Classification network architecture

ResNet50 is a classification model obtained by improving VGG19 [27] based on the residual learning mechanism. It retains the convolutional layer with a kernel size of 7 × 7 in VGG19 to learn more spatial information, and uses the maximum pooling layer for down-sampling. ResNet-50 has more layers and can learn deeper features. Because of the small size and rich spatial information of lung nodules, ResNet-50 is improved to obtain a classification network suitable for lung nodule diagnosis, which is named 3D ResNet50. Figure 5 shows the architecture of 3D ResNet50. The improvement methods are as follows.

Change the 2D network to the 3D network. The 3D morphological features of lung nodules have an important influence on its degree of malignancy. Moreover, for a single lung nodule, it is a challenging task to find the key slice that represents its malignancy.

Reduce the kernel size in the first convolutional layer and the last 2 residual blocks. Lung nodules are small and its edge shape is an important indicator for its diagnosis. In the calculation process, large convolution kernel introduces many padding voxels at the edge, which not only leads to the inefficient utilization of the edge voxels of lung nodules, but also increases the computational cost.

Abandon pooling layer and reduce stride of convolutional layers [28]. Most of lung nodules are small, abandoning the pooling layer ensures that network contains enough feature information.

The architecture of 3D ResNet50

In this study, negative log likelihood (NLL) loss was used for 3D ResNet50 to measure the difference between the output array and the one-hot vector of the label, which is defined as

where P_n and T_n represent the output array and the one-hot label respectively.