Computer-Aided Diagnosis of Pulmonary Nodules on CT Scan Images

Abstract

Lung cancer has become the leading cause of cancer death with the increasing morbidity and mortality all over the world. Early detection, early diagnosis, and early treatment are of great significance to increase the survival rate of lung cancer patients. Computer-aided Diagnosis (CAD) based on computed tomography (CT) is an important technique for early diagnosis of lung lesions. This paper proposes a four-step approach for segmentation and classification of pulmonary nodules on CT scan images. A lung parenchyma segmentation method is performed which overcomes the shortcoming of pleural nodule omission of simple thresholding segmentation method. Then, a modified U-net fully convolutional network is constructed to segment suspected pulmonary nodules. Followed that, gray-level, shape, interior morphological, external morphological, texture and spatial features of pulmonary nodules are extracted. Last, an ensemble model of Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) classifiers are employed to classify the suspected pulmonary nodules and the Area Under Receiver Operating Characteristic (ROC) Curve (AUC) is 0.93. Results confirm the feasibility and efficiency of the proposed approach. Taking both segmentation and classification of pulmonary nodules into account, we have thus created a doctor-assisted CAD approach for diagnosis of pulmonary nodules.