Abstract
Purpose :
Swept-source optical coherence tomography (SS-OCTA) is an effective modality for detecting choriocapillaris (CC) lesions in patients with posterior uveitis. Currently, manual identification remain time-consuming, inconsistent, and poorly reproducible. Here, we propose the development and validation of an automated lesion boundary deliniation algorithm.
Methods :
For five patients with posterior uveitis, 6x6 En-face SS-OCTA CC images were obtained on the PLEX Elite 9000 (Carl Zeiss AG). Automated flow deficits identification was performed using an established complex pixel intensity thresholding algorithm. Then, the boundary between normal and lesioned CC was identified using a sliding window threshold based on flow deficit size. Large flow deficits were joined into individual lesions. Boundaries were also manually outlined by two clinicans. Intra and inter-observer reproducibility of lesion area and boundary identification were quantitatively determined using the Dice coefficient.
Results :
Intra-observer Dice coefficient reproducibility was high on all images (> 0.90). Average inter-observer reproducibility betwen the two human graders and the algorithm was high for solitary lesions (>0.90), but lower for multifocal lesions (<0.90). An example of a solitary and complex lesion are shown in detail (figure 1). Intra-observer reproducibility of clinician 1 was 0.96 and clinician 2 was 0.98. Inter-observer agreement of clinician 1 and 2 was 0.93. The inter-observer agreement between clinician 1 and AA was 0.95 while between clinician 2 and AA was 0.93. Patient 2's multifocal CC lesions were outlined (figure 2). Intra-observer reproducibility of clinician 1 was 0.95 and clinician 2 was 0.93. Inter-observer agreement of clinician 1 and 2 were 0.87. The inter-observer agreement between clinician 1 and AA was 0.85 and between clinician 2 and AA was 0.86.
Conclusions :
Choriocapillaris lesions can be identified by this automated algorithm, and the algorithim performed similarly to two human graders. However, the automated analysis always provides the same lesion boundary, while variation is provided by the human graders. Ultimately, this algorithm could be used to provide quantitative information about disease severity for improved longitudinal monitoring of patients with posterior uveitis.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.