_history = []
        self.is_stopped = False
        
    def update(self, motion_score):
        self.motion_history.append(motion_score)
        if len(self.motion_history) > 10:  # Keep last 10 frames
            self.motion_history.pop(0)
        
        # Temporal smoothing
        smoothed_score = np.median(self.motion_history[-3:])  # 3-frame median
        
        # State machine logic
        if not self.is_stopped and smoothed_score > self.threshold:
            self.is_stopped = True
            return "STOP_RADIATION"
        elif self.is_stopped and smoothed_score < self.resume_threshold:
            self.is_stopped = False
            return "RESUME_RADIATION"
        
        return "CONTINUE"

Registration Quality:

• Use pyramid-based registration for large motions
• Validate registration with inverse consistency check
• Handle registration failures gracefully (flag bad frames)

Threshold Setting:

• Start with 2-3mm for cranial applications
• Use different thresholds for different anatomical regions
• Implement adaptive thresholds based on image quality

Temporal Filtering:

• Use median filter (3-5 frames) to reject outliers
• Implement exponential moving average for trend detection
• Apply different time constants for motion onset vs. recovery

Safety Margins:

• Always err on the side of radiation safety
• Implement watchdog timer for system failures
• Log all motion events for quality assurance

Registration Artifacts:

• Don't use registration on low-contrast regions
• Avoid over-smoothing that masks real motion
• Check for registration convergence failures

False Positives:

• Breathing motion can trigger false alarms - use respiratory gating
• Image noise changes between frames - apply consistent preprocessing
• Intensity variations from beam hardening - normalize images

Timing Issues:

• Don't use overly aggressive temporal filtering - delays detection
• Avoid hysteresis bands that are too narrow - causes oscillation
• System latency must be accounted for in safety margins

Mask Problems:

• Static masks don't account for patient positioning changes
• Over-restrictive ROIs can miss important motion
• Under-inclusive masks allow background noise to dominate