Learning apparatus, learning method, and learning program, graph structure extraction apparatus, graph structure extraction method, and graph structure extraction program, and learned extraction model
A learning unit derives, from a target image including at least one tubular structure, in a case where an image for learning and ground-truth data of a graph structure included in the image for learning are input to an extraction model which extracts a feature vector of a plurality of nodes constituting a graph structure of the tubular structure, a loss between nodes on the graph structure included in the image for learning on the basis of an error between a feature vector distance between nodes belonging to the same graph structure and a topological distance which is a distance on a route of the graph structure between the nodes, and performs learning of the extraction model on the basis of the loss.
1. A learning apparatus comprising:
at least one processor, wherein the processor is configured to:
input a learning image and ground-truth data of the learning image to an extraction model, wherein the ground-truth data of the learning image comprises an extraction result of nodes of a graph structure included in the learning image;
receive a feature map for learning outputted from the extraction model such that a feature vector distance between nodes belonging to a same graph structure included in learning image corresponds to a topological distance which is a distance on a route of the graph structure between the nodes;
derive, according to the feature map for learning and the ground-truth data for learning, a loss between the nodes on the graph structure included in the learning image on the basis of a difference between the feature vector distance and the topological distance; and
perform learning of the extraction model on the basis of the loss.
2. The learning apparatus according to claim 1 ,
wherein the processor is configured to further derive, for another learning image including at least two different graph structures, a loss such that a feature vector distance between nodes belonging to a same graph structure in the another learning image is decreased.
3. The learning apparatus according to claim 2 ,
wherein the processor is configured to further derive the loss such that a feature vector distance between nodes belonging to different graph structures in the another learning image is increased.
4. The learning apparatus according to claim 1 ,
wherein the extraction model is a fully convolutional neural network.
5. The learning apparatus according to claim 1 ,
wherein the target image and the image for learning are three-dimensional medical images.
6. The learning apparatus according to claim 1 ,
wherein the tubular structure is an artery and a vein.
7. The learning apparatus according to claim 1 ,
wherein the tubular structure is an artery, a vein, and a portal vein in a liver.
8. The learning apparatus according to claim 1 ,
wherein the tubular structure is a bronchus.
9. A graph structure extraction apparatus comprising:
at least one processor, wherein the processor is configured to:
acquire a target image including at least one tabular structure;
input the target image into the extraction model learned by the learning apparatus according to claim 1 ; and
receive an extraction result of a graph structure of the tabular structure included in the target image.
10. The graph structure extraction apparatus according to claim 9 , wherein the processor is further configured to;
label the tubular structure included in the target image according to the extraction result of the graph structure; and
display the target image in which the tubular structure is labeled, on a display.
11. A learning method comprising:
inputting a learning image and ground-truth data of the learning image to an extraction model, wherein the ground-truth data of the learning image comprises an extraction result of nodes of a graph structure included in the learning image;
receiving a feature map for learning outputted from the extraction model such that a feature vector distance between nodes belonging to a same graph structure included in learning image corresponds to a topological distance which is a distance on a route of the graph structure between the nodes;
deriving, according to the feature map for learning and the ground-truth data for learning, a loss between the nodes on the graph structure included in the learning image on the basis of a difference between the feature vector distance and the topological distance; and
performing learning of the extraction model on the basis of the loss.
12. A graph structure extraction method comprising:
acquiring a target image including at least one tabular structure;
inputting the target image into the extraction model learned by the learning method according to claim 11 ; and
receiving an extraction result of a graph structure of the tabular structure included in the target image.
13. A non-transitory computer-readable storage medium that stores a graph structure extraction program causing a computer to execute a process comprising:
acquiring a target image including at least one tabular structure;
inputting the target image into the extraction model learned by the learning method according to claim 11 ; and
receiving an extraction result of a graph structure of the tabular structure included in the target image.
14. A non-transitory computer-readable storage medium that stores a learning program causing a computer to execute a process comprising:
inputting a learning image and ground-truth data of the learning image to an extraction model, wherein the ground-truth data of the learning image comprises an extraction result of nodes of a graph structure included in the learning image;
receiving a feature map for learning outputted from the extraction model such that a feature vector distance between nodes belonging to a same graph structure included in learning image corresponds to a topological distance which is a distance on a route of the graph structure between the nodes;
deriving, according to the feature map for learning and the ground-truth data for learning, a loss between the nodes on the graph structure included in the learning image on the basis of a difference between the feature vector distance and the topological distance; and
performing learning of the extraction model on the basis of the loss.