Automatic pulmonary nodule detection from radiography using histograms of oriented gradients descriptors

Abstract

A chest X-ray examination is a painless, non-invasive, and cost effective medical examination performed at present day. A pulmonary nodule is a small round lesion or mass in the lungs which can be indicative of an infection or a neoplasm. Chest X-rays can be used to diagnose pulmonary nodules. State-of-the-art automatic pulmonary nodule detection techniques are agonized by the problems posed by noise, local-global feature dilemma, and the bias-and-variance dilemma. To evade these problems, this project proposes a three-layered framework to perform automatic diagnosis of pulmonary nodules. The first layer performs hybrid Haar-wavelet based image enhancement and contour-based lung field segmentation. The second layer extracts histogram of oriented gradient descriptors from a pre-processed X-ray image and compresses the high-dimensional descriptors onto a low dimensional manifold using codec manifold neural network. Finally, the third layer classifies whether the X-ray contains any signs of nodules using an ensemble of partial decision trees. Experiments have been carried out on three X-ray datasets. The proposed system was found to outperform the state-of-the-art systems The results demonstrate the efficacy of the proposed nodule detection framework. The proposed pulmonary nodule detection can be integrated with the existing X-ray equipment in hospitals in order to perform rapid diagnosis.