Noise-Resilient Pupil Segmentation in VNG Using RANSAC-Enhanced U-Net for Improved BPPV Diagnosis

As the number of patients experiencing dizziness has increased, particularly those affected by Benign Paroxysmal Positional Vertigo (BPPV), BPPV has gained attention. BPPV is a disorder where dislodged otoconia move from the utricle into the semicircular canals, leading to episodes of vertigo caused by the stimulation of the vestibular system. These movements result in short and repetitive nystagmus, which can be analyzed to diagnose the location of the vestibular dysfunction. Video-Nystagmo-Graphy (VNG) is a widely used diagnostic method for this purpose, but it relies heavily on expert interpretation and manual adjustments, presenting challenges in terms of accessibility and consistency. Recent research has focused on overcoming these limitations by utilizing deep learning techniques to automate the detection and analysis of nystagmus, thereby improving the accuracy and efficiency of diagnoses compared to traditional methods. However, previous studies have not adequately addressed the issue of noise in VNG recordings, a significant challenge due to the limited availability of VNG data and the nascent state of research in this area. In this paper, we aim to address this gap by evaluating algorithms under various noise conditions in VNG recordings. Specifically, we apply a RANSAC-based ellipse fitting algorithm to the U-Net model for pupil segmentation. This approach seeks to enhance the reliability of BPPV diagnosis by improving the accuracy of pupil segmentation in noisy VNG data, thus contributing to more consistent and reliable clinical outcomes.