Abstract
Purpose :
Microaneurysms (MA) are early clinical signs seen in patients with non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). They result from capillary wall outpouching, can bleed, and result in vision loss. The advent of optical coherence tomography angiography (OCTA) has allowed for the noninvasive detection of MAs as compared to fluorescein angiography. Herein, we developed an nnU-Net that automatically identifies MAs on OCTA en face images.
Methods :
Patients with early, intermediate, and severe NPDR as well as PDR, who were imaged on the ZEISS PLEX Elite 9000 using a field size of 6x6 mm, were retrospectively enrolled from the New England Eye Center. We isolated the automatically segmented superficial and deep capillary layers of the OCTA en face images. Two expert graders independently and manually labeled MA using the EXACT web-based labeling tool. Utilizing the labeled data, we trained a 2D nnU-Net (a self-configuring deep learning tool that utilizes heuristic and data-based rules to choose suitable hyper-parameters) to detect MA. Both the network topology and training process are guided by empiric rules that take into account image dimension, modality, and annotations. Furthermore, we adapted parameters of the nnU-Net to fit the specific needs of MA detection. We evaluated different loss functions and training parameters.
Results :
111 eyes from 60 patients with DR were included for analysis. Preliminary tests on the network utilized the superficial capillary layer labeled by one expert grader. 16 images were held back for testing and the network achieved a per-pixel accuracy score of 99% on the test data. These results are preliminary and do not include all of the data and annotations yet.
Conclusions :
Detection of MA in the superficial capillary plexus images from the PLEX Elite system is feasible using annotations generated from both the EXACT labeling tool and nnU-Net and it achieves a per-pixel accuracy score of 99%.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.