Automatic Detection of Horner Syndrome by Using Facial Images.

Area of Science:

• Ophthalmology
• Neurology
• Computer Science

Background:

• Horner syndrome, characterized by miosis, ptosis, and facial anhidrosis, signals potential damage to the oculosympathetic chain.
• Current diagnosis relies on subjective, operator-dependent clinical assessments.
• An objective, automated method is needed to improve diagnostic accuracy and reduce subjectivity.

Purpose of the Study:

• To present and verify an objective method for recognizing Horner syndrome from facial photographs.
• To develop and evaluate machine learning classifiers for automated Horner sign detection.
• To assess the accuracy, sensitivity, and specificity of the developed detection methods.

Main Methods:

• Collected and annotated 173 facial images, divided into training and testing sets.
• Developed a two-stage classifier using MediaPipe for landmark detection and geometric feature extraction, followed by machine learning classifiers.
• Trained a one-stage classifier using the YOLOv5 algorithm.

Main Results:

• The two-stage model with a Decision Tree Classifier achieved the highest accuracy (0.790), with sensitivity of 0.432 and specificity of 0.970.
• The one-stage YOLOv5 classifier demonstrated an accuracy of 0.65, sensitivity of 0.51, and specificity of 0.84.
• MediaPipe successfully detected facial landmarks in 92.2% of test images.

Conclusions:

• Automatic detection of Horner syndrome from facial images is feasible.
• The developed tool can serve as a secondary advisor for neurologists, enhancing diagnostic objectivity and accuracy.
• This approach has the potential to significantly aid in the early identification of serious underlying conditions indicated by Horner syndrome.