Automatic Cyst Detection in OCT Retinal Images Combining Region Flooding and Texture Analysis A. Gonz´ alez, B. Remeseiro, M. Ortega, M.G.Penedo VARPA Group, Department of Computer Science Universidade da Coru˜ na, Spain {ana.gonzalez, bremeseiro, mortega, mgpenedo}@udc.es P. Charl´ on Institute of Ophthalmology G ´ omez-Ulla Santiago de Compostela, Spain pablocharlon@institutogomez-ulla.es Abstract In this work Optical Coherence Tomography (OCT) retinal images are automatically processed to detect the presence of cysts. The methodology is composed by three phases: region of interest where cysts will be searched is delimited; a watershed algorithm is applied to find all the possible regions in the image which might conform cystic structures; finally, texture analysis is performed in each region from previous phase to final classification. Results show that accuracy achieved with this method is over 80%. 1. Introduction OCT retinal images are used by experts to diagnose di- seases, since their capture consists in a non-invasive, con- tactless method which gives a cross sectional image of the retina and its structures in a real time fashion[6]. Several diseases can be diagnosed nowadays with an OCT retinal analysis: OCT layer-thickness informa- tion is useful in diagnosing eye diseases like diabetic retinopathy[7]. The presence of cystic structures may de- note inflammation and diabetes[8]. The main problem currently is the high amount of information captured in these kind of images, which experts need to process. Therefore, several image processing-based applications are emerging with the purpose of expert assessment. Although some image processing techniques have been developed to detect anomalies in the macula[5] in an auto- matic way, there is not previous work establishing a global quantitative method for automatic cyst detection, with in- dependence of the kind of cyst, so this is the aim of this work. Cyst detection task is not immediate, given the diffe- rent problems present in OCT retinal images. Firstly, the presence of different pathologies in the same patient is pos- sible, resulting in several alterations on the retinal image. Some parts of the image may have lower quality due to the capture process. Besides the high variability in shape, size and orientation in the cyst, they can be placed on different layers of the retina, whose intensity properties are not simi- lar. There are also structures on the image with similar pro- perties, such as vessel shades. In addition, there are situa- tions in which the expert can not determine with a sufficient confidence level if a suspicious region is a cyst or not. This work proposes a methodology for an automatic detection of candidate regions to be cyst, in OCT retinal images. It is a novelty in the automatic assessment to ex- perts, as well as it provides an objective and repeatable method that allows the expert to improve his efficiency rates. In this case, HD-OCT images have been considered. This paper is organized as follows: in Section 2 the me- thodology for automatic cyst detection is exposed. Section 3 shows the results obtained with this technique and in Sec- tion 4 conclusions and future lines of work are presented. 2. Methodology The detection of cyst candidate regions has some com- plexity involved. Our proposal for this task is composed by different phases (Figure 1): a preprocessing to delimit the region of interest where cyst location will be performed; the detection of candidate regions using a watershed algorithm; finally, a classification process over texture descriptors ex- tracted from the regions to identify real cysts. The necessity of these phases is reflected in Figure 2, in which the region of interest and the cysts on the image have been marked. 2.1. Preprocessing The region of interest must be delimited. It is determined by the bounding layers of the retina, so the top part of the Inner Limiting Membrane (ILM) , as well as the bottom part of Retinal Pigment Epithelium (RPE), are segmented based on transforming the segmentation task into that of finding a minimum-cost closed set in a geometric graph[3]. In [2] the designing of the cost functions to segment these layers is presented. Figure 3 (a) shows the segmentation obtained. 978-1-4799-1053-3/13/$31.00 c 2013 IEEE CBMS 2013 397