Abstract
Purpose :
To determine if a classical foveal avascular zone (FAZ) segmentation method gives reliable measurements on large field of view (FOV) OCTA images compared to small FOV OCTA images.
Methods :
OCTA images were obtained on 25 eyes, including normal and diabetic retinopathy pathology, using the PLEX® Elite 9000 SS-OCT (ZEISS, Dublin, CA) with a small FOV of 3x3 mm (300 A-lines x 300 B-scans) and a large FOV of 9x15 mm (500 A-lines x 834 B-scans). The full retina layer was used for FAZ segmentation to encompass all boundary capillaries. To ensure images with the same physical size (in mm) correspond to the same pixel size, the large FOV images were resampled and the central 3x3 mm images were cropped for FAZ segmentation. An active contour algorithm [Ref 1] ensured a smooth FAZ boundary and functioned in two stages: 1) approximate location and size of the FAZ were determined from binary images of the full retina layer, and 2) this rough FAZ was expanded to encompass the entire FAZ using a fixed number of iterations of the active contour algorithm. Once the edge contour and interior of the FAZ were determined, the perimeter, area, and circularity index were calculated. Paired t-test quantified the differences in FAZ measurements between small and large FOV images.
Results :
Although large FOV images were resampled to align with the pixel size of small FOV images, the signal-to-noise ratio (SNR) was still low. A notable distinction was found in the FAZ area measurements (p<0.0001), where a reduced FAZ area (0.24±0.10 mm2) was found in large FOV images in comparison to small FOV images (0.32±0.11 mm2). Furthermore, a significant (p=0.007) decrease in FAZ perimeter was observed in large FOV images (2.32±0.62 mm) when compared to small FOV images (2.63±0.52 mm). No significant differences were found in circularity index (0.59±0.12 vs 0.56±0.13; p=0.31) between small and large FOV images. Fig. 1 and 2 present examples of FAZ segmentation in small and large FOV images, illustrating cases with good and poor agreement, respectively.
Conclusions :
A significant difference in FAZ area was found in large FOV images due to the low SNR. Caution should be taken when presenting FAZ segmentation in large FOV images using a classical approach. Deep learning methods have the potential to improve the accuracy of FAZ segmentation in large FOV images.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.