Spatial vascular connectivity network for deep learning construction of microcapillary angiography from single-scan-volumetric OCT

Abstract

Purpose : The purpose of this study to verify the feasibility of a deep learning approach for microcapillary angiography from single-scan-volumetric OCT using the principle of spatial vascular connectivity (SVC), i.e., brightness connectivity along the vessel direction.

Methods : In this study, we present ‘SVC-Net’, an encoder-decoder network. The input into SVC-Net is a three-channel image, consisting of three neighboring OCT B-scans from a single-scan-volumetric OCT, termed 3N. The output of SVC-Net is a single channel image, i.e., an OCTA B-scan. Figure 1 illustrates the pipeline of SVC-Net. To validate the effects of SVC, we also trained SVC-Net using a single channel OCT image, termed 1N. The ground truth corresponds to speckle variance OCTA B-scans determined from four repeated OCT B-scans. To prevent overfitting, data augmentation in the form of rotations, flips and zooming were implemented. Evaluation metrics for SVC-Net using 1N and 3N inputs included multi-scale-structural-similarity-index-measure (MS-SSIM) and peak-signal-to-noise-ratio (PSNR).

Results : For evaluation, we generate the enface projections from the superficial vascular plexus (SVP) and the deep capillary plexus (DCP). Representative examples of the SVC-Net 1N and 3N models are shown in Fig. 2. Qualitative observations show that the 3N model has better contrast compared to the 1N model. Quantitative comparison, the MS-SSIM and PSNR for the 1N SVP are 0.706 and 15.31, respectively, and for DVC are 0.583 and 12.24. Whereas the MS-SSIM and PSNR for 3N SVP are 0.760 and 16.73, respectively, and for DVC are 0.669 and 14.25, respectively. Quantitatively, the 3N model has improved performance over the 1N model.

Conclusions : In this study, a deep learning network utilizing SVC was developed for microvascular angiography construction from single-scan-volumetric OCT. This deep learning approach can improve clinical implementation of OCTA by reducing acquisition time, i.e., alleviate the requirement of multiple repetitions, thereby can alternatively improve transverse image resolution or field-of-view.

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

 

Figure 1. Schematic illustration of the deep learning pipeline for SVC based OCTA construction.

 

Figure 2. Representative en face OCTA images of ground truth and deep learning predictions.