Abstract
Purpose :
Optical coherence tomography angiography (OCTA) en face scans can be subject to artifacts from media opacities, eye movement, and dry eyes that may occur only in parts of the image. This study introduces a method to generate quantitative maps that describe the quality of an OCTA acquisition at each en face location. The values recorded can be averaged to provide a score that correlates with subjective quality grades.
Methods :
A machine learning linear regression model is trained to recognize the relationship between the texture properties and assigned quality grades of OCTA images. Grades were assigned on a 0-5 scale by two graders, with higher values indicating higher image quality (sharpness, visibility of clinical features, reduced artifacts). The training was done on 150 spectral-domain OCT (SD-OCT) angiography images captured on CIRRUS™ HD-OCT 5000 with AngioPlex® OCT Angiography (ZEISS, Dublin, CA) (57 3x3mm, 56 6x6mm, and 37 8x8mm scans from 29 eyes), and 452 swept-source OCT (SS-OCT) images captured on PLEX® Elite 9000 (ZEISS, Dublin, CA) (378 3x3mm, 37 6x6mm, 37 12x12mm from 39 eyes). The model was tested using 133 SD-OCT images (52 3x3mm, 44 6x6mm, and 37 8x8mm scans from 29 eyes) and 243 SS-OCT images (6x6mm scans from 49 eyes). The correlation of the quality scores predicted by the model and those assigned manually as well as the visual correlation of the quality maps with regional quality in individual acquisitions were assessed.
Results :
The model performed well for the CIRRUS images with mean absolute score error (MASE) of 0.49 ± 0.37 in the 3x3mm scans, 0.6 ± 0.4 in 6x6 scans, and 0.6 ± 0.41 in 8x8 scans. Overall MASE of the model for CIRRUS data was 0.56 ± 0.39. The MASE for the PlexElite was 0.6 ± 0.48. For scans of differing quality acquired on the same eye, the generated maps corresponded well with visual assessment of the images (Figure 1).
Conclusions :
The machine learning model was able to assess the quality of OCTA scans with relative accuracy and produce regional quality maps that correlated well with visual quality of the data. This method may be useful in providing quantitative feedback to operators when acquiring OCTA data, such as prompting the operator to reacquire an image if the quality metric is below a certain threshold.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.