Deep Learning Model for Static Ocular Torsion Detection Using Synthetically Generated Fundus Images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Static ocular torsion assessment aids in classifying vertical ocular misalignment.
• Current assessment methods are time-consuming and require specialized expertise.
• Developing efficient diagnostic tools is crucial for frontline providers.

Purpose of the Study:

• To develop deep learning models for assessing static ocular torsion using synthetic fundus images.
• To classify both the direction (intorsion vs. extorsion) and amount (physiologic vs. pathologic) of ocular torsion.
• To provide a more accessible and time-efficient diagnostic method.

Main Methods:

• Utilized a dataset of 276 right eye fundus images to generate synthetic images via rotation.
• Developed a binary classifier (intorsion vs. extorsion) and a multiclass classifier (physiologic vs. pathologic).
• Employed transfer learning with large synthetic datasets (12,740 images per model) and evaluated on unseen data.

Main Results:

• The binary classifier achieved 0.92 accuracy and 0.98 AUROC on synthetic data.
• The multiclass classifier achieved 0.77 accuracy and 0.94 AUROC on synthetic data.
• The binary classifier demonstrated strong generalization to real-world data with 0.94 accuracy and 1.00 AUROC.

Conclusions:

• Deep learning models can accurately detect the direction of static ocular torsion from synthetic fundus images.
• This technology is vital for differentiating between vestibular and ocular muscle misalignment.
• The strategy shows promise for AI research in rare neuro-ophthalmologic diseases with limited data.