Seite - 160 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 1: Three examples (scans 1, 2, and 5) of ground truth (left) and incorrect algorithmic (right) segmenta- tionsof the first retina layer (L1) inmid-stackslices. 3.Method In order to identify incorrectly segmented OCT scanswesuggest insection3.1toembedthesegmen- tationresults inasfewdimensionsaspossible. While this iscertainlymotivatedbycurseofdimensionality it is additionally motivated by an increase of inter- pretability–ophthalmologistsmaydesire tovisually relatea particular case tocases inspectedpreviously. Methods of outlier detection can be divided in three branches [6]. Supervised classification, when both inliers and outliers are labeled and in balance; unsupervised when training data of both inliers and outliers are unlabeled; and semi-supervised when training data consists only of observations describ- ing normal behavior. In section 3.2 we follow semi- supervised methods for the following reasons. First, there is no assessment of algorithmic segmentations available and we only can roughly estimate the class based on some metric (e.g., the Dice coefficient). Second, the outlier class (wrong segmentations) is expected to be under-represented. Third, it is likely thereareseveralsourcesofsegmentationerrorwhich could map to low-density clusters. We aim to detect outliers in low-density regions, too. We model the distribution of the inliers (correct segmentation) and compare the testpoints to thisdistribution. 3.1.Area curvesandtheirrepresentation Whileforeachretinal layera listof regionproper- ties can be thoughtof, for sakeof interpretability the slice-wiseareavaluesareofspecial interest. Further- more, thefocusof thisworkwasrestricted to layer1. This decision is based on the observation that a seg- mentationerror inL1 layerpropagates tosubsequent layers while correct L1 segmentations tend to corre- latewithcorrectly segmentedscans. ForeachsegmentedOCT,we introduce thevector a= [a0, .. . ,a199] > of layer-1 area values and refer to is as the area curve. Examples of how area curves look like for both ground truth and algorithmic seg- mentationsaregiven infigure2. Looking at the (orange) area curves calculated from the algorithmic segmentation, which are of the main interest, two types of shape appear: Those ex- hibiting a maximum (cf., scan 1, 2 or 5 of figure 2), or a minimum (cf. scan 0) around the middle of the slices. In healthy eyes, the layers get thinner around the cavity of the fovea [11], causing the area curves to exhibitaglobalminimumandtendtobeconvex. The first hypothesis about the curves with dominant con- cave bumps therefore was that they may correspond to pathologies where fluid intruded into the retinal layers andcaused themto thicken. Closer inspection of the corresponding scans and a comparison to the (blue) GT area curves, however, quickly disproved this hypothesis and revealed that the concave bumps tend to correspond to failures in segmentation. Further investigations revealed that the issue of a too thick segmented layer 1 appeared inall scans thatexhibitaglobalmaximuminthearea curveor tenddobeconcave. 160