Most contributing quantified ocular parameters for classification of glaucomatous optic disc shape

Abstract

Purpose : The purpose of this study was to find most contributing ocular parameters to classify the glaucomatous optic disc based on Nicolela’s classification model.

Methods : This study included 163 eyes of 105 open-angle glaucoma (OAG) patients (age: 62.3 ± 12.6, mean deviation: -8.9 ± 7.5 dB). All the eyes were classified into 4 disc types according to Nicolela’s classification (focally ischemic discs, myopic glaucomatous discs, generalized enlargement of the discs, senile sclerotic (SS) discs), by three ophthalmologists. Ninety-one quantified ocular parameters were acquired for each eye; 7 background data, 48 parameters from optical coherence tomography (OCT) related to cpRNFLT and optic disc topography, and 36 ocular circulation parameters from laser speckle flowgraphy. With randomly selected data (n=114), we trained a neural network (NN) which has just one hidden layer. Then we evaluated the classification model with the remaining data (n=49). During the training phase, we used minimum Redundancy Maximum Relevance (mRMR) and genetic algorithm based feature selection, to find the most contributing quantified parameters. For comparison, brute force feature selection after mRMR was applied, to check the validity of the selected parameters.

Results : The most contributing ocular parameters selected in our approach were the same as brute force feature selection selected, totally 9, including spherical equivalent, age, average rim disc ratio (nasal), average cup depth, horizontal disc angle, 6-sector superior-temporal cpRNFLT, superior-quadrant cpRNFLT, maximum cup depth and cup area. With these parameters, the validated accuracy with test data for NN was 87.8%.

Conclusions : With this novel approach, glaucomatous optic discs were classified objectively with high accuracy. Furthermore, the calculated confidence of predictions for each disc type can assist the glaucoma management.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.