Method and system for image registration using an intelligent artificial agent
Methods and systems for image registration using an intelligent artificial agent are disclosed. In an intelligent artificial agent based registration method, a current state observation of an artificial agent is determined based on the medical images to be registered and current transformation parameters. Action-values are calculated for a plurality of actions available to the artificial agent based on the current state observation using a machine learning based model, such as a trained deep neural network (DNN). The actions correspond to predetermined adjustments of the transformation parameters. An action having a highest action-value is selected from the plurality of actions and the transformation parameters are adjusted by the predetermined adjustment corresponding to the selected action. The determining, calculating, and selecting steps are repeated for a plurality of iterations, and the medical images are registered using final transformation parameters resulting from the plurality of iterations.
1. A method for deformable registration of medical images using an intelligent artificial agent, comprising:
receiving a first medical image and a second medical image;
calculating action-values for a plurality of actions based on a current state using a trained deep neural network, the current state determined from the first medical image and the second medical image;
selecting an action from the plurality of actions based on the calculated action-values;
computing a deformation field based on the selected action and a deformation model;
applying the deformation field to the first medical image to warp the first medical image; and
repeating the calculating, the selecting, the computing, and the applying for a plurality of iterations using the warped first medical image as the first medical image to register the first medical image and the second medical image.
2. The method of claim 1 , wherein calculating action-values for a plurality of actions based on a current state using a trained deep neural network comprises:
extracting features from the first medical image using first convolutional layers of the trained deep neural network;
extracting features from the second medical image using second convolutional layers of the trained deep neural network; and
comparing the extracted features from the first medical image and the extracted features from the second medical image to calculate the action-values for the plurality of actions.
3. The method of claim 2 , wherein the first convolutional layers and the second convolutional layers are implemented with separate weights.
4. The method of claim 1 , wherein computing a deformation field based on the selected action and a deformation model comprises:
encoding deformation for an entirety of the first medical image using a dense model.
5. The method of claim 4 , wherein applying the deformation field to the first medical image to warp the first medical image comprises:
decoding the deformation for the entirety of the first medical image; and
applying the decoded deformation to the first medical image.
6. The method of claim 1 , wherein computing a deformation field based on the selected action and a deformation model comprises:
encoding deformation at particular points in the first medical image that have corresponding points in the second medical image using a correspondence based model.
7. The method of claim 6 , wherein applying the deformation field to the first medical image to warp the first medical image comprises:
decoding the deformation at the particular points in the first medical image;
transforming a mesh of a segmented structure in the first medical image based on the decoded deformation;
computing a dense deformation field based on the transformation of the mesh; and
applying the computed dense deformation field to the first medical image.
8. The method of claim 1 , wherein computing a deformation field based on the selected action and a deformation model comprises:
modeling deformation of pixels at a location in the first medical image as having a Gaussian distribution; and
computing a deformation at a particular point in the first medical image based on a distance between the particular point and the location, a magnitude of the Gaussian distribution, and a standard deviation of the Gaussian distribution.
9. The method of claim 1 , wherein the trained deep neural network is trained to predict the action-values for the plurality of actions based on a plurality of training image pairs with known ground truth transformation parameters using reinforcement learning in which, for each training image pair, a reward for each action of the plurality of actions at a given state is used to train the deep neural network to learn an optimal registration policy.
10. An apparatus for deformable registration of medical images using an intelligent artificial agent, comprising:
means for receiving a first medical image and a second medical image;
means for calculating action-values for a plurality of actions based on a current state using a trained deep neural network, the current state determined from the first medical image and the second medical image;
means for selecting an action from the plurality of actions based on the calculated action-values;
means for computing a deformation field based on the selected action and a deformation model;
means for applying the deformation field to the first medical image to warp the first medical image; and
means for repeating the calculating, the selecting, the computing, and the applying for a plurality of iterations using the warped first medical image as the first medical image to register the first medical image and the second medical image.
11. The apparatus of claim 10 , wherein the means for calculating action-values for a plurality of actions based on a current state using a trained deep neural network comprises:
means for extracting features from the first medical image using first convolutional layers of the trained deep neural network;
means for extracting features from the second medical image using second convolutional layers of the trained deep neural network; and
means for comparing the extracted features from the first medical image and the extracted features from the second medical image to calculate the action-values for the plurality of actions.
12. The apparatus of claim 11 , wherein the first convolutional layers and the second convolutional layers are implemented with separate weights.
13. The apparatus of claim 10 , wherein the means for computing a deformation field based on the selected action and a deformation model comprises:
means for encoding deformation for an entirety of the first medical image using a dense model.
14. The apparatus of claim 13 , wherein the means for applying the deformation field to the first medical image to warp the first medical image comprises:
means for decoding the deformation for the entirety of the first medical image; and
means for applying the decoded deformation to the first medical image.
15. A non-transitory computer readable medium storing computer program instructions for deformable registration of medical images using an intelligent artificial agent, the computer program instructions defining operations comprising:
receiving a first medical image and a second medical image;
calculating action-values for a plurality of actions based on a current state using a trained deep neural network, the current state determined from the first medical image and the second medical image;
selecting an action from the plurality of actions based on the calculated action-values;
computing a deformation field based on the selected action and a deformation model;
applying the deformation field to the first medical image to warp the first medical image; and
repeating the calculating, the selecting, the computing, and the applying for a plurality of iterations using the warped first medical image as the first medical image to register the first medical image and the second medical image.
16. The non-transitory computer readable medium of claim 15 , wherein calculating action-values for a plurality of actions based on a current state using a trained deep neural network comprises:
extracting features from the first medical image using first convolutional layers of the trained deep neural network;
extracting features from the second medical image using second convolutional layers of the trained deep neural network; and
comparing the extracted features from the first medical image and the extracted features from the second medical image to calculate the action-values for the plurality of actions.
17. The non-transitory computer readable medium of claim 15 , wherein computing a deformation field based on the selected action and a deformation model comprises:
encoding deformation at particular points in the first medical image that have corresponding points in the second medical image using a correspondence based model.
18. The non-transitory computer readable medium of claim 17 , wherein applying the deformation field to the first medical image to warp the first medical image comprises:
decoding the deformation at the particular points in the first medical image;
transforming a mesh of a segmented structure in the first medical image based on the decoded deformation;
computing a dense deformation field based on the transformation of the mesh; and
applying the computed dense deformation field to the first medical image.
19. The non-transitory computer readable medium of claim 15 , wherein computing a deformation field based on the selected action and a deformation model comprises:
modeling deformation of pixels at a location in the first medical image as having a Gaussian distribution; and
computing a deformation at a particular point in the first medical image based on a distance between the particular point and the location, a magnitude of the Gaussian distribution, and a standard deviation of the Gaussian distribution.
20. The non-transitory computer readable medium of claim 15 , wherein the trained deep neural network is trained to predict the action-values for the plurality of actions based on a plurality of training image pairs with known ground truth transformation parameters using reinforcement learning in which, for each training image pair, a reward for each action of the plurality of actions at a given state is used to train the deep neural network to learn an optimal registration policy.