Implicit neural representation learning with prior embedding for sparsely sampled image reconstruction and other inverse problems
View Patent ↗A method for diagnostic imaging reconstruction uses a prior image x pr from a scan of a subject to initialize parameters of a neural network which maps coordinates in image space to corresponding intensity values in the prior image. The parameters are initialized by minimizing an objective function representing a difference between intensity values of the prior image and predicted intensity values output from the neural network. The neural network is then trained using subsampled (sparse) measurements of the subject to learn a neural representation of a reconstructed image. The training includes minimizing an objective function representing a difference between the subsampled measurements and a forward model applied to predicted image intensity values output from the neural network. Image intensity values output from the trained neural network from coordinates in image space input to the trained neural network are computed to produce predicted image intensity values.
1. A method for diagnostic imaging reconstruction comprising:
storing a prior image x pr from a scan of a subject, comprising image intensity at each coordinate in image space;
initializing parameters of a neural network using the prior image x pr ;
wherein the neural network maps coordinates in image space to corresponding intensity values in the prior image;
wherein initializing the parameters comprises minimizing an objective function representing a difference between intensity values of the prior image and predicted intensity values output from the neural network, thereby creating an implicit neural representation of the prior image;
performing a scan to acquire subsampled (sparse) measurements y of the subject;
training the neural network using the measurements y to learn a neural representation of a reconstructed image x, wherein the training comprises minimizing an objective function representing a difference between the measurements y and a forward model applied to predicted image intensity values output from the neural network;
computing image intensity values output from the trained neural network from coordinates in image space input to the trained neural network to produce predicted image intensity values.
2. The method of claim 1 wherein Fourier feature mapping is used to transform spatial coordinates to encoded coordinates prior to input to the neural network.
3. The method of claim 1 wherein the neural network is implemented by a deep fully-connected network or multi-layer perceptron (MLP).
4. The method of claim 1 wherein the MLP uses periodic activation functions after each fully-connected layer.
5. The method of claim 1 wherein performing a scan to acquire subsampled (sparse) measurements y of the subject comprises performing an MRI scan to acquire the subsampled (sparse) measurements y of the subject.
6. The method of claim 1 wherein performing a scan to acquire subsampled (sparse) measurements y of the subject comprises performing an CT scan to acquire the subsampled (sparse) measurements y of the subject.