During the last 9 months I have mainly worked on the development of a full model of the eye. In particular I have developed a poro-elastic model for the choroid and I have coupled it with a simple model of the vitreous. To obtain these results I have worked on the development of a library that can be used to couple different solvers. The idea is to develop adhoc software for the simulation of each compartment of the eye (for example: the choroid, the vitreous, the cornea, the anterior chamber, etc…) and make them interact with each other by using this library.
I have also finalised the publication on the reduced order technique that we have developed to speed up those simulations, alongside writing my PhD thesis.
In the final 6 months of the project I mainly focused on two things: Firstly, I closely collaborated with another ESR in order to apply mathematical modelling in estimating haemodynamic features. This was applied in the context of identifying important biomarkers of progression to diabetic retinopathy. Our common work concluded with writing up a journal paper, which was submitted and published. Secondly, I also focused on writing up my PhD thesis, recapping on all of the research that I conducted during my MSCA fellowship. The thesis was submitted at the end of my contract and I successfully passed my examination.
In the last reporting period, a model for predicting patients at risk of developing diabetic retinopathy was designed based on patient characteristics and clinical measurements. The model yielded a cross-validated AUC of 0.77 ± 0.04 and could be used to identify patients at risk of developing retinopathy, resulting to a significant reduction of the number of screening visits.
The below poster was also presented at the European Researchers’ LiGHTS Nights 2016 Science Festival hosted by the University of Lincoln:
This reporting period my main contribution has been my work in optic disc detection and curvature enhancement. Both methods proposed were able to outperform state of the art methods on publicly available datasets. Both methods are also very adaptable (this being the main feature of the tortuosity estimation). The curvature enhancement method has allowed for a new set of tortuosity metrics to be created. By changing the filter banks, these metrics can be adapted to better estimate the tortuosity evaluation of an individual grader.
Figure 1: Examples of different curvature enhancements on a set of synthetic vessels using the proposed method for curvature enhancement.
The progression from diabetes to diabetic retinopathy is associated with changes in retinal haemodynamics. In collaboration with Georgios, a longitudinal study of twenty-four subjects was conducted. In this study, we monitored the retinal haemodynamics during the three years before the appearance of diabetic retinopathy (DR) and in the first year of DR. We took vascular measurements from standard fundus images, and estimated fluidynamic parameters using a simple haemodynamic model. We show that there are statistically significant changes in some estimated haemodynamic parameters associated with the development of DR.
Please refer to the related paper “Hemodynamics in the retinal vasculature during the progression of diabetic retinopathy” (F. Calivá, G. Leontidis, P. Chudzik, A. Hunter, L. Antiga, B. Al-Diri), which was successfully submitted to the Journal for modeling in Ophthalmology.
During the last period, I have been focusing on revising our Medical Image Analysis paper presenting and validating the whole framework for automated corneal nerve image tortuosity estimation and interpretation. Moreover, I have been drafting and revising an IEEE Transactions on Medical Imaging paper on accelerating convolutional sparse coding for curvilienar structure segmentation and a new, more robust version of the SCIRD filters (see Figure) I proposed at MICCAI 2015. In the last part of the period I have been writing my PhD thesis and successfully passed the viva-voce examination on the 12th of Aug, subject to some very minor revisions. In parallel, I have tested the system I developed on a large data set of ~500 images from subject with different pathologies and a clinical paper is being written by our collaborators at Harvad Medical Schiool and Tufts Medical Center to report the main outcomes.
For more information, please go to: http://staff.computing.dundee.ac.uk/rannunziata
The final results of my project show that after occlusion, unidirectional flow towards the occlusion site was established. Diameters exhibited an immediate decrease, an ensuing increase with a time constant of 2.53 ± 0.77h (mean ± SD, n = 13 CAMs) and then a plateau of up to 60% above baseline without significant vascular tone change. WSS exhibited an immediate increase, a linear decrease to baseline approximately 12h post-occlusion and then remained unchanged. WSS increase upon occlusion and sustained diameter changes were correlated. WSS change only partly mirrored the diameter change rate through 24h post-occlusion.
The algorithm for the automatic detection of the ONH is improved and is now the fastest and second best (0.26% difference) in terms of performance compared to the best in literature, when tested on MESSIDOR dataset. Qualitative results are also improved with an elliptical approximation of the ONH boundary. This was possible by describing the photograph distortion is caused by the quasi-spherical shape of the eye.
A probability map is created based on a prospective database. The database includes 60 patients: 30 developed vision threatening DR / 30 did not. Approximately 10-year record from screening database for each patient (for an average of 7 visits each). The dataset includes more than 900 images and the analyses is made possible with the tools previously developed. The map returns the probability of developing vision threatening diabetic retinopathy given the location of the first detected microaneurysm for patients included in a diabetic retinopathy screening programme.
Example result from the ONH detection with elliptical correction:
The model of the eye used to quantify the distortion due to the orthogonal projection
of a curved surface:
Probability of developing vision threatening diabetic retinopathy given the location of
the first detected microaneurysm for patients included in a diabetic retinopathy
During this reporting period, Carlos conducted research regarding the utilisation of multi-modal cues as well as the accuracy of eye shape estimation for the retinal image registration method. Additionally, an image dataset for retinal image registration (FIRE: Fundus Image Registration Dataset) was made publicly available at http://www.ics.forth.gr/cvrl/fire/ containing 134 image pairs falling under 3 different categories and including ground truth for registration evaluation.
Type 2 diabetes represents the most prevalent cause for progression of diabetic retinopathy. However, there is a lack of appropriate translation animal models. Although the renin-angiotensin system plays an important role in the progression of diabetic retinopathy, its influence in diabetic retinopathy has not been systematically evaluated. Here we test the suitability of a new model, the TetO rat, addressing the role of angiotensin-II receptor 1 (AT1) blockade in experimental diabetic retinopathy.
Diabetes was induced by tetracycline-inducible small hairpin RNA (shRNA) knock-down of the insulin receptor in rats (TetO). Systemic treatment consisted of an AT1-antagonist (ARB) at the onset of diabetes. 4-5 weeks later, the retina was analyzed in vivo and ex vivo. Retinal function was assessed by Ganzfeld ERG.
Retinal vessels in TetO showed calibre differences together with gliosis. The total number as well as the proportion of activated mononuclear phagocytes was increased. TetO presented loss of retinal ganglion cells (RGC). ERG indicated photoreceptor malfunction. Both the inner and outer blood-retina-barrier were affected.
The ARB-treated group presented reduced gliosis and an overall amelioration of the retinal function together with RGC recovery. No statistically significant effects on vascular and inflammatory features were found.
TetO rat represents a promising translational model for early neurovascular changes of type 2 diabetic retinopathy. ARB treatment revealed an effect on the neuronal component but not on the vasculature.
A software tool was created for the manual supervision of lesions resulting from the automatic algorithm. The manually supervised detection of microaneurysms in the prospective database was concluded. The new method for the quick detection of the optic disk was improved, completed with an approximate segmentation and optimised for speed and accuracy. The method for the fully automatic detection of 5 clinically relevant regions was integrated with the new, fast algorithm for the optic disk detection. A new study showed that the distribution of microaneurysms in high-risk patients was higher in the upper than in the lower hemiretina (see Figure: Cumulative distribution of microaneurysms in the left and right eye).
In this reporting period, Francesco designed and implemented an algorithm that classified, as artery and vein, blood vessels along the two main retinal arcades. This study was submitted and accepted at the ARVO annual meeting.
He also collaborated with ESR 3.3, Giovanni Ometto, in a fruitful collaboration that led to the submission of a conference paper to the IEEE SPIE Medical Imaging conference.
Effort has been put during this reporting period to enhance the predictive analysis that has commenced since last semester. This has been achieved by including more variables in the analysis and currently, focusing on also including time-invariant information, i.e. measurements of clinical features. Some initial results indicate a decent model fit, particularly for the initial states, however, there is still space for improvement as the figures of model fit below suggest.
During the reporting period, simulations by means of the model of the retinal vascular network were completed. A manuscript on modelling was written and submitted to the Bulletin of Mathematical Biology.
The data obtained during animal experiments last summer was analysed. In February the manuscript was approved and published in Biomedical Optics Express.
Preparations for a new series of experiments were performed. These experiments are designed to study vasomotion in rat retinal vasculature. Results are planned to be submitted to a special issue in the Journal for Modeling in Ophthalmology.
Over the past six months, I have finished data collection from my experiment on early adaptation of arteriolar collaterals in chick chorioallantoic membrane (CAM) which will be presented in my paper that is to be submitted later this month. In general, our work shows that after occlusion, unidirectional flow towards the occlusion site was established. Diameters exhibited a significant immediate decrease, followed by an increase with a time constant of 2.5 ± 0.8h (mean ± SD, P<0.05) leading to a plateau up to 60% above the baseline value without significant vascular tone change. In contrast, WSS exhibited a significant increase immediately post-occlusion, followed by an approximately linear decrease returning to the baseline value after about 12h. Maximal increase in WSS and sustained diameter changes were significantly correlated (see figures below).
After joining the REVAMMAD project, I started a training about machine learning techniques focusing on ensemble methods. Later, in the frame of the data warehouse task, I started to plan an application for data and algorithm sharing together with the Orobix team. In order to best suit the needs of the collaboration, I developed a system that supports both centralized and peer-to-peer exchange of datasets and algorithms, also exploiting the software libraries currently being developed in an open source project (DAT project). Most of the time of this reporting period has been dedicated to the implementation and the deployment of a prototype of this application. More recently, I began to investigate the application of artificial neural networks for segmentation tasks in medical imaging.
This reporting period my main contribution has been my work in tortuosity. We have gone in a different direction than most methods, which rely on a rigid mathematical formulation of curvature, to a feature-based method which relies on enhancing curvature in an image. This allows for a tuneable method that doesn’t require a perfect segmentation or vessel centreline to accurately measure tortuosity in a vessel, and we hope to expand the idea to the image level.
The segmentation module of our fully automated tortuosity estimation system has been improved both in terms of modelling curvilinear structures and learning context filters. Moreover, a novel approach to accelerate convolutional sparse coding filter learning has been developed. This is expected to make filter learning much more discriminative and result in a more robust segmentation module. The tortuosity plane has been proposed as a better tool to quantify and interpret tortuosity (an example of how the tortuosity plane is used to map corneal nerve images in terms of tortuosity is included). Please refer to my webpage for further details. http://staff.computing.dundee.ac.uk/rannunziata/publications.html
In the last six months my work has focused on implementing an alternative method for summarising the retinal vessel calibres in more than one group (healthy, diabetics and DRs), which will yield and estimate the trunk vessel more accurately than the current method in literature. In addition to that, the different areas inside the retina were defined, using as landmarks the optic nerve head and fovea. These areas, that have been found to include the most lesions during the progression of DR, will be used for extracting geometric features in order to see whether any significant changes occur prior to the onset of DR. In collaboration with another ESR, vascular trees were segmented, connected and a series of hemodynamic features were calculated alongside the geometric ones, in order to conduct an overall analysis of the changes and the effect that the diabetes/DR have to the retinal vasculature during the last three years of diabetes until the first year of DR.
The last six months have been mainly dedicated to three different aspects. Firstly, finalising the ongoing publications: one paper on retinal autoregulation is now available online in the Journal for Modeling in Ophthalmology and another paper, more focused on the simplified fluid-structure interaction model used in the autoregulation modeling, has been accepted for publication in Comp. Meth. Appl. Mech. Engng. Secondly, the ROM techniques developed in view of their application on a more golbal model of the eye have been implemented and tested. A publication describing such techniques and their results has been submitted. Thirdly, the collaboration with the University of Strasbourg and in particular with Prof. Giovanna Guidoboni has started. The idea is to develop a more global model of the eye to address medical questions regarding the IntraOcular Pressure (IOP).
Netrins are a family of matrix-binding proteins that function as guidance signals. Netrin-4 displays pathologic roles in tumorigenesis and
neovascularization. To answer the question whether netrin-4 acts either pro- or anti-angiogenic, angiogenesis in the retina was assessed in Ntn-4−/− mice with oxygeninduced retinopathy (OIR) and laser-induced choroidal neovascularization (CNV), mimicking hypoxiamediated neovascularization and inflammatory mediated angiogenesis. The basement membrane protein netrin-4 was found to be localised to mature retinal blood vessels. Netrin-4, but not netrin-1 mRNA expression, increased in response to relative hypoxia and recovered to normal levels at the end of blood vessel formation. No changes in the retina were found in normoxic Ntn-4−/− mice. In OIR, Ntn-4−/− mice initially displayed larger avascular areas which recovered faster to revascularization. Ganzfeld electroretinography showed faster recovery of retinal function in Ntn-4−/− mice. Expression of netrin receptors, Unc5H2 and DCC, was found in Müller cells and astrocytes. Laser-induced neovascularization in Nnt-4−/− mice did not differ to that in the controls. Our results indicate a role for netrin-4 as an angiogenesis modulating factor in O2-dependent vascular homeostasis while being less important during normal retinal developmental angiogenesis or during inflammatory neovascularization.
In this reporting period, the risk of progression to the different states of diabetic retinopathy (R0, R1, R2 and R3) was estimated based on data collected from patients attending systematic retinal screening in the region of East Anglia. There was a total of 14,348 patients who were followed up for a time period of 6 years for whom the duration of diabetes, type of diabetes, age and gender were recorded.
The risk could be estimated for the cohort as well as for each individual based on their clinical profile. During the validation stage, the observed number of patients in each state was compared against the expected number of patients according to the risk estimation. There was a good approximation for the first two states and a mediocre approximation for the last two, primarily due to a small number of available data for these two states. The duration of diabetes, the type of diabetes and the gender were found to have a significant effect on the risk of transitioning in some states as can be seen in the picture attached.
For future steps, I wish to enhance the accuracy of the risk estimation by deploying more data from the more advanced states and more predictors.
This reporting period was mostly about collaborations. I was able to work with many of the other ESRs and co-author a few conference papers with them. It also gave me the opportunity to collect some data from the University of Dundee to help create a new tortuosity testing dataset. Once available, the dataset will allow others to publicly download and test new metrics while comparing to multiple clinician gradings. I was also able to create and publish my method for detecting vessels crossover abnormalities. This adds another tool to the retinal vessel analyser, the main topic of my project.
During this period of research, the following was achieved:
- Finalization of experiment protocol on adaptive remodelling of microvascular networks in CAM after acute microocclusion in arterioles and venules.
- In collaboration with ESR 2.4, Febro Guimaraes: Semi-automatic velocity measurement software tested.
- In collatoraction with ESR 2.1, Francesco Caliva, and Dr Bashir Al-Diri: Preliminary test on network reconstruction of CAM microvasculature.
During this research period, Carlos improved his retinal image registration method, which aligns retinal images. This is useful for studying the evolution of disease across examinations, or to enable other techniques such as the creation of super resolution images to perform more accurate measurements or study fluid dynamics on the retinal vessels when the images are taken in the same session.