Medical image segmentation from raw data using a deep attention neural network
Various approaches provide improved segmentation from raw data. Training samples are generated by medical imaging simulation from digital phantoms. These training samples provide raw measurements, which are used to learn to segment. The segmentation task is the focus, so image reconstruction loss is not used. Instead, an attention network is used to focus the training and trained network on segmentation. Recurrent segmentation from the raw measurements is used to refine the segmented output. These approaches may be used alone or in combination, providing for segmentation from raw measurements with less influence of noise or artifacts resulting from a focus on reconstruction.
1. A method for segmentation from raw data of a magnetic resonance imager, the method comprising:
acquiring, by the magnetic resonance imager, k-space data representing response from a patient;
segmenting an object of the patient represented in the k-space data by a machine-learned neural network, the machine-learned neural network having a segmentation network recurrently applied to output an output segmentation of the object in response to input of the k-space data to the machine-learned neural network; and
generating an image based on the segmentation.
2. The method of claim 1 wherein acquiring comprises acquiring with undersampling according to an undersampling mask, the machine-learned neural network including a data consistency layer configured to replace k-space lines from a reconstruction image with the k-space data for the k-space lines based on input of an undersampling mask.
3. The method of claim 1 wherein the machine-learned model comprises a sequence of reconstruction blocks and data consistency layers, a last one of the data consistency layers outputting to the segmentation network, wherein segmenting comprises outputting the output segmentation after two or more of the recurrent applications of the segmentation network.
4. The method of claim 1 wherein segmenting comprises segmenting with the machine-learned neural network having been trained with a segmentation loss and not a reconstruction loss.
5. The method of claim 1 wherein segmenting comprises segmenting with the segmentation network, a regularization block, and a data consistency layer being recurrently applied to generate the segmentation.
6. The method of claim 1 wherein segmenting comprises segmenting with the segmentation network recurrently applied two or three times.
7. The method of claim 1 wherein segmenting comprises segmenting with the machine-learned neural network including an attention network, the attention network generating multi-class attention maps from pre-output segmentations by the segmentation network.
8. The method of claim 7 wherein generating comprises generating by weighting images based on probability maps of different tissues as the pre-output segmentations, results of the weighting input to a regularization block of the segmentation network.
9. The method of claim 1 wherein segmenting comprises segmenting with the machine-learned neural network having been trained with training data generated by undersampling outputs of a magnetic resonance simulation from digital phantoms.
10. A method for segmentation from raw data of a magnetic resonance imager, the method comprising:
acquiring, by the magnetic resonance imager, k-space data representing response from a patient;
segmenting an object of the patient represented in the k-space data by a machine-learned neural network, the machine-learned neural network including an attention network, the machine-learned neural network outputting an output segmentation of the object in response to input of the k-space data to the machine-learned neural network; and
generating an image showing the segmentation.
11. The method of claim 10 wherein segmenting comprises segmenting by machine-learned neural network, the machine-learned neural network having a segmentation network recurrently applied to output the output segmentation.
12. The method of claim 10 wherein acquiring comprises acquiring with undersampling according to an undersampling mask, the machine-learned neural network including a data consistency layer configured to replace k-space lines from a reconstruction image with the k-space data for the k-space lines based on input of an undersampling mask.
13. The method of claim 10 wherein the machine-learned model comprises a sequence of reconstruction blocks and data consistency layers, a last one of the data consistency layers outputting to the segmentation network, wherein segmenting comprises segmenting with the machine-learned neural network having been trained with a segmentation loss and not a reconstruction loss.
14. The method of claim 10 wherein segmenting comprises generating, by the attention network, multi-class attention maps from pre-output segmentations by the segmentation network.
15. The method of claim 14 wherein generating comprises generating by weighting images based on probability maps of different tissues as the pre-output segmentations, results of the weighting input to a regularization block of a segmentation network.
16. The method of claim 10 wherein segmenting comprises segmenting with the machine-learned neural network having been trained with training data generated by undersampling outputs of a magnetic resonance simulation from digital phantoms.
17. A method for training segmentation from raw data of a reconstruction medical imager, the method comprising:
generating training data comprising pre-reconstruction measures and segmentation maps, the training data generated by medical imaging simulation of digital phantoms; storing the trained neural network; and
defining the neural network to include recurrent segmentation and an attention layer configured to output multi-class segmentation probability maps to a regularization block.
18. The method of claim 17 wherein generating comprises generating k-space or sinogram data as the pre-reconstruction measures, further comprising removing some of the k-space or sinogram data in a simulation of undersampling, and wherein training comprises training based on the training data resulting from the simulation of undersampling.
19. The method of claim 17 wherein training comprises training with a segmentation loss and not a reconstruction loss.