Using 3D Deep Learning for Classification of Multiple Retinal Diseases on Optical Coherence Tomography Images

Abstract

Purpose : Retinal diseases are assessed by trained medical personnel and triaged according to clinical urgency in ophthalmology clinics. This conventional approach is time-consuming and requires significant human resources and expertise. Here, we developed and tested a 3D deep-learning-based model to detect multiple retinal diseases on optical coherence tomography (OCT) images and compared the efficiency between the two architectures of the 3D model.

Methods : This was a retrospective analysis of Spectralis OCT images (Heidelberg Engineering) obtained from the Chinese University of Hong Kong Eye Centre and Hong Kong Eye Hospital. The inclusion of diagnosis and labeling of disease and the triage level was shown in Figure 1. We used 3D Residual Network (ResNet)-101 to develop a deep-learning model for the classification of multiple retinal diseases. We designed two architectures to compare the efficiency of the 3D model: a simple structure of a multi-outcome network (Architecture 1) and an ensemble of multiple binary-classification networks (Architecture 2) as shown in Figure 2. The model yielded the triage outcome based on the returned probability of the most urgent diseases at a given threshold. The model performance was tested by an external set with 1,010 volumetric scans collected from Prince of Wales Hospital, Hong Kong.

Results : 6,813 volumetric scans from 1,669 subjects (mean age [SD]= 68.5 [17.5], 44% Female) were included in training (70%), fine-tuning (10%), and internal validation (20%) sets. In the internal validation, Architecture 1 achieved area under curves (AUCs) above 0.941 in all diseases and normal, except drusen and ERM with AUCs of 0.731 and 0.845. Compared with Architecture 2, the discriminative performance was similar (all p-values >0.05 in DeLong's test). Next, we found Architecture 1 was tenfold faster than Architecture 2 (0.67s VS 6.13s) to output the prediction based on the same computational consumption. Furthermore, in the external testing, Architecture 1 had AUCs above 0.930 in all diseases and normal, except drusen and ERM.

Conclusions : The discriminative performance of the deep-learning model was good, which could be an add-on tool to facilitate the current triage workflow for retinal diseases potentially.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1. The Workflow of Image Collection and Image Label.

 

Figure 2. The 3D deep learning model for multiple retinal disease classification.