Purchase this article with an account.
LeAnne Young, Jongwoo Kim, Mehmet Yakin, Henry Lin, David Dao, Shilpa Kodati, Sumit Sharma, Aaron Y Lee, Cecilia S Lee, H Nida Sen; Automated detection of vascular leakage on fluorescein angiography. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2112.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Clinician interpretation of fluorescein angiograms (FA) can be subjective. In a translational study of FA images, we aimed to quantify variability of clinician FA segmentation and hypothesized a deep learning algorithm can segment FAs for vascular leakage and help detect clinically significant change.
200 uveitis patient FA images were obtained. As the ground truth, a team of clinicians segmented images for vascular leakage. A deep learning algorithm with a modified U-net architecture was trained to segment leakage. The Dice Similarity Coefficient (DSC) was used to compare the algorithm's segmentation results to the ground truth segmentation (the DSC ranges from 0 to 1, 0 denotes no overlap between 2 segmentations and 1 denotes perfect overlap). For interrater variability, 2 clinicians independently segmented 20 images and the average DSC was calculated. Lastly, 20 pairs of FA images were used to detect clinically significant changes in leakage (the gold standard being an expert clinician’s assessment). For each pair, the difference in percentage of the image occupied by the algorithm’s leakage segmentation was calculated and used to create a ROC curve and to determine a threshold for clinically significant change.
The 200 images came from 61 uveitis patients (median 2 images/patient, IQR 1-4). Diagnoses by anatomic location included anterior (2), intermediate (24), and posterior/panuveitis (35). The average FA timepoint was 361 seconds (SD 174). The algorithm achieved a best average DSC of 0.572 (Fig 1). Concordance between 2 clinicians was lower, with an average DSC of 0.374 (Fig 2). Lastly, a threshold of >0.8% change in the percent of the retinal area covered by the algorithm's segmentation had a 90% sensitivity and specificity for discerning clinically significant leakage (AUC: 0.95).
A preliminary deep learning algorithm was developed, had initial success in segmenting leakage in uveitis patient FAs, and was used to determine clinically significant change in vascular leakage. Further algorithm development is needed to improve segmentation accuracy. However, this is a first step to more standardized FA interpretation which can be useful as clinical trial outcomes.
This is a 2021 ARVO Annual Meeting abstract.
Algorithmic segmentation overlaid on ground truth segmentation. Example of higher (a) and lower (b) concordance between algorithm and ground truth segmentation
Example of higher (a) and lower (b) concordance between two clinicians
This PDF is available to Subscribers Only