Deep Learning-Based Analysis of Face Images as a Screening Tool for Genetic Syndromes

Area of Science:

• Medical Genetics
• Artificial Intelligence
• Computer Vision

Background:

• Rare genetic disorders affect approximately 4% of the global population, with most having a genetic basis.
• Identifying novel genetic syndromes and improving diagnostic yields are crucial due to the growing number of identified disease-associated genes.
• Accurate patient selection for genetic testing is vital given limited resources like clinical geneticists.

Purpose of the Study:

• To evaluate the performance of deep learning face recognition models in detecting dysmorphic features associated with genetic disorders.
• To assess the utility of these models in both multiclass (15 disorders vs. controls) and binary (disease vs. controls) classification tasks.
• To determine if the models can generalize and detect abnormalities in previously unknown or uncatalogued genetic disorders.

Main Methods:

• Utilized deep learning-based face recognition models for image analysis.
• Trained and tested classifiers on two distinct problems: a multiclass classification of 15 genetic disorders against controls, and a binary classification of disease versus controls.
• Evaluated the accuracy and generalization capabilities of the models, particularly their ability to identify novel disease presentations.

Main Results:

• The multiclass classifier achieved a maximum accuracy of 84% in identifying 15 genetic disorders.
• The binary classifier demonstrated a high accuracy of 96% in distinguishing between patients with genetic disorders and controls.
• Crucially, the binary classifier successfully identified dysmorphic features in patients with diseases not included in the training data, indicating strong generalization.

Conclusions:

• Deep learning models employing facial recognition are effective tools for detecting dysmorphic features linked to genetic disorders.
• The developed classifier exhibits the potential to screen for both known and previously unknown genetic syndromes by recognizing generalized facial abnormalities.
• This technology could serve as a valuable pre-screening tool, aiding in the efficient referral of patients to specialized genetic units.