Deep Learning Assisted Quantification of Neovascular Lesions with Optical Coherence Tomography Angiography (OCT-A)

Abstract

Purpose : OCT-A is a relatively new imaging modality allowing noninvasive visualization of macular neovascularization (MNV) in neovascular age-related macular degeneration (nAMD). Grader assessment of MNV area includes annotations on the angiogram using properly segmented OCT-A scans within defined slabs. MNV area is delineated after a comprehensive examination of the cube scan and visualization of flow in areas of retinal pigment epithelial detachment with or without a visible network on the angiogram. To reduce burden on human resources and improve efficiency, we developed an automated, deep learning (DL) algorithm for quantitative assessment of MNV area from OCT-A.

Methods : Training dataset included planimetry annotations of MNV area by multiple graders from 6x6mm OCT-A scans of 44 eyes generating a total of 104 lesion annotations. Screenshots of the annotated angiograms were used for training a U-Net DL model with a ResNet backbone for binary image segmentation. Predictions made by the algorithm were fitted with a binary mask to define borders and area measurements. As the images were of a fixed sized, pixel counts of mask sizes were used to measure predicted masked area. Validation dataset included 28 OCT-A for comparison of grader annotation with DL predictions using area measurements and Dice Similarity Coefficient (DSC).

Results : Mean difference in MNV area between graders was 0mm² (95% CI -1.12 to 1.12). Mean difference in MNV area between the DL algorithm and grader was 0mm² (95% CI -4.07 to 4.07) with an intra-class correlation of 0.731. Mean DSC between DL algorithm and grader was 0.45 +/- 0.26. Examination of OCT-A annotations with DSC < 0.6 (16 images) showed quilting artifacts and absence of MNV network on the angiogram.

Conclusions : There is a moderate correlation between DL quantification of MNV lesions and graders’ area measurements using OCTA. It is important to note that the algorithm’s decision is limited by the angiogram whereas the grader has access to structural b-scans and flow data. This study shows a very promising trajectory for the use of DL in the quantitative analysis of MNV lesions.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Original grader annotation (left), followed by a binarized mask (center), and a DL prediction fitted with a binary mask (right).