Purchase this article with an account.
Maciej Szkulmowski, Pawel Liskowski, Bartosz Wieloch, Krzysztof Krawiec, Bartosz Sikorski; Convolutional neural networks for artifact free OCT retinal angiography. Invest. Ophthalmol. Vis. Sci. 2017;58(8):649.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To demonstrate ability of deep convolutional neural network (CNN) to provide noninvasive visualization of retinal microcapillary network (RMN) in retinal diseases in with the use of data from a device combining Scanning Laser Ophthalmoscope (SLO) and Spectral Optical Coherence Tomography (SOCT). SLO system provides fast eye tracking system while SOCT delivers 3D data for knowledge-free vessel segmentation technique. The approach allows RMN to be presented in form of 3D visualization as well as in forms of angiographic maps of different retinal layers free of shadow artifacts blurring standard RMS visualizations.
The study was performed with SOCT laboratory setup (100,000 Ascans/sec, 4.5 um axial resolution, 91 dB detection sensitivity). Constant 30 Hz retinal preview is provided by the SLO device and is used to guide the SOCT scanning beam to the region of interest. RMS maps are created from 3D SOCT data using supervised machine learning algorithm exploiting deep convolutional neural network trained on exemplary data acquired from a set of 15 eyes (from both healthy volunteers and patients with retinal diseases) with vessels labeled by three independent skilled specialists. Training is performed using random-split and biased-split approaches to divide labeled data to training and test sets. Architecture of the CNN consists of 7 layers. Training is carried out by stochastic gradient descent with batch updates and momentum, which is equivalent to optimizing the multinomial logistic regression objective.
Trained CNNs provide sensitivity and specificity for RMN detection in training sets between 0.95 and 0.98 depending on training algorithm. We will show RMN maps obtained using CNNs with both proposed approach for 5 healthy volunteers and 12 patients with diabetic retinopathy, branch retinal vein occlusion and central retinal vein occlusion. We will compare the angiographic maps obtained for different retinal layers using CNN with maps obtained using standard phase-variance and complex difference angiographic algorithms.
Our results shows that CNN approach to RMN visualization provides accurate vessel detection incorporating a priori knowledge of skilled specialists and allows for increased sensitivity and specificity of SOCT based angiography. It also allows for generation of angiographic maps free of artifacts linked to vessel shadows in deeper layers.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only