A Mobile Application for the Diagnosis of Facial Nerve Palsy

Introduction – Facial Nerve Palsy

• Facial nerve palsy is a problem that involves the paralysis of any structures activated by the facial nerve
• Diagnostics today are very subjective
• There is a need for an objective method for evaluation of patients (sick/healthy, severity)

Review – Previous Project

• Finding facial keypoints using stickers
• Extracting facial features of right and left sides of face and comparing them
• Using machine learning algorithms todiagnose, based on features
• Establishing that using facial keypoints is useful for a stable diagnostics

Project Goals

• Finding facial keypoints without stickers
• Adding new features
• Performing diagnostics (sick/healthy) using the features found
• Improving diagnostics accuracy
• Implementing the new algorithm as a mobile application

Algorithm Requirements

• Fast computation time – for real time applications
• Robust to:
  • Face size
  • Slight rotation of face
  • Different facial expressions
• Low error between detected keypoints and actual facial features

Possible Solutions

• Active Appearance Model (AAM)
• Constrained Local Models (CLM)
• Extracting features using basic image processing tools

AAM – Active Appearance Model

• Creating a statistical model of face appearances: shape and texture (gray level)
• Matching the Appearance Model to new and unseen images
  • Appearance Model consisted shape and texture parameters
  • Update parameters iteratively
  • Goal: minimizing the difference between a real image and one synthesized by the model

AAM – Active Appearance Model -Results

CLM – Constrained Local Model

• Face model consisted of a face shape model and different patch models for different facial parts (around keypoints)
• Matching patches to given image in order to best locate keypoints while maintaining shape constraints (eyes above nose, nose above mouth…)
• Theoretically, will allow higher flexibility, less constrained by symmetry
• Implementation by Prof. Tim Cootes of the University of Manchester.

CLM – Constrained Local Model – Results

Extracting Features Using Basic Tools

• Using Viola and Jones to locate face, nose, eyes, mouth
• Using segmentation in different color channels and edge detection to extract interest points

Extracting Features Using Basic Tools -Results

Summery

• Algorithms for facial key point extractions exist, but do not work well for our cause:
• High computation time
• Relies on symmetry (Models based on healthy subjects)
• Found 12 key points using basic image processing tools
• Generated good results on specific subjects, need adjustments for different variations of:
  • Skin color
  • Hair color
  • Beard/mustache
  • Facial proportions

For Future Examination

• Building face models using examples from both healthy and ill subjects (for AAM or CLM)
• Creating a table of different parameters and thresholds for different subjects
• Using 3D analysis, depth images