Classification of lung nodules in CT images based upon a multiplane dense inception network

Lung cancer has been one of the leading causes of death in the world for decades. An effective computer-aided diagnosis (CAD) scheme for lung nodule analysis is critical in early detection of cancerous nodules.

This work is dedicated to the development of a CAD system based upon deep learning to predict the likelihood of nodule malignancy in lung computed tomography (CT) images.

Because lung nodules exhibit various sizes and shapes, handcrafted texture feature maps are associated with intensity CT images for network input. Ten selected texture features computed from the local binary pattern (LBP), gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), and gray-level size zone matrix (GLSZM) methods are exploited as the concatenation candidates. A new lung nodule classification framework based upon a multiplane dense inception network (MPDINet) is investigated. The proposed model takes advantage of DenseNet for feature condensation and GoogLeNet for feature extraction. Three parallel branches in the axial, coronal, and sagittal planes, including the perinodular zone reinforce nodule characterization while maintaining computation efficacy.

Our MPDINet was evaluated on the public Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) challenge dataset, where 1235 (802 benign and 433 malignant) nodules were selected, based upon 10-fold cross validation. The proposed model with the inverse difference moment (IDM) feature concatenation input achieved high AUC (0.9821 ± 0.0234), sensitivity (0.9426 ± 0.0979), specificity (0.9732 ± 0.0363), and precision (0.9499 ± 0.0657) rates, which demonstrated accurate lung nodule classification.

The developed MPDINet architecture with the handcrafted feature concatenation input is promising in many lung nodule classification applications with CT images.