NROI based feature learning for automated tumor stage classification of pulmonary lung nodules using deep convolutional neural networks

Abstract

Identifying the exact pulmonary nodule boundaries in computed tomography (CT) images are crucial tasks to computer-aided detection systems (CADx). Segregation of CT images as benign, malignant and non-cancerous is essential for early detection of lung cancers to improve survival rates. In this paper, a methodology for automated tumor stage classification of pulmonary lung nodules is proposed using an end-to-end learning Deep Convolutional Neural Network (DCNN). The images used in the study were acquired from the Lung Image Database Consortium and Infectious Disease Research Institute (LIDC-IDRI) public repository comprising of 1018 cases. Lung CT images with candidate nodules are segmented into a 52 × 52 pixel nodule region of interest (NROI) rectangle based on four radiologists’ annotations and markings with ground truth (GT) values. The approach aims in analyzing and extracting the self-learned salient features from the NROI consisting of differently structured nodules. DCNN are trained with NROI samples and are further classified according to the tumor patterns as non-cancerous, benign or malignant samples. Data augmentation and dropouts are used to avoid overfitting. The algorithm was compared with the state of art methods and traditional hand-crafted features like the statistical, texture and morphological behavior of lung CT images. A consistent improvement in the performance of the DCNN was observed using nodule grouped dataset and the classification accuracy of 97.8%, the specificity of 97.2%, the sensitivity of 97.1%, and area under the receiver operating characteristic curve (AUC) score of 0.9956 was achieved with reduced low false positives.

Keywords

• Deep learning
• Convolutional neural network
• Segmentation
• Data augmentation
• Dropouts