Ocular Torsion Assessment from Fundus Images and Video-oculography
Team: Team Hot Dog
- Program: Biomedical Engineering
- Course: Precision Care Medicine
Project Description:
Ocular torsion is the eye rotation about the line of sight. Pathologies in the vestibular-ocular-motor pathway, including time-sensitive medical conditions like brainstem stroke, often result in changes in the pattern of ocular torsion. The two types of ocular torsion are static torsion and dynamic torsion. Static torsion is clinically assessed by analyzing the disc-foveal angle (DFA) in the fundus images, and dynamic torsion is clinically evaluated through the ocular counter-roll test during in-person visits, or recorded by video-oculography (VOG). However, manually analyzing fundus images is time-intensive, while in-person ocular counter-roll tests create an access barrier in a telemedicine setting. Though automatic torsion assessment technologies exist, they are limited in speed due to the need for extensive data pre-processing. As a result, a major obstacle in the field is that clinicians fail to provide rapid and automatic ocular torsion assessment in both in-person and telemedicine settings due to the lack of computer-based eye movement analysis technologies, delaying the process of treating time-sensitive medical conditions. The objective of our project is to build deep learning models for automatic assessment of static and dynamic ocular torsion from fundus images and oculography videos.
Project Photo:
The logo features the torsional movement of the eye, which is what our project focuses on.
Project Poster
Project Post Summary:
We developed algorithms to automatically assess ocular torsion patterns based on both fundus images and oculography videos, using deep-learning based image classification, image segmentation, and video classification models. The feasibility of our project is supported by established deep-learning based models that can perform image segmentation on fundus images and video motion detection on oculography videos. With data synthesis, augmentation and model transferring, ocular torsion can be well detected from fundus images. The model performance can be further improved when more image data becomes available for training and fine-tuning. Label generation method for video classification is reliable, but more machine learning training and validation is needed in the future.
Project Mentors, Sponsors, and Partners
- Kemar Green
- David Zee
- Jorge Otero-Millan
- Amir Kheradmand
- Casey Overby Taylor
- Joseph L Greenstein