Purchase this article with an account.
Jessica Loo, Cindy Xinji Cai, Emily Chew, Martin Friedlander, Glenn J Jaffe, Sina Farsiu; Deep learning-based automatic segmentation of retinal cavitations on OCT images of MacTel2. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1616.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop an automatic method to segment retinal cavitations on optical coherence tomography (OCT) images of macular telangiectasia type 2 (MacTel2)
Retinal cavitations are a MacTel2 biomarker visible on OCT images as hypo-reflective spaces that are often angular or irregular in shape. They differ from cysts seen in exudative conditions which tend to be round or oval and are less irregular. The dataset consisted of 9501 B-scans from OCT volumes of 99 eyes enrolled in an international, multi-center, phase 2 interventional MacTel2 clinical trial (NCT01949324). Cavitations tend to be located in the macula, and in particular, the temporal fovea. In this dataset, they were present in 8.1% of the B-scans. Therefore, we undertook a two-stage approach to automatically segment the images. First, a convolutional neural network (CNN1) was trained to detect B-scans containing cavitations. Next, a second network (CNN2) was trained to segment the cavitations in B-scans. The CNNs were trained and validated with B-scans randomly sampled from 88 volumes and tested on all B-scans from the remaining 11 volumes. During testing, CNN1 was used to classify the probability that each B-scan contained cavitations. For those with a probability ≤ 0.5, CNN2 automatically set the final segmentations to not contain any cavitations. Otherwise, CNN2 predicted probability maps of the cavitations, which were thresholded and post-processed to obtain the final segmentations. Performance was evaluated using gold standard manual segmentations by a trained Reader and compared to the alternative one-stage approach, whereby all B-scans were segmented by CNN2 without being first classified by CNN1.
Using 9-fold cross-validation and the proposed approach, the sensitivity and specificity of CNN1 were 0.93 and 0.80, respectively, and the average Dice similarity coefficient (DSC) of CNN2 was 0.94 ± 0.21 across all 9501 B-scans. Figure 1 shows some examples. Using the alternative one-stage approach, the average DSC was 0.85 ± 0.34.
Overall, there was good agreement between the manual and automatic segmentations. This algorithm will be useful to quantify retinal cavitations and to assess longitudinal structure-function correlations in MacTel2 as they can potentially be a negative predictor of visual acuity.
This is a 2020 ARVO Annual Meeting abstract.
Example segmentations of retinal cavitations by a trained Reader (manual) and the proposed algorithm (automatic).
This PDF is available to Subscribers Only