Transformer-Based Detection and Clinical Evaluation System for Torsional Nystagmus

Area of Science:

• Ophthalmology
• Neurology
• Medical Imaging

Background:

• Benign paroxysmal positional vertigo (BPPV) diagnosis relies on precise assessment of torsional nystagmus.
• Conventional videonystagmography (VNG) struggles with accurate quantification of rotational eye movements.
• Existing automated methods show limited efficacy in detecting torsional nystagmus.

Purpose of the Study:

• To develop an automated system for accurate and quantitative detection of torsional nystagmus.
• To introduce the Torsion Transformer model for direct torsion angle estimation from iris images.

Main Methods:

• A self-supervised learning framework with Decoder and Finder modules to estimate torsion angles.
• Time-series torsion angle data analyzed by a 1D convolutional neural network (1D-CNN) classifier.
• Evaluation using video recordings from 127 BPPV patients.

Main Results:

• The Torsion Transformer achieved 89.99% sensitivity, 86.36% specificity, 88.82% F1-score, and 87.93% AUROC.
• Performance is comparable to established methods for horizontal and vertical nystagmus detection.
• The model effectively quantifies torsional nystagmus.

Conclusions:

• The Torsion Transformer demonstrates significant technical improvement in torsional nystagmus detection.
• This system shows promise as a clinical decision support tool for BPPV diagnosis.
• Enhances diagnostic accuracy for healthcare professionals.