Automatic pancreas ct segmentation method based on a saliency-aware densely connected dilated convolutional neural network

A pancreas CT automatic segmentation method based on a saliency-aware densely connected dilated convolutional neural network, comprising the following steps of:

• (1) preprocessing of training set, comprising the following steps of:
• collecting CT volume data and making a standard pancreas segmentation result of the data;
• denoting 3D CT volume data as X, and slice number of the volume data as L, a corresponding standard segmentation being Y=(yj,j=1, . . . , |X|),yj={0,1}, where |X| represents a number of all voxels in X, yj=1 or yj=0 represents that voxel j belongs to the pancreas or a background, respectively;
• Slicing each volume X into two-dimensional image slices alone axial view;
• and
• combining three consecutive images into a three-channel pseudo-color image, denoted as XA,l(l=1, . . . , L)
• Slicing Y into two-dimensional image slices alone axial view, and combining three consecutive label images into a three-channel label image, denoted as YA,l(l=1, . . . , L);
• adjusting a contrast of each two-dimensional image XA,l;
• cropping each pair of two-dimensional images XA,l and YA,l into fixed size image block as input of a coarse segmentation network;
• (2) coarse segmentation stage, comprising the following steps of:
• constructing a deep convolutional neural network for coarse segmentation for pancreas CT, and training the network by training samples to obtain a pancreas initial segmentation model;
• feeding the test image into the trained network model to assign each pixelin the image to be assigned a probability value PA,lC of belonging to the pancreas;
• binarizing the probability value to obtain an initial segmentation result ZlC of the pancreas;
• cropping the region of interest on the original input image according to the initial segmentation result, and denoting the cropped images as XA,lF;
• Similarly, cropping the label image and denoting it as YA,lF.
• (3) calculating a saliency map based on a geodesic distance, comprising:
• calculating a geodesic distance map according to the original image XA,lF and the initial segmentation result ZlC;
• performing a saliency transformation on the geodesic distance map to obtain a saliency map S(XA,lF);
• (4) fine segmentation stage, comprising the following steps of:
• constructing a saliency-aware densely connected dilated convolutional neural network:
• adding a saliency-aware module after each dense block of the densely connected dilated convolutional neural network to introduce saliency information based on the geodesic distance map. Denoting F(XA,lF) as output feature map of the dense block, and taking the saliency map S(XA,lF) as a weight map to act on the feature map:
• Lc(XA,lF)=Fc(XA,lF)└S(XA,lF),
• where c∈{1, 2, . . . , C} is an index of feature map channel, and └ represents element-wise multiplication;
• combining the obtained Lc(XA,lF) with the original feature map Fc(XA,lF) through an identity transformation and a parameter η:
• Hc(XA,lF)=Fc(XA,lF)└η*Lc(XA,lF),
• where the symbol └ represents element-wise addition, and the parameter η is an adjustment coefficient, which is obtained through network training. The output Hc(XA,lF) of the obtained saliency-aware module is used as an input of next transition layer to participate in the calculation;
• feeding the original image XA,lF and the corresponding label image YA,lF into the constructed saliency-aware densely connected dilated convolutional neural network for parameter training, and obtaining a fine segmentation network model;
• (5) fusing multi-layer two-dimensional pancreatic segmentation results to obtain a three-dimensional pancreatic segmentation result,
• for a test image Xtest, slicing Xtest along an axial view to obtain a two-dimensional image sequence, and combining three consecutive gray-scale images into a three-channel pseudo-color image XA,ltest(l=1, . . . , L) ;
• feeding each image into the trained coarse segmentation model and the fine segmentation model successively, and obtaining a prediction result PA,l for each image about pancreas area;
• performing multi-layer probability value fusion on the prediction result PA,l , and a predicted segmentation result ZA(l) of the two-dimensional original image of each layer is the average value of predicted values of three pseudo-color images in this layer.