Un-supervised convolutional neural network for distortion map estimation and correction in MRI

1. A method for magnetic resonance imaging (MRI) comprising:

performing by an MRI apparatus an echo planar imaging acquisition using inverted up/down phase encoding directions and reconstructing acquired images having geometric distortions along the phase encoding directions due to off-resonant spins;

feed-forward estimating by a convolutional neural network (CNN) a phase distortion map from the acquired images;

wherein the CNN has a U-net architecture and is trained to minimize a similarity metric between un-warped up/down image pairs;

performing geometric distortion correction of the acquired images using the phase distortion map to unwarp the acquired images; wherein performing geometric distortion correction of the acquired images comprises unwarping the acquired images by 1d cubic interpolation along a phase encoding direction with density weighting to account for signal pile-up that takes place during training of the CCN.

2. The method of claim 1 wherein the U-net architecture of the CNN has an encoder section with encoder convolution layers having decreasing feature map dimensions, followed by a decoder section with decoder transposed convolution layers having increasing feature map dimensions.

3. The method of claim 2 wherein the decoder transposed convolution layers receive information flow from encoder convolution layers of corresponding dimensions, wherein the information flow comprises feature maps.

4. The method of claim 2 wherein at least one of the decoder transposed convolution layers receives an up-sampled distortion map estimated from a previous layer.

5. The method of claim 1 wherein the density weighting is estimated by counting contributions of neighboring points weighted by a linear distance metric.