International Conference on Engineering Technologies (ICENTE'19) Konya, Turkey, October 25-27, 2019 Segmentation of Capillaroscopic Images K. TUTUNCU1 and M. BUBER1 1 Selcuk University, Konya/Turkey, ktutuncu@selcuk. edu.tr 1 Selcuk U niversity, Konya/Turkey, mbuber@selcuk.edu.tr Abstract- Capillaroscopy device shoot videos of capillaries of oral mucosa and nailfold of patient over the related skin without any pain. The image frames of videos are used by experts for early detection or treatment of some diseases such as diabetics, rheumatism and etc. Since this process is implemented in manually, decision support systems that helps the experts for diagnosis have been subjects of studies of biomedical researches. First step of these systems is the successful segmentation process on these images that will be used for classification of disease depending on 8 parameters such as the number of capillaries in a certain area, the distance between the two vessels, the size of the capillaries and etc. This study aims to contribute decision support system for experts by presenting a successful segmentation. In this study Otsu, Fuzzy C-mean, Fast Marching, Region Growing and H-Minima methods have been used for segmentation of capillarocopic images. The segmentation accuracy ratios of upper mentioned methods were obtainedas %80,47, %67,44, %63,23, %44,11 and 96.76%, respectively. When the results were examined, it was observed that the H-Minima method, which had not previously been applied in capillary images, reached the highest accuracy parameter value. Keywords: Image Processing, Capillaroscopy, Capillary Video, Capillary, Segmentation, H-Minima Method. I. INTRODUCTION Image segmentation is a very challenging problem that needs to be specifically designed according to the application area. The success of algorithms developed for this purpose also depends largely on the correct identification of similarity criteria used for zone homogeneity. Image segmentation techniques are often classified as threshing-based, edge determination, region-based, and clustering-based techniques [1-3]. Image segmentation, which is used in many fields, is also heavily used in the medical field and is of great importance in the diagnosis and treatment of many diseases [4]. The tests performed with a capillaroscopy device are one of these samples. Capillaroscopy is the process of taking pictures or video images of capillaries with the help of a device. With these images, capillaries are examined and changes in veins can be a precursor to some diseases. In particular, nailfold capillaroscopy makes it easy to diagnose rheumatological and diabetic diseases [5]. II. LITERATURE REVIEW Tama et al. (2015) conducted a study to measure morphological parameters for each capillary by image processing. Morphological parameters were determined by preprocessing, dualization, skeletal (skeleton) inference and skeletal (skeleton) segmentation. These parameters are the width of the capillary, the height of the capillary, the morphology of the vascular nodes, the width of the terminal nodes, the calibre and the curvature. With the application they developed, up to 96% sensitivity was achieved in the automatic measurement of all parameters [6]. Vucic (2015) in his thesis work used the support vector machine method as an image processing technique, resulting in an accuracy parameter of 92% [7]. Bellavia et al. (2014) prefixed images with wavelet analysis and mathematical morphology. Later they applied segmentation in order to minimize in-class lighting value differences of capillary and background images. Mean sharpness, mean recall and Jaccard index were calculated as 0.924, 0.923 and 0.858, respectively [8]. Isgro et al. (2013) also performed segmentation on sequential videos of capillaries at the tip of the nail. First, rough segmentation was applied, then a series of automatic thresholds were passed and a segmentation map was drawn with the STAPLE algorithm. The accuracy, sensitivity and authenticity parameters compared to manually segmented images were obtained as 97%, 96% and 98% respectively [9]. Goffredo et al. (2012) developed a new method for parsing the color field in digital nailfold bottom capillaroscopy analysis. The threshold values of the image were determined using Otsu technique and 7x7 median filter. In nailfold capillary images, segmentation is indicated to have the highest performance predictably in the green channel. As a result, values of 75~87%, 85~90% and >80% were found for sensitivity, authenticity and accuracy parameters, respectively [10]. Kwasnicka et al. (2007) presented a preliminary study of the use of automatic note insertion methods for computer aided diagnosis and analysis of capillary images. In summary, capillary images were taken, their features were removed, these features were processed with a predefined note splitter and they obtained an automatic note spliced capillary image. As a result of the study, five separate images were obtained. A total of six images were created with the first image passed through the Gaussian filter. These images were applied grid segmentation to create all regions together with the segmentation. After that, automatic note insertion system was applied to images using multiple class machine learning, balanced average and continuous relevance model. As a result of this application, the accuracy parameter value was obtained as 77% [11]. Riano-Rojas et al. (2007) conducted a study to infer E-ISBN: 978-605-68537-9-1 312