| Title |
Preprocessing Optimization for a Horizontal Nystagmus Detection Model for BPPV |
| Authors |
강태훈(Taehun Kang) ; 조윤상(Yoonsang Cho) ; 한주혁(Ju-Hyuck Han) ; 조용석(Yongseok Cho) |
| Keywords |
Nystagmus; Benign paroxysmal positional vertigo; Deep learning; ResU-net; DenseNet |
| Abstract |
This study proposes a horizontal auto detection model for the diagnosis of benign sudden head sphincter (BPPV).This study proposes an automatic horizontal nystagmus detection model for the diagnosis of Benign Paroxysmal Positional Vertigo (BPPV). While conventional preprocessing methods based on U-Net achieve high accuracy, they are limited by long processing times, making them less suitable for use in emergency diagnostic situations. In this study, the preprocessing stage was removed, and grayscale eye images were directly used as input, significantly improving time efficiency. The proposed ResU-Net and DenseNet models demonstrated comparable classification performance to existing CNN-based models. In particular, ResU-Net achieved an accuracy of 79.9% and an F1-score of 74.8%, despite the absence of a preprocessing step. Furthermore, it achieved approximately 2.84 times faster processing speed compared to the preprocessing-based approach, demonstrating clinical practicality beyond mere performance metrics. These results confirm that ResU-Net can serve as a practical alternative that simultaneously satisfies both speed and accuracy requirements in emergency diagnostic scenarios. |