Purchase this article with an account.
Yi-Zhong Wang, Daniel Galles, Martin Klein, Kirsten G Locke, David G Birch; A Deep Convolutional Neural Network (CNN) Model for Automatic Measurement of Ellipsoid Zone (EZ) Width in Retinitis Pigmentosa (RP). Invest. Ophthalmol. Vis. Sci. 2019;60(9):1527.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Previous studies suggested that EZ width is an effective biomarker for disease progression in RP (Birch et al. 2013). Conventional automated OCT image analysis software often incorrectly identifies EZ in RP, thus its accurate delineation requires manual correction which is time-consuming and costly. Here, we developed and evaluated a deep CNN model for automatic measurement of EZ width from OCT B-scan images.
A CNN to classify tiny images (Fang et al., 2017) was adopted and implemented in MATLAB. Horizontal midline B-scan with visible EZ from 88 RP patients and 20 normal subjects were used to generate image patches (33x33 pixels) for training and testing the model. Prior to being imported to MATLAB, B-scan images were segmented automatically using Spectralis software, then manually corrected for the boundaries of ILM, EZ, RPE, and BM. Positive training image patches were those centered on a boundary. A total of 600,000 patches generated from one eye of each subject were used for training (80%) and validation (20%). A separate test dataset of patches generated from fellow eyes was used to evaluate the trained CNN model. All pixels in B-scan images of the fellow eyes of RP patients were then classified to obtain probability maps for each boundary class. A local connected area searching algorithm was developed to process the maps for reconstructing boundaries and measuring EZ width within central 8 mm. Correlation and Bland-Altman analyses were conducted to compare EZ width measured by the model to those by manual correction.
At the end of CNN training (45 epochs), the accuracy (the model correctly identified the class of image patches in validation set) was 97%. For the test dataset, the accuracy of the model to identify ILM, EZ, RPE, and BM patches was 96%, 97%, 97%, and 97%, respectively. The correlation between the EZ width measured by the model with that by manual correction was 0.96 (p<0.05). Bland-Altman analysis revealed a mean difference of 0.20 mm between the model and the manual measurements of EZ width.
These results demonstrated the capability of deep machine learning methods for automatic EZ width measurement in RP, suggesting that well-trained and validated CNN models can be used to quantify structural deficits for detecting disease progression and for evaluating treatment effect in future clinical trials for RP.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only