Deep learning volumetric deformable registration
View Patent ↗A method and system for automated deformable registration of an organ from medical images includes generating segmentations of the organ by processing a first and second series of images corresponding to different organ states using a first trained CNN. A second trained CNN processes the first and second series of images and the segmentations to deformably register the second series of images to the first series of images. The second trained CNN predicts a displacement field by minimizing a registration loss function, where the displacement field maximizes colocalization of the organ between the different states.
1 . A method for automated deformable registration of an organ from images, the method comprising:
receiving in a computer processor configured for executing a trained convolutional neural network (CNN) image data to comprising computed tomography (CT) series pairs comprising a first series of images of the organ in a first state and a second series of images of the organ in at least one second state;
propagating the first and second series of images of the organ using a first trained CNN to generate segmentations of the organ;
propagating the first and second series of images and the segmentations of the organ using a second trained CNN to deformably register the second series of images to the first series of images, wherein the second trained CNN is trained to predict a displacement field by minimizing a registration loss function, wherein the displacement field is configured to maximize colocalization of the organ between the first state and the at least one second state; and
generating an output comprising a deformed image corresponding to the at least one second state;
wherein the organ comprises one or more lung, the first state is inspiratory and the at least one second state is expiratory, and the CT series pairs comprise volumetric lung CT scans.
2 . The method of claim 1 , wherein the registration loss function is a combination of one or more loss function selected from the group consisting of cross-correlation loss, displacement field loss, Dice loss, and Jacobian loss.
3 . The method of claim 1 , wherein the image data comprises voxels.
4 . The method of claim 3 , wherein the first and second trained CNNs are 3D U-Net CNNs.
5 . The method of claim 1 , further comprising extracting lung measurements from the segmentations.
6 . The method of claim 1 , further comprising generating a disease map from a visual comparison of the first series of images and the deformed image.
7 . The method of claim 1 , further comprising pre-processing the image data by one or more of resizing to a standard resolution, scaling voxel attenuations, and affine-registering the second series to the first series.
8 . A system for automated deformable registration of an organ from medical images, comprising:
at least one computer processor configured to:
acquire image data comprising computed tomography (CT) series pairs comprising a first series of images of the organ in a first state and a second series of images of the organ in at least one second state;
execute a first convolutional neural network (CNN), wherein the first CNN trained to generate segmentations of the organ using the first and second series of images;
execute a second CNN, wherein the second CNN is trained to deformably register the second series of images to the first series of images using the first and second series of images and the segmentations of the organ, wherein the second CNN is further configured to predict a displacement field by minimizing a registration loss function, wherein the displacement field is configured to maximize colocalization of the organ between the first state and the at least one second state; and
generate an output comprising a deformed image corresponding to the at least one second state;
wherein the organ comprises one or more lung, the first state is inspiratory and the at least one second state is expiratory, and the CT series pairs comprise volumetric lung CT scans.
9 . The system of claim 8 , wherein the registration loss function is a combination of one or more loss function selected from the group consisting of cross-correlation loss, displacement field loss, Dice loss, and Jacobian loss.
10 . The system of claim 9 , wherein the image data comprises voxels.
11 . The system of claim 9 , wherein the first and second CNNs are 3D U-Net CNNs.
12 . The system of claim 8 , wherein the at least one computer processor is further configured for extracting lung measurements from the segmentations.
13 . The system of claim 8 , wherein the at least one computer processor is further configured for generating a disease map from a visual comparison of the first series of images and the deformed image.
14 . The system of claim 8 , wherein the at least one computer processor is further configured for pre-processing the image data by one or more of resizing to a standard resolution, scaling voxel attenuations, and affine-registering the second series to the first series.