Lung Boundary Detection and Classification in Chest X-Rays Images Based on Neural Network

Abstract

The isolation of different structures is often performed on chest radiography (CXR) and the classification of abnormalities is an initial step in detection systems as computer-aided diagnosis (CAD). The shape and size of lungs may hold clues to serious diseases such as pneumothorax, pneumoconiosis and even emphysema. More than 500,000 people die in the United States every year due to heart and lung failure, often being tested for the normal CXR film. With an increasing number of patients, the doctors must over-work, hence they cannot provide the advice and take care of their patients correctly. In this case, the computer system that supports image classification and boundary CXR detection is needed. This paper presents our automated approach for lung boundary detection and CXR classification in conventional poster anterior chest radiographs. We first extract the lung region, size measurements, and shape irregularities using segmentation techniques that are used in image processing on chest radiographs. For the CXR image, we extract 18 various features using the gray level co-occurrence matrix (GLCM) which enables the CXR to be classified as normal or abnormal using the probabilistic neural network (PNN) classifier. We measure the performance of our system using two data sets: the Montgomery County (MC) x-ray dataset and the Shenzhen X-ray dataset. The proposed methodology has competitive results with relatively shorter training time and higher accuracy.