Pulmonary Nodule Segmentation in Lung CT Images by Post Processing U-Nets Using Average Ensemble Learning Technique

No Thumbnail Available

Date

2020-06-24

Journal Title

Journal ISSN

Volume Title

Publisher

Addis Ababa University

Abstract

Pulmonary nodules are potential manifestation of lung cancer and accurate segmentation of different pulmonary nodules from lung Computed Tomography (CT) scan images is important clinical relevance in diagnosis, prognosis and treatment of lung cancer. However, due to the highly heterogeneous type, size, location and shape of nodules, segmentation of pulmonary nodules is very challenging. In this study we propose and present an improved pulmonary nodule segmentation method based on thresholding and morphological operators, lungs region segmentation algorithm, modified U-Net model for pulmonary nodules detection and segmentation, and average ensemble learning as a post processing technique. First, as a preprocessing stage that includes lungs region segmentation, normalization, cropping and resizing has done on the raw input CT scan images. Thresholding and morphological operators algorithm is simple and yields good result in accurately segmenting the lung parenchyma to reduce the search space. Then modified U-Net model is loaded to detect and segment pulmonary nodules. U-Net is a widely used Convolutional Neural Network (CNN) structure for end-to-end segmentation and can be used on the entire image classification labels over the entire input pixels so that it is more efficient and expected to yield better performance. Furthermore, instead of picking the best U-Net network structure, we applied average ensemble learning method as a post processing technique. An ensemble of three U-Net models having different network structure, trained on the same dataset with different hyper-parameters, can generally improve the overall segmentation performance. The performance of our proposed method is trained, tested, and evaluated using 858 lung CT images and their corresponding ground truth nodule masks obtained from Lung Nodule Analysis 2016 (LUNA16) dataset and achieved evaluation results of 0.848 mean Dice Similarity Coefficient (DSC), 0.965 mean accuracy, 0.826 mean sensitivity, and 0.983 mean specificity. Moreover, we compared our test results with other methods results to show our approach’s performance. Experiments and these preliminary results showed that our proposed method can effectively improve the segmentation accuracy of pulmonary nodules and the effectiveness of our approach. Generally, the algorithms used here are simple, effective and practical.

Description

Keywords

CT, LUNA16, thresholding and morphological operators, pulmonary nodules, U-Net, average ensemble learning, DSC

Citation