Abstract
We present a robust deep learning framework for the automatic localisation of cone photoreceptor cells in Adaptive Optics Scanning Light Ophthalmoscope (AOSLO) split-detection images. Monitoring cone photoreceptors with AOSLO imaging grants an excellent view into retinal structure and health, provides new perspectives into well known pathologies, and allows clinicians to monitor the effectiveness of experimental treatments. The MultiDimensional Recurrent Neural Network (MDRNN) approach developed in this paper is the first method capable of reliably and automatically identifying cones in both healthy retinas and retinas afflicted with Stargardt disease. Therefore, it represents a leap forward in the computational image processing of AOSLO images, and can provide clinical support in on-going longitudinal studies of disease progression and therapy. We validate our method using images from healthy subjects and subjects with the inherited retinal pathology Stargardt disease, which significantly alters image quality and cone density. We conduct a thorough comparison of our method with current state-of-the-art methods, and demonstrate that the proposed approach is both more accurate and appreciably faster in localizing cones. As further validation to the method’s robustness, we demonstrate it can be successfully applied to images of retinas with pathologies not present in the training data: achromatopsia, and retinitis pigmentosa.
Highlights
There has already been extensive research on automating cone localisation in images of healthy retinas, with state-of-the-art algorithms obtaining similar-to-human performance[4,5,6,7]
In the proposed approach the 95% confidence interval for the average difference in cone counts was 0.18 ± 1.47, with an upper limit-of-agreement (ULOA) given by 7.34 ± 2.55, and a lower limit-of-agreement (LLOA) of −6.99 ± 2.55; for the retrained convolutional approach we found an average difference of 1.75 ± 2.70, an ULOA of 15.34 ± 4.83, and a LLOA of −11.84 ± 4.83
Through the use of MultiDimensional Recurrent Neural Network (MDRNN) architectures, this paper presents the most robust automatic cone detection algorithm to date
Summary
There has already been extensive research on automating cone localisation in images of healthy retinas, with state-of-the-art algorithms obtaining similar-to-human performance[4,5,6,7]. The innovative introduction of MDRNNs allows the network to consider the entire image whilst classifying a single pixel, as well as being able to take advantage of the highly correlated classifications of neighbouring pixels This is in contrast to the deep learning approach presented in Cunefare et al.[7], which uses a sliding-window convolutional network. A detailed comparison of our approach with the state-of-the-art[4,7] methods on images from healthy volunteers and of volunteers afflicted by Stargardt disease demonstrates that the proposed MDRNN framework is more accurate and significantly faster Following this we show, qualitatively, that, in some cases, the network is able to generalise to images of retinas with pathologies that were not present in the training set, retinitis pigmentosa and achromatopsia
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
