Purchase this article with an account.
Warren Lewis, Arindam Bhattacharya, Luis de Sisternes; Increased vessel length index in deep retinal layer angiography en face slabs via convolutional neural networks. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB058.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Assessment of average vessel length in deep retinal layer (DRL) projections is complicated by speckle and brightness variations, which can cause vessels to appear broken in binarized skeleton traces. Hessian “vesselness” filters have been used to correct this, but have limited ability to improve connectivity, and may create spurious vessels from background noise. This is a study of the use of convolutional neural networks (CNN’s) as an alternative means to restore vessel continuity.
A modified 5-layerCNN with deep supervision was trained on OCTA images of the deep capillary plexus, using averaged scans of the same subject as ground truth. Approximately 30,000 patches of individual images were trained against patches of the averaged image. A PLEX® Elite 9000 SS-OCT System (ZEISS, Dublin, CA) was used to acquire 3-4 repeated 6x6 mm OCTA scans of subjects having vascular pathology. These scans were registered and averaged, providing high quality volumes, which were then segmented to produce en face projections of the DRL. The individual volumes were also segmented to produce DRL projections. The CNN was used to process each of these individual projections. Vessel traces were calculated from the CNN output images and compared with traces made using a traditional method using Hessian filters. The average vessel length was compared between the two groups.
See Figure and Table. The CNN output produces a vessel trace with much better continuity. Examination of the figure shows that the improved continuity is a good reflection of the original structure as captured in the scan. Table shows average vessel length calculated for scans of 3 subjects. Although there was significant variation in the subject vasculature, this is not apparent in the results obtained with Hessian vessel enhancement. The CNN output shows a much clearer variation among the scans, which is reflective of the original data.
CNN-based vessel enhancement is a new, promising alternative to “vesselness ” filters for improved vessel continuity.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
Figure. 1st row, Original, CNN output; 2nd row, vessel traces from Hessian filtered original image and CNN output, 3rd row, sets 2 and 3 DRL slabs.
<p align="center" style="margin: 0px 0px 13.33px; text-align: center;"><font color="#000000" face="Calibri" size="3">Table. Average vessel length (mm) calculated usingHessian and CNN processing</font></p>
This PDF is available to Subscribers Only