Abstract

Purpose: Cortical/cerebral visual impairment (CVI) is a leading cause of pediatric visual impairment and is characterized by abnormalities of higher-order visual processing and visual attention. However, children with CVI have limited ability to cooperate with standard methods for assessing these abnormalities. In this study, we combined machine learning (saliency modeling) with a passive (ie, no active task requirements) eye tracking protocol to evaluate visual attention to low-and high-level features in children with CVI.
Methods: 42 children with CVI and 29 age-matched controls (≤ 12 years old) were prospectively recruited. Eyes were tracked while participants viewed a series of images on a monitor. Saliency maps were generated using SegCLIP, a deep neural network for image segmentation based on large-scale language pre-training (CLIP). SegCLIP allows prompting any feature of interest, resulting in a corresponding saliency map (Fig 1). Both low-level (eg, texture) and high-level (eg, faces) maps were generated. For each participant, fixation saliency values were calculated for each trial per each map (Fig 2). Data from children with CVI and controls were compared using Mann-Whitney tests.
Results: CVI patients produced atypical fixation patterns compared to controls. Among others, CVI patients fixated less to human faces (p= 0.0004) and textures (p= 0.001). Instead, they preferred fixating on backgrounds (p= 0.001). Children with CVI also produced significantly shorter fixations to regions of increased complexity (p= 0.007).
Conclusions: Children with CVI exhibit distinct fixation patterns, focusing on backgrounds rather than features …

Date

June 17, 2024

Authors

Melinda Chang, Kleanthis Avramidis, Rahul Sharma, Mark Borchert, Shrikanth Narayanan

Journal

Investigative Ophthalmology & Visual Science

Volume

65

Issue

7

Pages

1501-1501

Publisher

The Association for Research in Vision and Ophthalmology