A deep learning-based diagnostic tool for identifying various diseases via facial images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• The COVID-19 pandemic increased patient reluctance for in-person consultations.
• Diseases can manifest distinct facial features, suggesting potential for remote diagnosis.
• Existing computer-aided facial diagnosis systems often lack deep learning integration and rely on controlled environments.

Purpose of the Study:

• To develop a novel deep learning system, FaceDisNet, for computer-aided facial diagnosis.
• To create a public dataset of facial images from unconstrained environments for disease detection.
• To enable the detection of both single and multiple diseases from facial images.

Main Methods:

• FaceDisNet integrates spatial and spatial-spectral features from various convolutional neural network architectures.
• Feature selection techniques were employed to reduce dimensionality after feature fusion.
• An ensemble classifier based on stacking was utilized for final disease classification.

Main Results:

• FaceDisNet demonstrated effectiveness in diagnosing both single and multiple diseases.
• The system achieved a maximum accuracy of 98.57% for binary classification and 98% for multiclass classification.
• Performance was validated after ensemble classification and feature selection steps.

Conclusions:

• FaceDisNet serves as a reliable tool for automated facial diagnosis, mitigating challenges associated with manual methods.
• The system facilitates accurate diagnoses without direct physical patient contact, supporting remote healthcare.
• This approach offers a promising avenue for disease screening and physician support in a contactless manner.