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.

ABSTRACT

Purpose: Previously, we have shown the capability of a hybrid deep learning (DL) model that combines a U-Net and a sliding-window (SW) convolutional neural network (CNN) for automatic segmentation of retinal layers from OCT scan images in retinitis pigmentosa (RP). We found that one of the shortcomings of the hybrid model is that it tends to underestimate ellipsoid zone (EZ) width or area, especially when EZ extends toward or beyond the edge of the macula. In this study, we trained the model with additional data which included more OCT scans having extended EZ. We evaluated its performance in automatic measurement of EZ area on SD-OCT volume scans obtained from the participants of the RUSH2A natural history study by comparing the model's performance to the reading center's manual grading. Materials and Methods: De-identified Spectralis high-resolution 9-mm 121-line macular volume scans as well as their EZ area measurements by a reading center were transferred from the management center of the RUSH2A study under the data transfer and processing agreement. A total of 86 baseline volume scans from 86 participants of the RUSH2A study were included to evaluate two hybrid models the original RP240 model trained on 480 mid-line B-scans from 220 patients with retinitis pigmentosa (RP) and 20 participants with normal vision from a single site, and the new RP340 model trained on a revised RP340 dataset which included RP240 dataset plus an additional 200 mid-line B-scans from another 100 patients with RP. There was no overlap of patients between training and evaluation datasets. EZ and apical RPE in each B-scan image were automatically segmented by the hybrid model. EZ areas were determined by interpolating the discrete 2-dimensional B-scan EZ-RPE layer over the scan area. Dice similarity, correlation, linear regression, and Bland-Altman analyses were conducted to assess the agreement between the EZ areas measured by the hybrid model and by the reading center. Results: For EZ area > 1 mm2, average dice coefficients ± SD between the EZ band segmentations determined by the DL model and the manual grading were 0.835 ± 0.132 and 0.867 ± 0.105 for RP240 and RP340 hybrid models, respectively (p < 0.0005; n = 51). When compared to the manual grading, correlation coefficients (95% CI) were 0.991 (0.987-0.994) and 0.994 (0.991-0.996) for RP240 and RP340 hybrid models, 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 mm2 and 0.082 ± 0.825 mm2, respectively. Conclusion: Additional training data improved the hybrid model's performance, especially reducing the bias and narrowing the range of the 95% limit of agreement when compared to manual grading. The close agreement of DL models to manual grading suggests that DL may provide effective tools to significantly reduce the burden of reading centers to analyze OCT scan images. In addition to EZ area, our DL models can also provide the measurements of photoreceptor outer segment volume and thickness to further help assess disease progression and to facilitate the study of structure and function relationship in RP.