Seite - 159 - in Joint Austrian Computer Vision and Robotics Workshop 2020

TowardsIdentificationof IncorrectlySegmentedOCTScans VerenaRennerandJirˇı´ Hladu˚vka PatternRecognitionand ImageProcessingGroup TUWien e1527272@student.tuwien.ac.at jiri@prip.tuwien.ac.at Abstract. Precise thickness measurements of retinal layers are crucial to decide whether the subject re- quires subsequent treatment. As optical coherence tomography (OCT) is becoming a standard imaging method in hospitals, the amount of retinal scans in- creases rapidly, automated segmentation algorithms are getting deployed, and methods to assess their performanceare indemand. In this work we propose a semi-supervised frame- work to detect incorrectly segmented OCT retina scans: ground-truth segmentations are (1) embed- ded in2Dfeature spaceand(2)used to trainanout- lier scoring function and the corresponding decision boundary. We evaluate a selection of five outlier detection methods and find the results to be a promising start- ing point to address the given problem. While this work and results are centred around one concrete segmentationalgorithmwesketch thepossibilitiesof how the framework can be generalized for more re- centormore precise segmentationmethods. 1. Introduction It is known that frequent eye screening helps to early-diagnose the diabetic macular edema (DME) [14] and therefore raises the effectiveness of needed treatments. Additionally, the number of age-related macular degeneration (AMD) patients is increasing, because of ageing population [9], as well as those suffering fromDMEdue to the risingnumberofdia- betescases. OCTtechnologyisnowadaysminimally invasive, very fast, and therefore widely spread, so that a large number of OCT scans needs to be pre- processed automatically. Ophthalmological depart- ments are developing or deploying systems to deal with the large amount of OCT data produced. One such instance to segment retinal layers from OCT scans is based on the work [5]. While accurate in most of cases, the method occasionally exhibits im- perfections. Animprovement isdesirable,as thecor- rect segmentation is essential for further automatic evaluation of OCT scans. This is because the thick- nessof the retinal layers ishighly related to thepres- ence of diseases, like AMD or DME [5]. They are caused by intraretinal and subretinal fluids, leading to a swelling of the retinal layers [10], exerting pres- sure on the light-receptors damaging them and thus eyesight. Imperfections in segmentation can be caused by different reasons such as bad contrast of parts of the scan, noise, artefacts or an unsupported edge-case of the segmentationalgorithm. This work aims to support the identification of in- correctly segmented OCT scans with a two-fold pur- pose in mind. First, it is of interest to increase the trust of ophthalmologists in the algorithm by flag- gingsegmentationsthatmaypotentiallyrequireman- ual inspection. Second, to improve segmentation al- gorithms, it is desirable to automatically identify in- correctsegmentationsofpreviouslyunseenscansand focuson improvements for suchcases. 2.Dataset A set of 100 OCT scans, each accompanied with both manual ground truth (GT) and algorithmic (A) segmentation [5] have been provided for this study. Each OCT scan is a stack of 200 1024×200 gray scale images. Both theground truthand thealgorith- mic segmentation are available as slice-wise bound- aries of 13 retina layers. Figure 1 shows boundary examples of the first retina layer (L1). There is no expert assessment available on whether the algorith- mic segmentationsareacceptedascorrectornot. For legal issues, this dataset is currently unavail- able forpublicuse. 159