3-D analysis of multiple OCTA acquisitions for the automatic diagnosis of diabetic retinopathy

Abstract

Purpose : Blindness can result from diabetic retinopathy (DR). Optical Coherence Tomography Angiography (OCTA) provides a detailed view of the retinal and choroid vascularization at the microvascular level. This study investigates the performance of two different specifications of PLEX® Elite 9000 (Carl Zeiss Meditec Inc., Dublin, CA, USA) OCTA acquisitions (6x6mm with 500x224x500 voxels and 15x15mm with 834x224x834 voxels) in the automatic classification of DR.

Methods : 801 eyes from 432 diabetic patients were imaged for the study. This collection was divided as follows: 567 eyes (from 306 patients) for training, 137 eyes (from 73 patients) for validation, and 97 eyes (from 53 patients) for testing. DR severity in each eye was graded by a retina specialist. Resnet50 3-D classifiers were trained to analyze structure and flow volumes within 224x224x224x2 voxel patches. For training, random crop is used, and for prediction, multiple random crops are used, and the most pathological patch is used as the prediction result. Separate models were trained for the two different specifications of OCTA and a third model was trained to analyze them jointly with a two-branch intermediate fusion ResNet50. Finally, the predictions of these models were combined through max or average late fusion, and classification performance was assessed using the Area under the ROC Curve (AUC).

Results : The 6x6mm OCTA model performed better at detecting whether DR was present (Task0). In contrast, the 15x15mm OCTA model performed better at predicting whether the severity level was at least moderate non-proliferative DR (Task1), whether it was at least severe NPDR (Task2) and whether it was proliferative or treated DR (Task3). Intermediate fusion led to limited performance improvement. However, late fusion always led to improvement compared to the analysis of a single acquisition: AUC = 0.912 for task 0, as opposed to 0.832 using 6x6mm alone; AUC = 0.829 for task 1, as opposed to 0.769 using 15x15mm alone; AUC = 0.812 for task 2, as opposed to 0.772 using 15x15mm alone; AUC = 0.900 for task 3, as opposed to AUC = 0.845 using 15x15mm alone.

Conclusions : The joint analysis of 6x6mm and 15x15mm OCTA acquisitions improves DR classification performance. Late fusion, which clearly outperforms detection using a single acquisition, is particularly promising.

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

 

Table1: Severity distribution of eyes

 

Table2: Different fusion methods' AUC results