Cross-Device AI Fusion: Enhancing Diabetic Retinopathy Diagnosis with Combined Clarus and Optos Images

Abstract

Purpose : Ultra-widefield Retinography (UWF-R) is increasingly pivotal in ophthalmic diagnostics, distinguished by its broad field of vision and advanced imaging methodologies. This research acknowledges the distinct acquisition technologies employed by the Clarus 500 (Carl Zeiss Meditec Inc., Dublin, CA, USA) and the 200Dx (Optos, Dunfermline, UK) UWF-R systems, suggesting the possibility of complementary benefits. The study investigates whether a deep learning (DL) model trained on the fusion of UWF-R images from these technologically divergent systems can enhance automatic Diabetic Retinopathy (DR) severity assessment compared to using images from a single system individually.

Methods : A DL model was trained on a dataset that included images from Clarus and Optos devices for each eye, with 411/105/113 images for the training/validation/test sets. This data was collected in the frame of the Evired project. The severity of DR on each image was determined by an expert using the ICDR scale. The proposed model has two streams, initialized with ResNet50, to extract features. A CrossAttention module was integrated to refine feature extraction by emphasizing areas with the most diagnostic significance from each image. We have also trained a ResNet50 for each device separately to compare results.

Results : We considered four binary classification tasks, using four severity cutoffs, and selected the best model for each task based on the validation area under the receiver operating characteristic curve (AUC). The fusion model utilizing combined Clarus and Optos images demonstrated enhanced diagnostic capabilities. For the classification tasks of mild nonproliferative DR (NPDR) or more and severe NPDR or more, the combined model outperforms the Clarus- and Optos-only models. Notably, for detecting proliferative DR (PDR) or panretinal photocoagulation (PRP), the combined model provides a statistically significant improvement from Clarus and Optos. These results underscore the benefit of image fusion in identifying advanced DR stages.

Conclusions : The proposed AI can automatically grade DR in UWF-R images with good precision. The cross-device fusion model capitalized on the strengths of both Clarus and Optos imaging systems, resulting in a more robust and precise evaluation of DR severity, and can be used as an annotation assistance system in a multimodal reading center.

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

 

Fig 1. Proposed method

 

Tab 1. AUCs results