AUTHOR=Wang Yi-Zhong , Birch David G. 

TITLE=Performance of Deep Learning Models in Automatic Measurement of Ellipsoid Zone Area on Baseline Optical Coherence Tomography (OCT) Images From the Rate of Progression of USH2A-Related Retinal Degeneration (RUSH2A) Study

JOURNAL=Frontiers in Medicine

VOLUME=Volume 9 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2022.932498

DOI=10.3389/fmed.2022.932498

ISSN=2296-858X

ABSTRACT=Purpose: Previously, we have showed the capability of a hybrid deep learning (DL) model for automatic segmentation of retinal layers from OCT images in retinitis pigmentosa (RP). One of the shortcomings of the model is that it tends to underestimate ellipsoid zone (EZ) area when EZ extends toward or beyond the edge of the macula. In this study, we trained the model with additional data and evaluated its performance in automatic measurement of EZ area on SD-OCT volume scans obtained from the participants of the RUSH2A study by comparing the model’s performance to the reading center’s manual grading. Methods: De-identified Spectralis high-resolution volume scans and their EZ area measurements by a reading center were obtained from the RUSH2A study. Eighty-six baseline volume scans from 86 participants were included to evaluate two hybrid models: the original RP240 model trained on 480 B-scans from 220 participants with RP and 20 participants with normal vision, and the new RP340 model trained on a revised RP340 dataset including RP240 plus additional 200 B-scans from another 100 participants with RP. EZ and apical RPE were automatically segmented by the hybrid models and EZ areas were determined. Dice similarity, correlation, linear regression, and Bland-Altman analyses were conducted to assess the agreement between the EZ areas measured by the models and by the reading center. Results: For the RP240 hybrid model and the RP340 hybrid model, dice coefficients ± SD with manual grading for EZ area>1 mm2 were 0.835±0.132 and 0.867±0.105, respectively. Correlation coefficients (95% CI) were 0.991 (0.987–0.994) and 0.994 (0.991–0.996), respectively. Linear regression slopes (95% CI) were 0.918 (0.896–0.940) and 0.995 (0.975–1.014), respectively. Bland-Altman analysis revealed a mean difference±SD of -0.137±1.131 and 0.082±0.825 mm2, respectively. Conclusions: Additional training data improved the hybrid model’s performance. Automatic EZ area measurements generated from DL models significantly correlate with those by the reading center. The close agreement of DL models to manual grading suggests that DL may provide effective tools to significantly reduce the burden of segmentation and provide retinal layer measurements to help assess disease progression and to facilitate the study of structure function relationship in RP.