A hybrid deep learning framework for keratoconus detection based on anterior and posterior corneal maps.

Abstract

Purpose : To develop and assess the accuracy of a deep learning framework for keratoconus detection based on eccentricity and elevation maps

Methods : We collected 1,196 corneal images from 299 eyes of 168 patients who underwent tomography examination (Pentacam^tm, Oculus) at the Cornea Division of EPM-UNIFESP, Brazil. We developed a deep learning framework to investigate anterior eccentricity (AEC), posterior eccentricity (PEC), anterior elevation (AEL), and posterior elevation (PEL) maps for detecting keratoconus. We first generated deep features using the RestNet-50 deep learning architecture, then employed a Support Vector machine (SVM) classifier to diagnose keratoconus. We used t-Distributed Stochastic Neighbor Embedding (t-SNE) to visualize deep features of all four maps to assess the quality of learning subjectively. We employed 5-fold cross validation of classification to assess the quality of learning objectively.

Results : A total of 148 eyes were normal and 151 eyes had keratoconus. The clinical diagnoses of the cases were generated by three specialists. We extracted 1,000 deep features from the ResNet50. Qualitative examination of the t-SNE visualizations indicates a significant separation of deep features between normal and KCN eyes (Fig.1). The accuracy of the hybrid deep learning model in detecting keratoconus from the AEC, PEC, AEL and PEL maps were 98.3%, 97.3%, 95%, and 97.3%, respectively. The area under the ROC curve was equal to 0.99 for the anterior eccentricity map (Fig.2). The time for running the entire framework was equal to only 1.5 minutes.

Conclusions : We developed a deep learning model to extract deep features from corneal images and a conventional machine learning classifier to learn deep features for keratoconus detection. The model achieved high accuracy in identifying keratoconus based on corneal eccentricity and elevation maps, highlighting the critical role of these maps in clinical settings. The proposed framework was also time efficient with low computation complexity.

This is a 2021 Imaging in the Eye Conference abstract.

 

Visualization of 1000 deep features extracted by ResNet50 deep learning network using t-SNE. A) Anterior eccentricity, B) Posterior eccentricity, C) Anterior elevation, and D) Posterior elevation.

 

ROC curve of the model in detecting keratoconus from anterior eccentricity map.