Abstract
Purpose :
Averaging is known to improve vessel continuity in ocular Optical Coherence Tomography Angiography (OCTA) en face images but is time consuming, since it involves repeatedly scanning the patient. Alternatively, improving continuity in individual images is possible by using a Hessian “vesselness” filter. Although this can improve continuity, it can compromise fidelity to the anatomy. This work investigates the use of a convolutional neural network (CNN) on individual scans to improve continuity without this loss of accuracy.
Methods :
A CNN having a modified 5-layer architecture with deep supervision was trained on OCTA images of the superficial capillary plexus, using averages of repeated scans of the same subject as ground truth. Approximately 30,000 patches of individual images were trained against corresponding patches of the averaged image. A PLEX® Elite 9000 SS-OCT System (ZEISS, Dublin, CA) was used to acquire 4 repeated 6x6 mm OCTA scans each of 12 subjects, 6 normals and 6 with vascular pathology. Superficial angiography slabs from these repetitions were averaged, providing 12 high quality images, using one of the repetitions as the reference for each average. The CNN was used to process each reference slab, providing a “pseudo-average” frame. In parallel, a Hessian filter was applied to these reference scans to improve vessel continuity. Each averaged image was then correlated with its corresponding CNN image, the Hessian image, and the original reference image to assess similarity.
Results :
Figure 1 shows examples of a reference image before processing, the CNN processed image, the Hessian filtered image, and the averaged image. The averaged image has improved vessel continuity, but required a series of 4 scans. Both the CNN output and the Hessian filter improved the vessel continuity from a single scan. Table 1 shows correlations between the averaged image and the reference image, the CNN processed image, and the Hessian-processed image for each subject. It is clear that the former two are both better correlated to the average than the latter.
Conclusions :
CNN-based image processing can improve vessel continuity without the loss of fidelity associated with Hessian filters.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.