Engineering Applications of Artificial Intelligence 81 (2019) 133–144 Contents lists available at ScienceDirect Engineering Applications of Artificial Intelligence journal homepage: www.elsevier.com/locate/engappai K-nearest neighbor driving active contours to delineate biological tumor volumes ✩ Albert Comelli a,b,c , Alessandro Stefano b,∗ , Giorgio Russo b,d , Samuel Bignardi a , Maria Gabriella Sabini d , Giovanni Petrucci c , Massimo Ippolito e , Anthony Yezzi a a Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA b Institute of Molecular Bioimaging and Physiology, National Research Council (IBFM-CNR), Cefalù (PA), Italy c Department of Industrial and Digital Innovation (DIID) – University of Palermo (PA), Italy d Medical Physics Unit, Cannizzaro Hospital, Catania, Italy e Nuclear Medicine Department, Cannizzaro Hospital, Catania, Italy ARTICLE INFO Keywords: Cancer segmentation Active contour algorithm K-nearest neighbor FDG and MET PET imaging Biological target volume ABSTRACT An algorithm for tumor delineation in positron emission tomography (PET) is presented. Segmentation is achieved by a local active contour algorithm, integrated and optimized with the k-nearest neighbor (KNN) classification method, which takes advantage of the stratified k-fold cross-validation strategy. The proposed approach is evaluated considering the delineation of cancers located in different body districts (i.e. brain, head and neck, and lung), and considering different PET radioactive tracers. Data are pre-processed in order to be expressed in terms of standardized uptake value, the most widely used PET quantification index. The algorithm uses an initial, operator selected region containing the lesion, and automatically identifies an operator-independent optimal region of interest around the tumor. Successively, a slice-by-slice marching local active contour segmentation algorithm is used. The key novelty of the proposed approach consists of a novel form of the energy to be minimized during segmentation, which is enhanced by incorporating the information provided by a KNN classifier. The delineation process and its termination are fully automatic, so that intervention from the user is reduced to a minimum. Due to the high level of automation, the final segmented lesion is independent of inter-operator variation in the initial user input, making the entire process robust and the result completely repeatable. In order to assess the performance under different contrast ratio scenarios, we first evaluate the proposed method on five phantom datasets. Next, we assess the applicability of the method in the radiotherapy environment by investigating fifty clinical cases and two different PET radio-tracers. Our investigation shows that the proposed method can be applied in clinical settings and produces accurate and operator-independent segmentations, attaining good accuracy in realistic conditions. 1. Introduction In radiation therapy planning, Magnetic Resonance Imaging (MRI) provides high quality detailed images and excellent soft-tissue con- trast, while Computerized Tomography (CT) images provides attenu- ation maps. In this context, Positron Emission Tomography (PET) is a non-invasive imaging technique which has the advantage, over mor- phological imaging techniques, of providing functional and metabolic information about the patient’s condition. Inclusion of PET images in radiotherapy protocols is desirable because it provides crucial in- formation to accurately target the oncological lesion and to escalate the radiation dose without increasing normal tissue injury (Newbold ✩ No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.engappai.2019.02.005. ∗ Corresponding author. E-mail address: alessandro.stefano@ibfm.cnr.it (A. Stefano). et al., 2008). For this reason, PET may be used for improving the Planning Treatment Volume (PTV) (Niyazi et al., 2013). In addition, PET quantitative assessment in oncological patients, conveys functional information which is predictive of treatment response (Alongi et al., 2016) and faster changing than the morphological response (Wahl et al., 2009). As such, PET can greatly improve the clinical cancer treatment decision making (Allegra et al., 2017). In particular, the maximum Standardized Uptake Value (SUV max ) is the most widely used quantification parameter giving a punctual measure of cellular metabolism (Fletcher and Kinahan, 2010). To provide further informa- tion about the cancer, additional quantitative parameters have been introduced, such as biological tumor volume (BTV) and tumor lesion https://doi.org/10.1016/j.engappai.2019.02.005 Received 10 May 2018; Received in revised form 7 December 2018; Accepted 5 February 2019 Available online xxxx 0952-1976/© 2019 Elsevier Ltd. All rights reserved.