Dense Convolutional Network and Its Application in Medical Image Analysis

Figure 1

DenseNet structure. (a) The basic structure of DenseNet consists of dense block, transition layer, convolutional layer, and fully connected layer. (b) Denseblock consists of densely connected dense units with nonlinear mapping functions of BN, ReLU, and Conv, which are designed with preactivation strategy to make network training easier and generalization performance better. Dense unit input is spliced and merged with all outputs of the previous dense units, and new features generated also need to be passed to subsequent dense units, so that shallow features of dense block are repeatedly reused and effectively utilized, which can alleviate gradient disappearance to a certain extent, and a large number of features can be generated with a small number of convolution kernels; final DenseNet model is relaticely in scale. (c) Transition layer is the structure between adjacent dense blocks, which consists of convolution and average pooling layer, compressing dense block input and all extracted feature information, reducing feature map size and dimensionality, which can effectively reduce the number of dense block parameters and prevent network from overfitting. The fully connected layer is classification prediction layer, which reduces the influence of feature location on classification by integrating category feature information in network features, and classifies feature information after weighting.

(a)

(b)

(c)