Horizontal nystagmus identification with joint SAM segmentation and time series classification

Nystagmus is an involuntary jerky eye movement. It is an oculomotor sign of asymmetry in the vestibular pathway. With the rapid advancement of deep learning and its widespread applications, numerous deep learning-based methods for nystagmus detection have been proposed based on eye movement videos. This would enhance diagnostic efficiency. We propose a horizontal nystagmus detection model with joint SAM segmentation and time series classification.

In this study, a convolutional neural network is employed to eliminate the invalid interference frames in the video. Subsequently, SAM is utilized to extract the motion trajectory of the pupil from the video, aiming to comprehensively investigate the motion characteristics of horizontal nystagmus and mitigate potential sources of interference. In addition, the horizontal nystagmus was determined by spatial attention and a multi-scale one-dimensional time series convolutional classifier.

Experiments were conducted on a clinically collected horizontal nystagmus video dataset. The experiments are mainly divided into two parts: pupil localization and nystagmus detection. In the experiments of pupil localization, the accuracy of the proposed method reaches 79.53%. For nystagmus detection, the precision of our method reaches 81%, which achieves significantly better performance than other approaches.

This paper presents an efficient diagnostic approach for the detection of horizontal nystagmus, offering substantial improvements in detection accuracy. Furthermore, it introduces a novel and clinically applicable solution for the early screening and intervention of vestibular disorders, enhancing the potential for timely diagnosis and management in clinical practice.