Performance of two ultra-widefield retinal imaging systems for the automatic diagnosis of diabetic retinopathy

Abstract

Purpose : Ultra-widefield retinography (UWF-R) examination is a technological evolution of conventional retinography examination by allowing extensive retina views and providing more information about signs of pathology in the periphery. Different UWF-R devices exist which produce visually different images. This study investigates the performance of two different UWF-R devices, the Clarus 500 (Carl Zeiss Meditec Inc., Dublin, CA, USA) and 200Dx (Optos, Dunfermline, UK), for the task of automatic Diabetic retinopathy (DR) severity assessment, using deep learning.

Methods : A total of 731 eyes of 324 diabetic patients were imaged using Optos, and 1131 eyes of 509 diabetic patients using Clarus. The severity of DR on each image was determined by an expert using the ICDR scale. Each image was classified into one of the following categories: no DR, mild nonproliferative DR (NPDR), moderate NPDR, severe NPDR, proliferative DR (PDR), and panretinal photocoagulation (PRP). This data set was divided into 499/119/113 eyes for train/validation/test for Optos and 826/192/113 for Clarus. The test set was the same for both devices (same patients). In a second experiment, a subset of Clarus images, Clarus--, was used to match the number of training and validation eyes in each class as Optos. The retina was automatically segmented and cropped in Optos images. Different models were trained for each dataset and evaluated using the area under the receiver operating characteristic curve (AUC).

Results : We considered four binary classification tasks, using four severity cutoffs, and selected the best model for each task based on the validation AUC. Clarus performed better in the test set for all tasks. The best AUCs were 0.8359, 0.8332, 0.8218, and 0.9666 for task0 (mild NPDR or more), task1 (moderate NPDR or more), task2 (severe NPDR or more), task3 (PDR or PRP), respectively. These AUC values suggest that the models are discriminating well.

Conclusions : The proposed AI can automatically grade DR in UWF-R images with good precision. The performance of the trained models on the Clarus dataset was promising. When tested on the Clarus2 dataset, a smaller dataset defined for a fair comparison with Optos, the performance degrades but remains better than Optos. This highlights the influence of the amount of data and the acquisition device on the automatic DR grading.

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

 

Tab 1. AUCs results

 

Fig 1. proposed method