Method based on deep neural network to extract appearance and geometry features for pulmonary textures classification
Provided is a method based on deep neural network to extract appearance and geometry features for pulmonary textures classification, which belongs to the technical fields of medical image processing and computer vision. Taking 217 pulmonary computed tomography images as original data, several groups of datasets are generated through a preprocessing procedure. Each group includes a CT image patch, a corresponding image patch containing geometry information and a ground-truth label. A dual-branch residual network is constructed, including two branches separately takes CT image patches and corresponding image patches containing geometry information as input. Appearance and geometry information of pulmonary textures are learnt by the dual-branch residual network, and then they are fused to achieve high accuracy for pulmonary texture classification. Besides, the proposed network architecture is clear, easy to be constructed and implemented.
1. A method based on deep neural network to extract appearance and geometry features for pulmonary textures classification, wherein comprising the following steps:
1) original data preparation: the original data includes CT image patches of pulmonary textures, corresponding image patches containing geometry information and corresponding category labels, for training, validation, and testing;
2) construction of dual-branch residual network: based on skip-connection structure proposed in residual network, a dual-branch residual network with 18 layers is constructed;
3) train the dual-branch residual network from step 2),
wherein, the construction of the dual-branch residual network in the step 2) comprises the following steps:
2-1) the constructed dual-branch residual network is composed of 18 layers, including 15 convolutional layers and 3 fully connected layers; a top branch uses appearance information in image as input, a bottom branch uses corresponding image patches containing geometry information as input;
2-2) the skip-connection structure calculates the residual error between the input information and the information calculated by convolutional layer, through learning residual error by the network, the residual error is formulated as follows:
y=F ( x i { W i })+ x
in the formulation, ydenotes final output result after convolutional layer, Fdenotes convolutional calculation function, x denotes input of the network, idenotes the index of convolutional layer with domain of definition set as [1,15],W i denotes the weight matrix of i-th convolutional layer obtained from training the network;
2-3) the output of last convolutional layer of the top and bottom branches are directly concatenated after a max-pooling layer to fuse information of top and bottom branches;
2-4) a max-pooling layer is utilized after the first and last convolutional layers; and
2-5) the stride of 6-th, 10-th, and 10-th convolutional layers are set to be 2.