Detecting Early Choroidal Changes Using Piecewise Rigid Image Registration and Eye-Shape Adherent Regularization

Choroidal and retinal thickness changes can occur in patients with refractive errors (e.g. myopia)
or ocular diseases (e.g. central serous chorioretinopathy, glaucoma, etc.) and must therefore be
detected as early as possible and monitored. Image acquisition is usually done with the help of
optical coherence tomography (OCT), which allows 2- and 3-dimensional images with micrometer
resolution. Segmentation-based image analysis methods are used to detect and quantify thickness
changes. However, segmenting the choroid is often a challenging task because of low contrast,
loss of signal and the presence of artifacts in the acquired images. In particular, in vivo imaging
of the choroid-sclera interface (CSI), the border separating the choroid from the sclera, is prone
to these image degradations.
In this thesis, we present CRAR, a novel method for the early detection of choroidal changes
based on piecewise rigid image registration. CRAR allows elastic modeling of the relatively soft
choroid without affecting the more rigid properties of the surrounding sclera and retina. Rather
than insisting on finding the exact position of the CSI, we focus on the changes of the entire
choroid-sclera border. This enables us to circumvent the aforementioned difficulties because,
using this approach, an exact recognition of the choroid-sclera boundary is not required.
In this approach, we focus on juvenile myopia (also called “school myopia”), which in Asian
regions, and especially in China, has reached almost epidemic dimensions by now. Since juvenile
myopia correlates with changes in the thickness of the choroid, but not with its structure as
such, we restrict the transformation model to the anterior-posterior (z-) direction. The proposed
regularization allows respecting the eye's natural shape. In this context, the local homogeneity
of the transformations in nasal-temporal (x-) and superior-inferior (y-) direction are boosted by
penalizing their radial differences.
However, a comprehensive evaluation of the performance in detecting such changes is challenging,
as a ground truth for comparison with the in vivo situation does not exist. In order to
overcome this limitation, we present a statistical validation framework for automated choroidal
thickness changes detection, in which a method purely based on the common agreement between
the algorithm and all experts is combined with an exhaustive power analysis approach. We show
the strengths of the framework with the example of CRAR: the framework demonstrates if an
algorithm functions at an expert level, while the integrated power analysis allows concluding
whether the algorithm performs even better than the experts.
We further applied CRAR to macular telangiectasia type 2 (MacTel2). The analysis of follow-up
images of this disease suggests that there might be a correlation between changes in the
choroidal thickness and the further development of MacTel2. The further refinement of the
presented method CRAR can provide an objective and sensitive tool to analyze and monitor the
progress of myopia, and beyond.