JPWO2020041503A5 - - Google Patents

Claims (37)

A non-temporary processor-readable storage medium that stores at least one of a processor executable instruction or data.
With at least one processor communicably coupled to said at least one non-temporary processor readable storage medium.
It is a machine learning system equipped with
While the at least one processor is in operation
Receiving training data containing multiple batches of unlabeled image sets, each image set contains a source image and a target image each representing a medical image scan of at least one patient.
One or more between multiple batches of the unlabeled image set that trains one or more convolutional neural network (CNN) models to allow registration of the target image to the source image. Learn the conversion function of
The one or more trained CNN models are stored in the at least one non-temporary processor readable storage medium of the machine learning system.
Machine learning system. The machine learning system of claim 1, wherein the at least one processor trains the one or more CNN models using an unsupervised training algorithm. The machine learning system of claim 2, wherein the unsupervised training algorithm comprises a loss function that is calculated from the source image and the pair of target images and not from any explicit human-generated annotation on the image. .. The machine learning system of claim 3, wherein the loss function comprises a root mean square error per pixel between the source image and the target image. The machine learning system of claim 3, wherein the differentiable objective function comprises a mutual information loss between the source image and the target image. The machine learning system of claim 3, wherein the differentiable objective function comprises an L2 loss between the source image and the target image. The machine learning system of claim 3, wherein the differentiable objective function comprises a center weighted L2 loss function between the source image and the target image. The machine learning system of claim 3, wherein the differentiable objective function comprises a normalized cross-correlation loss function between the source image and the target image. The machine learning system of claim 1, wherein the plurality of batches of the unlabeled image set comprises one or both of 2D or 3D images. The machine learning system of claim 1, wherein the transformation function comprises one or both of an affine transformation or a dense non-linear correspondence map. 10. The machine learning system of claim 10, wherein the transformation function comprises the high density nonlinear correspondence map comprising a high density deformation field (DDF). The machine learning system according to claim 1, wherein the one or more CNN models include a global network model, and the global network model receives the training data and outputs an affine transformation matrix. The machine learning system according to claim 12, wherein the affine transformation matrix is calculated on the target image with respect to the source image. 12. The machine learning system of claim 12, wherein the source image and the target image include all possible combinations of image pairings. 15. The machine learning system of claim 14, wherein the source image and the target image include all images in a single cardiac MR scan. 14. The machine learning system of claim 14, wherein the source image and the target image include all images from one or more heterogeneous MR scan volumes. 12. The machine learning system of claim 12, wherein the global network model comprises a contraction path comprising a group of at least one layer including at least one convolution layer, a maximum pooling layer, a batch normalization layer, and a dropout layer. 17. The machine learning system of claim 17, wherein the global network model comprises a rectifier or a leaky rectifier that follows at least one of the groups of layers in the contraction path. The machine learning system according to claim 12, wherein the affine transformation matrix output by the global network model includes an affine space transformation layer. The machine learning system according to claim 12, wherein the affine transformation of the affine transformation matrix is limited by a scaling coefficient. The machine learning system according to claim 12, wherein the affine transformation matrix includes a regularization operation. 21. The machine learning system of claim 21, wherein the regularization operation comprises bending energy loss. 21. The machine learning system of claim 21, wherein the regularization operation comprises gradient energy loss. The machine learning system according to claim 1, wherein the one or more CNN models include a local network model that receives the training data and outputs a high density deformation field of the local network. 24. The machine learning of claim 24, wherein the at least one processor warps the target image to provide a warp target image, and the warp target image is acquired by applying an affine transformation field to the original target image. system. The local network model comprises a contraction path and an expansion path, the contraction path comprises one or more convolutional layers and one or more pooling layers, each pooling layer precedes at least one convolutional layer, said expansion. The path comprises several convolution layers and several upsampling layers, each upsampling layer precedes at least one convolutional layer, and each upsampling layer comprises at least one upsampling operation and trained kernel. 24. The machine learning system according to claim 24, comprising a translocation convolution operation for executing the interpolation operation, or an upsampling operation following the interpolation operation. 24. The machine learning system of claim 24, wherein the output of the high density deformation field of the local network comprises a free-form similarity space converter. 27. The machine learning system of claim 27, wherein the free-form similarity space converter comprises an affine transformation. 27. The machine learning system of claim 27, wherein the freeform similarity space converter comprises high density freeform transformation field warping. 24. The machine learning system of claim 24, wherein the output of the high density transformation field of the local network comprises a regularization operation. 30. The machine learning system of claim 30, wherein the regularization operation comprises bending energy loss. 30. The machine learning system of claim 30, wherein the regularization operation comprises gradient energy loss. The machine learning system of claim 1, wherein the one or more CNN models include a global network, a local network, and an output high density deformation field. The machine learning system of claim 1, wherein the at least one processor uses an Adam optimizer that uses an unsupervised differentiable loss function to optimize the one or more CNN models. 34. The machine learning system of claim 34, wherein the at least one processor calculates the unsupervised loss function between the source image and the warp target image. 35. The machine learning system of claim 35, wherein the warp target image is acquired by applying a high density deformation field to the original target image. The machine learning system of claim 1, wherein the image set includes the cardiac short axis CINE MR series .