Paper Submission & Registration
8th Dutch Bio-Medical Engineering Conference
16:30   Eye
Chair: John van Opstal
15 mins
Optimal control may underlie Listing’s Law for eye movements
John van Opstal, Carlos Aleluia, Akhil John, Alexandre Bernardino
Abstract: The six extra-ocular muscles give the eye three degrees of rotational freedom, but only two d.o.f. are required to look in any direction. As a result, the eyes are controlled by Donders' law, which restricts ocular cyclotorsion to a 2D surface describing the horizontal and vertical components of gaze. With the head erect and still, and gazing at far-visual space, Donders' surface becomes a plane: Listing's plane (LP), restricting ocular cyclo-torsion to zero. Interestingly, fast ocular saccades obey Listing's law (LL) also during their trajectories, indicating that they are controlled as single-axis rotations. However, because saccades can be generated with 3 d.o.f. when the eyes are brought out of LP (e.g., during head tilts, vestibular stimulation, spontaneously, or by brainstem microsimulation), there are strong reasons to assume a neural 3D control of eye movements, as no mechanical system (e.g., pulley-based) accountsfor these behaviours. In addition, saccade kinematics are determined by their nonlinear ‘main-sequence’ relations for peak eye velocity, saccade duration, and shape of the velocity profiles. Finally, saccade trajectories are nearly straight (in line with single-axis rotations), so that torsional, horizontal, and vertical angular velocity components are scaled versions of each other (component cross-coupling). We developed a computational model, that faithfully represents the physics and geometry of a biomimetic robotic eye, with six extra-ocular elastic 'muscles', controlled in antagonistic pairs by three independent motors, subjected to multiplicative noise. We show that open-loop feedforward optimal control of the model eye fully accounts for the 3D behavior (LL), nonlinear saccade kinematics and straight trajectories, by minimising a cost function that contains: (i) the total force exerted by the muscles at any fixation direction ('minimising co-contraction'), (ii) saccade duration (hyperbolic discount of reward acquisition), and (iii) 2D accuracy at saccade end (i.e., eye on target). We show that the muscle insertion points on the globe may underlie the observed tilt of Listing's plane with respect to straight ahead. In conclusion, Listing's law and nonlinear kinematics may be due to a central control strategy that aims to optimise speed-accuracy trade-off and fixation effort
15 mins
Automatic vessel segmentation and wall-to-lumen ratio quantification in adaptive optics ophthalmoscopy
Roan van Bakel, Elise Bakker, Felix Dikland, Thomas Poldervaart, Danilo Andrade De Jesus, Luisa Sanchez Brea, Stefan Klein, Theo van Walsum, Daniela Castro Farías, Kate Grieve, Michel Paques
Abstract: Retinal vascular diseases are a leading cause for blindness and partial sight certifications [1]. By combining adaptive optics (AO), a technique that counteracts optical aberrations in the eye, with fundus photography, high resolution images of the retinal microvasculature can be acquired [2,3]. The wall-to-lumen ratio (WLR) is a quantitative vascular biomarker which is significantly different between hypertensive patients and healthy subjects [2]. The current algorithms for calculating WLR based on AO ophthalmoscopy are limited to single local measurements manually chosen by the clinician. Also, different vessel diameters are not taken into consideration, even though greater vessel diameters are correlated with a lower WLR in both normotensive and hypertensive subjects. In this work, an algorithm has been developed to tackle the current limitations of the WLR estimation. The proposed algorithm outputs the mean WLR of three vessel diameter classes. For a single image, the WLR was calculated along the entire vessel. Firstly, the midline of the retinal vessels was found by segmenting the axial reflection. Then, the distances to inner and outer diameter on either side of the midline were estimated by localising changes in pixel intensity using a perpendicular scanning method and optimised with an iterative linear regression model. The mean WLR was then calculated from the inner and outer diameter. For the algorithm development, a set of 23 retinal AO volumes of 23 subjects was used, with the volumes ranging from 10 to 60 images. The results showed that the proposed algorithm is able to tackle the current limitations associated with the WLR estimation. A comparison of the calculated WLR with a reference algorithm using a Bland-Altman analysis was also performed, resulting in a mean difference of 16.6% (p<0.0001). In conclusion, WLR can be measured in an automated way from AO ophthalmoscopy images. Such measurements may help to improve the diagnosis and follow up of vascular retinal pathologies. Nevertheless, further work needs to be done for the algorithm validation in clinical practice. References [1] C. S. Brand. “Management of retinal vascular diseases: a patient-centric approach”. In: Eye (Lond) 26 Suppl 2 (2012), S1–16. [2] E. Koch et al. “Morphometric analysis of small arteries in the human retina using adaptive optics imaging: relationship with blood pressure and focal vascular changes”. In: J Hypertens 32.4 (2014), pp. 890-8. [3] M. Paques et al. “Adaptive optics ophthalmoscopy: Application to age-related macular degeneration and vascular diseases”. In: Prog Retin Eye Res 66 (2018), pp. 1–16.
15 mins
Glaucoma prediction based on 3D-OCT speckle imaging
Vania Bastos Silva, Luisa Sánchez Brea, Theo van Walsum, Stefan Klein, João Cardoso, Robert Iskander, Monika Danielewska, Malgorzata Kostyszak, João Barbosa Breda, Jan Van Eijgen, Ingeborg Stalmans, Danilo Andrade De Jesus, Pedro G. Vaz
Abstract: Glaucoma is an insidious and unpredictable disease, being the leading cause of irreversible blindness worldwide. However, its damage is preventable, making an early diagnostic and close progression monitoring of the utmost importance. Optical Coherence Tomography (OCT) is a low-coherence interferometry technique widely used as a tool to analyze relevant biomarkers for glaucoma diagnosis. During image acquisition, an optical beam is directed at the retina, and most of the light is scattered in different angles rather than directly reflected. The light travels through different optical paths, with different lengths, until it reaches the image plane, where the intensity of each point is the superposition of all the waves in that point. This interference phenomenon creates granular patterns called speckle patterns. Speckle patterns in OCT signals can have a dual role, both as source of noise, signal-degrading speckle, and as carrier of information, signal-carrying speckle [1]. Speckle analysis has many useful biomedical applications, but early changes in patterns of OCT images from glaucoma patients have yet to be studied. These patterns have the potential to add new information to the features currently used and improve glaucoma prediction methods. Two and three dimensional speckle fields can be simulated with algorithms which accurately represent reality [2]. Such models allow the study of data properties and help develop and validate image-based algorithms, such as spatial contrast, that retrieve speckle-related information. These algorithms can later be translated to real data, to extract additional features and understand the variability of speckle measurements in glaucomatous patients 3D-OCT data. In this work, a new 3D-model to study and infer OCT speckle properties in glaucomatous data is being developed. The speckle-related features and other clinically relevant data (e.g age, retinal fiber layer thickness, microvascular density, etc) will be used to develop a predictive model based on discriminative Event-Based modelling (EBM): a data-driven technique to estimate the sequence in which biomarkers for a disease become abnormal. The data of 300 subjects from the Leuven Eye Study cohort will be used. Concluding, in this project, we expect to contribute to glaucoma prediction by developing a new model combining OCT-speckle information and other clinically relevant features
15 mins
Automated lamina cribrosa segmentation in optical coherence tomography scans of healthy and glaucomatous eyes
Rita Marques, Luisa Sánchez Brea, Theo van Walsum, Stefan Klein, João Cardoso, Jan Van Van Eijgen, Ingeborg Stalmans, João Breda, Mariana Dias, Pedro G. Vaz, Danilo Andrade De Jesus
Abstract: Glaucoma is the main cause of irreversible blindness worldwide and it starts to manifest through damage to the retinal ganglion cell (RGC) axons as they exit the eye at the optic nerve head (ONH). Lamina cribrosa (LC), on the other hand, is a mesh-like structure that fills the posterior scleral foramen through which unmyelinated RGC axons pass before converging in the ONH [1]. Therefore, the LC is an important anatomical part in the pathogenesis of glaucoma and recent advances in optical coherence tomography (OCT) imaging have made it easier to study in greater detail [1]. A number of studies have already explored the segmentation of the retinal layers and choroid in OCT imaging using deep learning (DL) techniques. However, the current literature on automated segmentation of the ONH and adjacent structures with DL methods is scarce [2]. In this work, a DL-based algorithm is going to be developed in order to attempt an automatic segmentation of several structures at the ONH including the lamina cribrosa, the retinal pigment epithelium, the retinal nerve fiber layer (RNFL), the RGC layer, and the choroid. For this purpose, data from 300 subjects from the Leuven Eye Study cohort will be used for training and testing the algorithm. The segmented structures will be used to extract clinically relevant data such as LC depth, LC thickness, LC curvature, prelaminar tissue, RNFL thickness, Bruch’s membrane opening minimum rim width and focal defects. In summary, the main goal of this project is to develop a robust framework to use in 3D segmentation of the ONH built on a strong training set that can be validated through multiple OCT devices. The features extracted from the automated segmentation will aid the comprehension, early diagnosis, and follow-up of glaucoma disease in clinical practice.

end %-->