Using 3D Convolutional Networks with Shortcut Connections for Improved Lung Nodules Classification

Lung cancer is the leading cause of cancer-related death worldwide. Due to the difficulty of artificial extraction of medical image features and the development of artificial intelligence in the field of medical image, various deep learning methods for lung nodules classification have been proposed to help doctors diagnose and detect lung cancer in the early stage. The traditional 2D CNN cannot make use of the 3D spatial characteristics of CT data, while the 3D CNN has many parameters, which leads to low model efficiency. Residual Network (ResNet) is a residual structure using skip connection that makes deep classification network easier to train. Therefore, motivated by the work of 3D CNN and ResNet, in this paper, a VGG based 3D residual connection network, called VGG+ResCon, is proposed to mine the vertical information of tumor CT images and accelerate the training efficiency of the model. Besides, after enhancing the dataset, focal loss is used to replace the traditional cross-entropy as the loss function, which solves the problem of uneven distribution of positive and negative samples of medical data (more negative samples than positive samples). And it also makes the model more focused on difficult-to-classify samples. This methodology was evaluated on the LUng Nodule Analysis 2016 (LUNA16) dataset, with the best precision of 93.62%, recall of 92.48%, specificity of 96.83% and f1-score of 93.04%. Experimental results demonstrate the effectiveness of the proposed method in classifying malignant and benign pulmonary nodules.