Retinal optical coherence tomography (OCT) is increasingly used for quantifying neuroaxonal damage in diseases of the central nervous system such as multiple sclerosis. High-quality OCT images are essential for accurate intraretinal segmentation and for correct quantification of retinal thickness changes. The quality of OCT images depends largely on the operator and patient compliance. Quality evaluation is time-consuming, and current OCT image quality criteria depend on the experience of the grader and are therefore subjective. The automatic graderindependent real-time feedback system for quality evaluation of retinal OCT images, AQuA, was developed to standardize quality evaluation and data accuracy. It classifies by signal quality, anatomical completeness and segmentation plausibility and has been validated by experienced graders. However, it is currently limited to OCT scans taken with one device from a single vendor. The aim of this work is to improve the capability of the AQuA quality classifier to generalize to new data, by developing a convolutional neural network (CNN), AQuANet. Moreover, this CNN may serve as a basic quality classifier, that can be adapted to specific problems by transfer learning. AQuANet is trained on A-Scan batches with quality labels automatically obtained with AQuA. Thus, a large set of training data of about 13000 A-Scan batches could be used, leading to an accuracy of 99.53%.

Date Published: 26 February 2019
PDF: 11 pages
Proc. SPIE 10868, Advanced Biomedical and Clinical Diagnostic and Surgical Guidance Systems XVII, 1086814 (26 February 2019); doi: 10.1117/12.2510393

Published in SPIE Proceedings Vol. 10868:
Advanced Biomedical and Clinical Diagnostic and Surgical Guidance Systems XVII
Anita Mahadevan-Jansen, Editor(s)