The challenge
The task was to build an application that could accurately assess eyelid position and identify abnormalities related to proptosis, a condition in which the eye is displaced forward.
To do this, the application had to calculate key ophthalmological metrics: Margin Reflex Distance 1 and 2 (MRD1 and MRD2), the distances from the centre of the pupil to the upper and lower eyelid margins; eyebrow height, the distance from the centre of the pupil to the eyebrow; proptosis distance, from the centre of the pupil to the lateral orbital bone; and eye movement distances in various directions. Accurate measurements are essential for identifying and monitoring proptosis.
Our solution
We used deep learning to build Eye Metrics, a cross-platform web application that runs on iOS and Android mobile devices as well as desktop and laptop browsers. Users, primarily patients, capture facial images and the application calculates the relevant ophthalmological metrics. The predictive models continue to improve as new training samples are added.
We then extended the application with functions for eye surgical assessment, aimed at giving healthcare professionals a more comprehensive and efficient analysis workflow.
- Calculation of MRD1, MRD2, eyebrow height, proptosis distance and eye movement distances from facial images
- Cross-platform access on iOS, Android and desktop browsers
- Continuous model improvement with new training samples
- Batch upload of before-and-after surgery images, with automated comparison of LAB colour values of the sclera
- Live video recording to estimate the change in sclera LAB values over a set interval, typically 1 to 5 minutes
The outcome
Eye Metrics gives patients and healthcare professionals a practical tool for ophthalmological assessment. Its model outperforms Google's face mesh detection, helping clinicians identify proptosis and allowing patients to take an active role in monitoring their eye health.
The automated before-and-after comparison saves doctors time in surgical assessment by replacing manual review of image data.
