Facial Scanning With a Digital Camera

Harold Choi, Weisi Lin, Seng Chee Loon, Clement Tan, Wanling Wong, Jovina See, Zhouye Gu, Chee Keong Kwoh, Paul ChewJournal of Glaucoma 2013 · DOI

TL;DR

This paper explores the potential of 3D facial scanning to identify individuals at risk for Primary Angle Closure (PAC) glaucoma. By analyzing specific facial parameters captured by a 3D camera, researchers developed a machine learning model that can distinguish PAC patients from healthy controls with high accuracy, suggesting a novel, non-invasive screening tool.

What Problem Does This Paper Address?

Primary Angle Closure (PAC) glaucoma is a significant cause of blindness, particularly in Asian populations. While risk factors like age and sex are known, and ocular factors like shallow anterior chambers are identified, there's a lack of understanding regarding distinct facial characteristics associated with PAC. Current screening methods for PAC can be time-consuming or require specialized equipment. This study aims to bridge this gap by investigating whether common, significant facial characteristics in PAC patients can be identified using 3D facial scanning technology, potentially leading to a more accessible and efficient screening method.

Methods

The researchers employed a case-control study design. They utilized a 3dMDface System, a 3D digital camera setup, to capture high-resolution facial images of both PAC patients and healthy controls. This system uses "active stereo photogrammetry" with industrial-grade hardware and projected light patterns to create detailed 3D facial models.

Facial features were then analyzed using a Neural Network Toolbox in MATLAB. Specifically, they extracted thirteen facial parameters (e.g., distances between facial landmarks like canthi, palpebral apertures, facial height, and width) and calculated ratios of some of these parameters. A machine learning approach, specifically a 3-layer neural network (input, hidden, output layers), was trained to classify individuals as either PAC patients or controls based on these facial parameters. The network used hyperbolic tangent sigmoid and log-sigmoid transfer functions and was trained using a back-propagation algorithm. A 10-fold cross-validation was performed to ensure robust evaluation of the model's performance.

Datasets

The study used a dataset comprising:

  • 55 patients with Primary Angle Closure (PAC)
  • 38 control subjects

These subjects were recruited from the National University Hospital Department of Ophthalmology. The paper does not mention the use of any pre-existing public datasets for training or evaluation.

Key Insights and Ideas

  • Facial Morphology as a Biomarker: The core idea is that PAC might be associated with subtle, yet quantifiable, differences in facial structure that can be captured by advanced imaging.
  • 3D Photography for Medical Screening: The study demonstrates the feasibility of using 3D facial scanning, a technology with rapid advancements, for medical screening purposes beyond its traditional applications in plastic surgery.
  • Machine Learning for Complex Pattern Recognition: The application of neural networks highlights the power of machine learning in identifying complex, non-linear relationships between facial features and a specific medical condition.
  • Potential for Non-Invasive Screening: The findings suggest a future where screening for PAC could be as simple as taking a 3D photograph, making it more accessible and less intimidating for patients.

Key Results

  • The study successfully developed a machine learning model capable of distinguishing PAC patients from controls using 3D facial parameters.
  • The model achieved an accuracy of 86.1% and a sensitivity of 93.0%.
  • A confusion matrix revealed 53 true positives, 28 true negatives, 9 false positives, and 4 false negatives.
  • While individual facial parameters did not show statistically significant differences between groups, the combination of these parameters through the neural network model proved effective.
  • The mean age of the PAC patient group (68.8 years) was statistically significantly different from the control group (62.0 years), but the sex and race distributions were well-matched.

Limitations and Future Work

Limitations: * Small Sample Size: The authors acknowledge that this is a pilot study with a limited sample size, necessitating larger studies for confirmation. * Observer Bias: The researchers involved in photography and data management were not masked, potentially introducing observer bias. * Spurious Features: Neural networks can sometimes identify spurious features that may not generalize well to independent populations. * Age Matching: While subjects were adults, the PAC group was statistically older than the control group, which could have a minimal impact. * External Factors: Factors like obesity and facial expression could influence facial measurements, although efforts were made to standardize.

Future Work: * Conducting larger-scale studies to confirm the findings. * Developing a facial recognition software that incorporates these identified characteristics for a quick, economical, and effective screening tool for PAC. * Investigating how facial expressions influence facial parameters and their relationship to PAC. * Correlating these findings with other decision-making tools for a more comprehensive evaluation.

Summary updated

Abstract

PAC patients have distinct facial features that may be differentiated from those of normal individuals mathematically using 3D photography. There is potential for this 3D facial scanning technology in screening for PAC in the general population.
BibTeX
@article{choi2013facial,
  author = {Harold Choi and Weisi Lin and Seng Chee Loon and Clement Tan and Wanling Wong and Jovina See and Zhouye Gu and Chee Keong Kwoh and Paul Chew},
  title = {{Facial Scanning With a Digital Camera}},
  journal = {Journal of Glaucoma},
  year = {2013},
  doi = {10.1097/ijg.0000000000000017},
}
Figure