CPC G06T 7/11 (2017.01) [G06T 3/4046 (2013.01); G06T 5/50 (2013.01); G06T 7/0012 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/20221 (2013.01); G06T 2207/30088 (2013.01)] | 6 Claims |
1. A method executed by a computer, the method comprising:
1) acquiring a medical image data set;
2) acquiring, from the medical image data set, an original image and a real segmentation image of a target region in the original image in pair to serve as an input data set of a pre-built constant-scaling segmentation network, the input data set comprising a training set, a verification set and a test set;
3) training the constant-scaling segmentation network by using the training set to obtain a trained segmentation network model, and verifying the constant-scaling segmentation network by using the verification set, the constant-scaling segmentation network comprising a feature extraction module and a resolution amplifying module, wherein, during decoding, each decoder layer is connected to a corresponding tailored feature map from a corresponding layer of an encoder; and
4) inputting the original image to be segmented into the segmentation network model for segmentation to obtain a real segmentation image;
wherein in 3), in the constant-scaling segmentation network, a loss function is set as a set similarity measure function, which is expressed by the following formula:
![]() where |A∩B| represents common elements between a set A and a set B; |A| represents a number of elements in the set A; |B| represents a number of elements in the set B; the elements in the set A are real segmentation images obtained after segmentation by the constant-scaling segmentation network using the input data set; and, the elements in the set B are real segmentation images of the target region in the original image; and
to calculate the set similarity measure function of a predicted real segmentation image, |A|+|B| is approximate to a dot product of multiplication of an actually segmented image and the real segmentation image, and values of each pixel point in the set A and the set B are added; and, when the loss function is minimal, the training is stopped to obtain the trained segmentation network model;
wherein in 3), stopping training when the loss function is minimal to obtain the trained segmentation network model comprises:
3.1) initializing weight parameters of the constant-scaling segmentation network in each stage on the basis of an Adam optimizer, and randomly initializing the weight parameters by using a Gaussian distribution with an average value of 0;
3.2) for each sample image that is input into the training set of the segmentation network model and comprises a composite image and the original image, calculating a total error between the real segmentation image obtained by the constant-scaling segmentation network and the real segmentation image of the target region in the original image by forward propagation, then calculating a partial derivative of each weight parameter by back propagation, and updating the weight parameters by a gradient descent method; and
3.3) repeating 3.1) and 3.2) until the loss function is minimal, to obtain the trained segmentation network model.
|