Polyp Detection from Structural and Angiographic OCT Using Deep Learning

Abstract

Purpose : We aim to automatically detect polyps characteristic of polypoidal choroidal vasculopathy (PCV) from structural and angiographic OCT using a deep learning network.

Methods : This study included 66 eyes diagnosed with PCV using a combination of indocyanine green angiography (ICGA) and structural OCT. From these eyes, 13 were used for validation, 13 for testing, and the rest for training. Thirteen healthy controls were also included for performance evaluation. Volunteers were scanned with a commercial instrument using either a 3×3-mm or 6×6-mm scan pattern at the location of polyps as revealed by ICGA. No scans were excluded due to image quality. We used a faster region-based convolutional neural network (faster RCNN) architecture, which is a region proposal network that detects features within learned regions (Fig. 1). To aid in feature identification, we restricted region proposals to locations between the inner limiting and Bruch’s membranes. Ground truth region proposals were determined by cross-referencing ICGA images with the structural OCT scans. Inputs to the network consisted of sets of 10 adjacent structural OCT and the corresponding OCT angiography (OCTA) B-frames. The network outputs regions of interest and the probability that each contains a polyp (Fig. 2).

Results : We evaluated performance by determining the area under receiver operating characteristic curve (AROC = 0.938) (Fig. 2) for polyp detection by comparison with the healthy controls. The network achieves a detection sensitivity of 73.1% at 95% specificity.

Conclusions : A deep learning approach can reliably detect polyps from structural OCT and OCTA data. This approach may help enable improved monitoring and screening for PCV by removing the need for invasive and expensive ICGA procedures.

This is a 2021 Imaging in the Eye Conference abstract.