CELL RECONSTRUCTION UNDER VORONOI AND ENCLOSING ELLIPSES FROM 3D MICROSCOPY Geraldo L. B. Ramalho 1 , Daniel S. Ferreira 1 , Andrea G. C. Bianchi 2 , Claudia M. Carneiro 2 , Fátima N. S. Medeiros 3 , Daniela M. Ushizima 4,5 gramalho@ifce.edu.br, daniels@ifce.edu.br, andrea@iceb.ufop.br, carneiro@ef.ufop.br, fsombra@ufc.br, dushizima@lbl.gov, 1 Federal Institute of Education, Science and Technology of Ceará, Maracanaú, Brazil 2 Federal University of Ouro Preto, Ouro Preto, MG, Brazil 3 Federal University of Ceará, Fortaleza, CE, Brazil 4 Berkeley Institute of Data Sciences, University of California, Berkeley, CA, USA 5 Lawrence Berkeley National Laboratory, Berkeley, CA, USA ABSTRACT Cell occlusion, staining variation, particulate forms and diversity of cervical cells are some of the challenges in automating cervical cytology. This paper tackles some of these issues, including the detection of nucleus and cytoplasm from a new standardization for specimen preparation through mono/thin-layer technology. Our approach consists of three main steps: (a) rough segmentation of subcellular compartments using super pixel combined to Voronoi diagrams, (b) structural refinement of the cytoplasm boundary through calculus of variations, and (c) morphological reconstruction combined to optimization methods to determine minimum enclosing ellipse. We test our implementation on real 3D cervical cell images, containing several cells at different occlusion levels and variable contrast. Our results show both qualitative and quantitative assessment of the datasets, using a completely automated computer program. The quantitative performance presents average Dice Coefficient higher than 87%. Index Terms— Cervix cytology, Voronoi, Segmentation, Occlusion 1. INTRODUCTION Cervical cancer is the fourth most common cancer among women, with about 527,000 new cases each year in the world, and nearly 80% of cases take place in low-income countries. This cancer was responsible for the deaths of 265,000 women in 2012, and 87% of these deaths occurred in developing countries [1, 2]. In Brazil, it is expected that 15.590 new cases of cervical cancer will occur every year, with an associated risk of about 15.33 cases per 100,000 women [3]. With the exception of skin cancer, this tumor has the greatest potential for prevention and cure when diagnosed early because it has a pre-cancerous condition that can be recognized and treated early. However, its incidence will continue to increase, especially in developing countries if preventive measures are not broadly applied. The examination of conventional Pap (Papanicolaou) smears is the main strategy of screening programs worldwide. However, the Pap test is based on human visual analysis, a procedure that is not accurate enough to ensure prevention. Therefore, the development of computational tools to ensure greater consistency in these analyzes and reduction mainly of false negative results is very important to ensure the quality of the Pap smear. In order to address the digital images from Pap smears, this paper focuses on boundary extraction of individual cytoplasm and nuclei from overlapping cervical cytology images acquired at different focal planes (FOV). Our method differs from other approaches in the literature [4, 5] because it processes not only the extended-depth-of- field (EDF) image representation but also the FOV images to search for borders of overlapped cytoplasms. We combine intertwined algorithms to deliver an efficient pipeline for automated nucleus and cytoplasm segmentation: superpixel combined to Voronoi Diagrams or SPVD [5], followed by algorithms using calculus of variations to construct edge maps, processed using mathematical morphology methods, such as reconstruction allied to ellipse detectors and finally deliver a refined cytoplasm boundary.