Device antagonist hybrid fusion model for suspect keratoconus detection

Abstract

Purpose : To develop and evaluate a hybrid fusion deep learning model for suspect keratoconus detection based on pachymetry corneal map

Methods : At Jichi Medical University in Japan, we collected 5317 corneal maps using CASIA optical coherence tomography (OCT)-based device (SS-1000, Tomey, Japan), while in Egypt, we gathered 1000 corneal maps using a Pentacam Scheimpflug imaging apparatus (Oculus Optikgerate GmbH). With the 5317 maps from the CASIA dataset, we created a deep learning fusion model consisting of two architectures, i.e. Xception and InceptionResnetv2, to identify suspect keratoconus. Features were extracted from Casia dataset using the two deep learning models, then they were used to train to Support Vector machine (SVM) classifier. The independent test set included only the maps from Pentacam device.

Results : The CASIA dataset contained 1541 map suspect keratoconus (SUSP) eyes and 3776 normal (NOR) eyes (annotated based on the CASIA built-in Ectasia Screening Index; ESI). The independent Pentacam dataset included 500 eyes that were normal and 500 eyes that were suspect keratoconus (annotated based on clinical evaluations through two specialists). By using the Pentacam dataset as a test set, the suggested deep learning model's accuracy and Area under the Receiver operator Characteristics Curve (AUC) were 80.2% and 0.845, respectively, for diagnosing suspect keratoconus from pachymetry map. The confusion matrix of using the fusion of deep learning features and SVM classifier is illustrated in Fig.1.

Conclusions : We created a hybrid fusion deep learning framework to identify suspect keratoconus based on CASIA dataset from Japan, and we tested it using independent corneal maps from Pentacam device from Egypt. The proposed fusion method illustrates the generalizability of the created deep learning model. The suggested model has a respectable level of accuracy and AUC in identifying suspect keratoconus using two separate devices from different ethnicity. This approach could improve the ophthalmological care for suspect keratoconus.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.

 

Fifure 1. The confusion matrix of using the fusion of Xception and InceptionResnetv2 deep learning features and SVM classifier