Dense Convolutional Binary-Tree Networks for Lung Nodule Classification

Abstract

This paper investigates the problem of benign or malignant diagnosis of pulmonary nodule with original thoracic computed tomography images, and presents a novel end-to-end deep learning architecture named dense convolutional binary-tree network (DenseBTNet). Besides introducing center-crop operation into the DenseNet, the DenseBTNet splits isolated transition layers of the DenseNet and merges them with dense blocks, then adjusts feature-maps transition mode to compact the model. The DenseBTNet has several compelling advantages: 1) the DenseBTNet not only preserves densely connected mechanism of the DenseNet to extract features of lung nodules in different level, but also further reinforces this mechanism to a level of dense blocks and enriches multi-scale features and 2) The DenseBTNet owns high parameter-efficiency and is lightweight in the scale of parameters as well. Experimental results show that the DenseBTNet largely boosts the performance of the DenseNet and achieves higher accuracies on the task of lung nodule classification in comparison with state-of-the-art approaches.

Keywords

• Lung nodule classification
• computed tomography (CT) imaging
• densely connected convolutional networks