Assessing the Efficacy of Deep Learning and Machine Learning Features in Detecting Keratoconus from Corneal Maps

Abstract

Purpose : To develop and validate a hybrid model integrating deep learning and classical machine learning features for detecting keratoconus (KC) from corneal maps collected from dual Scheimpflug analyzer

Methods : A total of 363 corneal maps were collected from 121 eyes of 61 patients using GALILEI G4 instrument (Ziemer Ophthalmic Systems AG) in Sousas Eye Clinics , São Paulo, Brazil. We used corneal thickness (CT), anterior elevation Best-Fit Toric Aspheric (antBFTA) and posterior elevation Best-Fit Toric Aspheric (postBFTA) maps for the downstream analysis. We first preprocessed the maps and removed non-image objects and extracted features based on InceptionV3 deep learning architecture. We then fed features to a Support Vector Machine (SVM) classifier and tested the performance based on 5-fold cross validation (Fig.1). We then employed t-distributed stochastic neighbor embedding (t-SNE) to evaluate the quality of the extracted features subjectively (Fig. 2). We evaluated the performance of the model based on accuracy and area under the receiver operating characteristic curve (AUC) metrics

Results : The dataset included 49 normal eyes and 72 eyes with KC that were previously diagnosed by two corneal specialists. The accuracy of the proposed model in detecting KC from CT, antBFTA and postBFTA maps was 90%, 95% and 96.7%, respectively, and the AUC was 0.96, 0.98 and 0.99, respectively.

Conclusions : We developed a hybrid machine learning model to detect KC from CT, antBFTA and postBFTA maps based on the InceptionV3 architecture. Findings suggest that machine learning models are promising in detecting KC, as well as underscore how antBFTA and postBFTA maps are crucial for this purpose. While the model was developed based on relatively a small number of samples, it was accurate and generalizable. This approach is more appropriate for the scenarios with small subsets and may improve KC research and clinical practice.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. Block diagram of the proposed framework based on DL features and SVM applied to the corneal thickness, anterior and posterior elevation Best-Fit Toric Aspheric(BFTA) maps

 

Figure 2. Visualization of deep features extracted by Inception 3 deep learning mdoel using t-SNE. A) CT, B) antBFTA, and C) postBFTA.