VAULT

Vault Accuracy Using deep Learning Technology

An AI-driven tool to accurately predict postoperative vault of the EVO Visian ICL™ (Implanted Collamer® Lens).

PATENT PENDING

Team

Taj Nasser, MD

Greg Parkhurst, MD

Matt Hirabayashi, MD

Gurpal Virdi, MD

Objectives

To create an accurate, repeatable, and continuously improving machine-learning based tool for the prediction of post-operative ICL™ Vault using various imaging modalities (e.g., Ultrasound Biomicroscopy and Anterior Segment OCT). This approach is novel and image-based, training the model on the unique anatomy of each eye. The models are trained with data from refractive surgery cases in the United States at a high volume surgery center.

Methodology

3,059 images from 437 eyes of 221 consecutive patients who underwent ICL implantation were included. A neural network was trained on pre-operative Very High Frequency (VHF) digital ultrasound images, patient demographics, and postoperative vault. A neural network was chosen and extensively adapted for the purposes of this project.

Results – UBM

12.1 MAE: 66.3 µm

12.6 MAE: 103 µm

13.2 MAE: 66.3 µm

VAULT Prediction Error Stratified by Size and Magnitude of Prediction Error
ICL Size	Percent of Predictions within Error Range
	≤ 250 µm	≤ 400 µm	≤ 500 µm
12.1	99.7%	99.8%	100%
12.6	97.4%	99.0%	99.0%
13.2	92.4%	95.1%	96.6%

Comparison to Current Literature
Study	Error	Location	Sample Size
Rocamora et al.	132.0 μm (MAE)	Argentina	115 Eyes 59 Patients
Kim et al.	104.7 μm (MAE)	Korea	892 Eyes 471 Patients
Kang et al.	106.88/143.69 μm (MAE)	Korea	2756 Eyes
Shen et al.	159.03 μm (RMSE)	China	6297 Eyes 3536 Patients
Chen et al.	129.89 μm (MAE)	China	1941 Eyes 1941 Patients
Russo et al.	96.94 µm (MAE)	Italy	561 Eyes 300 Patients
VAULT	12.1: 66.3 µm (MAE) 12.6: 103 μm (MAE) 13.2: 91.8 µm (MAE)	USA	3059 Images 437 Eyes 221 Patients

Read the JCRS Publication for VAULT: UBM

Next Steps

We are actively training the model on anterior segment OCTs. View the results below, pending publication.

Results – OCT

VAULT Prediction Error Stratified by Size and Magnitude of Prediction Error
ICL Size	Percent of Predictions within Error Range
	≤ 250 µm	≤ 400 µm	≤ 500 µm
12.1	99.7%	99.8%	100%
12.6	97.4%	99.0%	99.0%
13.2	92.4%	95.1%	96.6%

Vision

An user friendly interface will allow surgeons to upload their images and patient data for the model to process and return a clean output with predicted vault by ICL size. Eventually, the model can be integrated into UBM or OCT machines.

User Interface

(input page)

(results page)

(printout)