Automated Multi-Class Facial Syndrome Classification Using Transfer Learning Techniques.

Area of Science:

• Medical Genetics
• Artificial Intelligence
• Computer Vision

Background:

• Genetic disorders impact over 6% globally, necessitating early diagnosis for effective management.
• Current screening methods for rare facial genetic disorders are often insufficient, delaying diagnosis.
• Facial dysmorphic features are key indicators of many genetic syndromes.

Purpose of the Study:

• To evaluate deep learning models for recognizing dysmorphic characteristics in facial photographs.
• To develop and assess a multi-class facial syndrome classification framework for specific genetic disorders.
• To compare the performance of various pre-trained convolutional neural network (CNN) models for this task.

Main Methods:

• Fine-tuning pre-trained CNN models: VGG16, ResNet-50, ResNet152, and VGG-Face.
• Utilizing facial photographs for multi-class classification of Down syndrome, Noonan syndrome, Turner syndrome, Williams syndrome, and healthy controls.
• Evaluating model performance based on accuracy and F1-Score.

Main Results:

• The fine-tuned VGG-Face model demonstrated superior performance compared to other evaluated CNN models.
• The VGG-Face model achieved an accuracy of 90% in classifying the specified genetic disorders.
• An F1-Score of 90% was attained, indicating robust detection capabilities.

Conclusions:

• Deep learning, particularly the fine-tuned VGG-Face model, shows significant promise for accurate and early detection of specific genetic disorders using facial images.
• This approach offers a potential advancement over existing screening techniques, facilitating timely medical intervention.
• The study highlights the efficacy of AI in analyzing complex visual patterns for diagnosing rare genetic conditions.