A Lung Cancer Detection Approach Based on Shape Index and Curvedness Superpixel Candidate Selection
Author: Jeovane Honório Alves (Lattes)
E-mail: jhalves@inf.ufpr.br, jeohalves@gmail.com, j.h.alves@ieee.org
Supervisor: Lucas Ferrari de Oliveira
Abstract
Cancer is one of the causes with more mortality worldwide. Lung cancer is the most common type (excluding non-melanoma skin cancer). Its symptoms appear mostly in advanced stages, which difficult its treatment. For patient diagnostic, computer tomography (CT) is used. CT is composed of many slices, which maps a 3D region of interest. Although it provides many details, its analysis is very exhaustive, which may has negatively influence in the specialist’s diagnostic. The objective of this work is the development of lung segmentation and nodule detection methods in chest CT images. These images are segmented to separate the lung region from other parts and, after that, nodule detection using superpixel methods is applied. The Axes’ Labeling had a mean of nodule preservation of 93.53% and the Monotone Chain Convex Hull method presented better results, with a mean of 97.78%. For nodule detection, the Felzenszwalb and SLIC methods are employed to group nodule regions. A nodule candidate selection based on shape index and curvedness is applied for superpixel reduction. Then, classification of these candidates is realized by the Random Forest. The LIDC database was divided into two data sets: a development data set composed of the CT scans of patients 0001 to 0600, and a untouched, validation data set, composed of patients 0601 to 1012. For the validation data set, the Felzenszwalb method had a sensitivity of 60.61% and 7.2 FP/scan.
Files
Dissertation: (PDF – in english)