ESCRS - FPT08.01 - Deep Learning Detection And Quantification Of Lens Density Using Ss Oct

Deep Learning Detection And Quantification Of Lens Density Using Ss Oct

Published 2022 - 40th Congress of the ESCRS

Reference: FPT08.01 | Type: Free paper | DOI: 10.82333/g9dm-m618

Authors: Christophe Panthier* 1 , Pierre Zeboulon 1 , Damien Gatinel 1

1ophthalmology,Fondation Rothschild,Paris,France

Purpose

To automatically detect the lens edge on the SS OCT scans provided by the Anterion and determine after a training session, if a pixel was a “cataract pixel” or a “normal pixel”.

To perform a reliable algorithm for detection and quantification in daily pratice.

Setting

monocentric prospective study at the Rothschild Foundation. Utilization of neuronal network for deep learning imaging treatment.

Methods

The fraction cataract (FC) is the number of cataract pixel divided by the total pixel number of the lens. The ocular scatter index (OSI) was measured with a double-pass aberrometer (OQAS®) and compared to the FC.

Results

2622 SS-OCT scans were included (437 eyes). 687 SS-OCT scans were used to perform the automatic detection. 585 images (241 with cataract and 344 normal) were used in the training set. 1350 images of 225 eyes (95 cataract and 130 normal) were included in the validation set.

A FC greater than 0.67 was the cut-off threshold for cataract, with 94.6% sensitivity and 95.4% specificity. The FC area under the ROC curve (AUROC) was 0.979.

Conclusions

FC measured by our automatized algorithm based on SS-OCT is a repeatable and reliable objective cataract grading method. In cases of FC greater than 0.67, it is reasonable to discuss cataract surgery in symptomatic patients.