Abstract
Purpose :
Image artifacts in optical coherence tomography (OCT) are common problems that compromise the quality of the images, potentially misleading clinical interpretation. Each OCT device provides its proprietary image quality index that is used as the threshold for acceptable image criteria. However, an acceptable measure of such a quality index does not eliminate the presence of various image artifacts. This can lead to unreliable measurements and/or require extra steps to filter artifacts out to build a clean image dataset for research. The purpose of this study was to investigate the prevalence of OCT image artifacts on images qualified for manufacturer’s recommended quality index assessment and see how a deep learning (DL) model can identify bad quality images affected by image artifacts.
Methods :
5,144 OCT images of optic nerve head scans (200x200x1024 samples over 6x6x2 mm3, Cirrus HDOCT, Zeiss, Dublin, CA) were examined retrospectively (597 patients). Image artifacts were classified into eye movement, shadowing, vignetting, banding, and other minor artifacts. The manufacturer’s recommended signal strength cutoff (≥ 6) was used to subset the image dataset, including only acceptable scans, which comprised 4,555 scans of 546 patients. Each scan was subjectively graded from 1 (bad) to 5 (good) based on the number, severity, and location of artifacts on the corresponding en face image. A DL model (EfficientNetV2) was fine-tuned to distinguish good quality (grade 3 and above) from bad quality (grade 1 and 2) images.
Results :
Prevalence of artifacts and the grading outcomes are summarized in Table 1. Figure 1 shows the typical artifacts. Most common artifacts affecting the signal strength qualified images was shadowing followed by eye movement. Of the images, 7.1% and 13.6% of the images were labeled as bad quality or of relatively poor quality. The DL model successfully classified acceptable and not acceptable images at AUROC 96.6%.
Conclusions :
While the manufacturer’s recommend quality score is an important indicator for image quality assessment, in this study, we found a high prevalence of artifacts on en face OCT images with an acceptable quality score. The DL model approach was useful to refine the qualification of OCT images.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.