Abstract

This article describes a novel ensemble approach to achieve high quality segmentation of Chest X-Ray (CXR) images. Specifically, the Japan Society of Radiological Technology (JSRT) dataset consisting of 247 CXR images has been used to study the effectiveness of this approach. The study has been carried out in three phases: preprocessing, segmentation, and validation by classification. The novelty of the study comes from combining a Denoising Autoencoder with Contrast Limited Adaptive Histogram Equalization (CLAHE) for preprocessing; an ensemble method that combines Canny edge detection with morphology, Otsu thresholding, automatic thresholding, and granule segmentation for segmentation; and the Gray-Level Co-occurrence Matrix (GLCM) for feature extraction to aid in classification. The quality of classification of lung CXR images as nodules vs. non-nodules, and malignant vs. benign, using this unique combination of various standard approaches for segmentation coupled with a variety of classifiers such as Random Forest, Decision Tree, Naive Bayes, Support Vector Machine (SVM), Logistic regression, K-Nearest Neighbor (KNN), and XGBoost has been validated with Receiver Operating Characteristic (ROC) curves. This study attains high accuracy of 97.67% for enhanced ensemble segmentation and accuracy of 98.93% for nodule/nonnodule classification and 99.18% accuracy for malignant/benign classification with SVM.

Authors

V. Thamilarasi¹, A. Asaithambi², R. Roselin³
Sri Sarada College for Women, India^1,3, Florida Polytechnic University, United States of America²

Keywords

Medical Image Analysis, Chest X-Ray Images, Segmentation, Classification, Nodule, Non-Nodule, Thresholding, Granule