Evoked Potential-Evidenced Visual Impairment Categorization using Vision Mamba on ERG Correspondence Features

Visual impairment categorization plays a critical role in determining support levels for individuals with low vision. Traditional assessment methods, which rely on visual acuity and field measurements, often suffer from subjectivity and variability. To address this, we explore the use of electroretinography (ERG) as an objective alternative. However, challenges such as signal noise, complex signal patterns, and limited dataset availability hinder the accuracy of measurements. In this work, we propose a robust deep learning framework for visual impairment categorization. To address the signal noise, we introduce the ERG correspondence feature, capturing interocular consistency to mitigate unilateral noise effects. To efficiently learn discriminative features, we apply Continuous Wavelet Transform (CWT) with characterized mother wavelets to convert 1D ERG signals into 2D time-frequency representations, which are then processed by a Vision Mamba model. Furthermore, a new dataset containing various low-vision disease cases is collected and a challenging categorization task is defined. Experimental results demonstrate that our method achieves better performance compared to existing ERG-based classification approaches, offering a promising direction for objective and reliable visual impairment assessment.