Archives

  • 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • br The Authors Production and

    2020-08-18


    2019 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
    1. Introduction
    Lung cancer in men is the most common cause of death world-wide (WHO, 2017; Siegel et al., 2013; American Cancer Society, 2017). Only 70% of patients diagnosed with lung cancer are detected even when they reach a high stage (Camarlinghi, 2013; Jemal, 2017; Dhara et al., 2012). There are several steps to diagnose a patient who is suspected of lung cancer starting with symptoms of physical examination. Radiological examination is needed when there are no symptoms in a physical examination with a doctor’s recommendation (Metastatic Cancer - National Cancer Institute, 2017). CT scan is one of the most commonly used modalities for
    ⇑ Corresponding author at: Department of Electrical Engineering and Information Technology, Universitas Gadjah Mada, Jalan Grafika 2, Kampus UGM, Yogyakarta 55281, Indonesia.
    E-mail addresses: [email protected] (R. Nurfauzi), [email protected] ac.id (H.A. Nugroho), [email protected] (E.L. Frannita).
    q This work is funded by Directorate General of Higher Education, Ministry of Research, Technology and Higher Education, Republic of Indonesia through the ‘‘Penelitian Tim Pasca Sarjana” Research Grant.
    Peer review under responsibility of King Saud University.
    Production and hosting by Elsevier
    detection of nodules as one of the earliest symptoms in diagnosing lung cancer. Generally, lung CT images are reconstructed in an axial Concanamycin-A of 512 512 with more than a thousand slices depending on the specification of the CT machine. As a result, direct assessment of lung CT images will take longer time before the radiologists determine final diagnosis.
    Manual observation on 3D lung CT images by radiologists may cause some errors because of its tediousness and due to human factors, such as fatigue and difference in experience (Kundel and Revesz, 1976; Berbaum et al., 1990; Renfrew et al., 1992; Petrick et al., 2013; McNitt-Gray et al., 2007). One example of reliance on expertise occurs in a large lung CT database, namely LIDC-IDRI. In the LIDC-IDRI database (Armato et al., 2011; LIDC-IDRI), a diagnostic evaluation called ground truth was provided by four radiologists and only 34.7% of cases had the same diagnostic eval-uation results (McNitt-Gray et al., 2007). Computer-aided detec-tion (CADe) systems aim to overcome this limitation of manual examination and to speed up evaluation time (Doi, 2005). How-ever, a CADe design must follows the rules of the radiologist in rec-ognizing nodules (Valente et al., 2016; Lederlin et al., 2013).
    There are four types of lung nodules based on their position, namely well-circumscribed nodule, juxta-vascular nodule, nodule with pleural tail and juxta-pleural nodule, as shown in the Fig. 1 (Dhara et al., 2012). Juxta-vascular nodule, nodule with pleural tail and juxta-pleural nodule are the most common types of nodules. These nodules have similar intensity with the attached organ in
    This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
    Please cite this article as: R. Nurfauzi, H. A. Nugroho, I. Ardiyanto et al., Autocorrection of lung boundary on 3D CT lung cancer imagesq, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.02.009
    2 R. Nurfauzi et al. / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
    lung wall area. Hence, the lungs with these nodules cannot be seg-mented only by conventional methods. An advanced method is required to separate them from the lung boundary. The presence of this nodule types allows the challenge for many CAD lung devel-opers to obtain the lung area correctly without losing nodule areas (Diciotti et al., 2011).
    The simplest and the most widely used method to overcome this challenge is morphological closing (Javaid et al., 2016; Teramoto and Fujita, 2013; Gupta et al., 2015). However, the accu-racy and computational time depend on the size of the morpholog-ical kernel (Pu et al., 2008). When the kernel size is too small, the under segmentation (US) area will be larger and the computational time will be shorter. On the contrary, when the kernel size is get-ting bigger, the over segmentation (OS) area will be larger and the computational time will be longer. The US is defined as an area of nodules that are not covered by the system. The OS is defined as an over covered area as result of segmentation process by the system.
    In Pu et al. (2008), adaptive border marching (ABM) is proposed to solve the limitation of morphological closing (Javaid et al., 2016; Teramoto and Fujita, 2013; Gupta et al., 2015). The method is sim-ple and very useful. The main task of this method is to connect every border pixel point. The support vector machine (SVM) rules are applied to eliminate the pixel pair connection that not sus-pected as nodule basin. The SVM works by considering the distance of pixel pair connection and the depth of curvature to develop the rules. The limitation of this method is of high computational since the system must predict each pixel pair.