Siemens Healthineers Ag

Siemens Healthcare Gmbh

Mahmoud Mostapha

Mariappan S Nadar

Simon Arberet

Boris Mailhe

Marcel Dominik Nickel

Thomas Benkert

Pascal Ceccaldi

Youngjin Yoo

Gregor Körzdörfer

Dorin Comaniciu

Esther Raithel

Nirmal Janardhanan

Hongki Lim

Systems and methods for reconstruction for a medical imaging system. An adapter is used to adapt scan data so that different quantities of repetitions or directions may be used to train and implement a single multichannel backbone network.

Multichannel deep learning reconstruction of multiple repetitions

For reconstruction in medical imaging, such as reconstruction in MR imaging, scanning is accelerated by under-sampling. In iterative reconstruction, the input to the regularizer is altered provide for correlation of non-local aliasing artifacts. Duplicates of the input image are shifted by different amounts based on the level of acceleration. The resulting shifted images are used to form the input to the regularizer. Providing an input based on shifts allows the regularization to suppress non-local as well as local aliasing artifacts.

Accelerated medical image reconstruction with improved aliasing artifact reduction

For reconstruction in medical imaging, such as reconstruction in MR imaging, an iterative, hierarchal network for regularization may decrease computational complexity. To further maintain computational complexity while improving robustness, auxiliary information is used in the regularization and corresponding reconstruction. The auxiliary information is in put to the machine-learned network.

Iterative hierarchal network for regulating medical image reconstruction

A computer-implemented method includes, based on an input dataset defining an input image, determining a reconstructed image using a reconstruction algorithm, and executing a data-consistency operation for enforcing consistency between the input image and the reconstructed image. The data-consistency operation determines, for multiple K-space positions at which the input dataset comprises respective source data, a contribution of respective K-space values associated with the input dataset to a K-space representation of the reconstructed image.

Data-consistency for image reconstruction

For reconstruction, a machine-learned model is adapted to allow for reconstruction based on the repetitions available in some scanning. The reconstruction for one or more subsets is performed during the scanning. The machine-learned model is trained to reconstruction separately or independently for each repetition or to use information from previous repetitions without requiring waiting for completion of scanning. The reconstructed image may be displayed much more rapidly after completion of the acquisition since the reconstruction begins during the reconstruction.

Deep learning-based realtime reconstruction

For reconstruction in medical imaging, self-consistency using data augmentation is improved by including data consistency. Artificial intelligence is trained based on self-consistency and data consistency, allowing training without supervision. Fully sampled data and/or ground truth is not needed but may be used. The machine-trained model is less likely to reconstruct images with motion artifacts, and/or the training data may be more easily gathered by not requiring full sampling.

Self-supervised machine learning for medical image reconstruction

For reconstruction in medical imaging, user control of a characteristic (e.g., noise level) of the reconstructed image is provided. A machine-learned model alters the reconstructed image to enhance or reduce the characteristic. The user selected level of characteristic is then provided by combining the reconstructed image with the altered image based on the input level of the characteristic. Personalized or more controllable impression for medical imaging reconstruction is provided without requiring different reconstructions.

Reconstruction with user-defined characteristic strength

For correction of an image from an imaging system, an inverse solution uses an imaging prior as a regularizer and a physics model of the imaging system. An invertible network is used as the deep-learnt generative model in the regularizer of the inverse solution with the physics model of the degradation behavior of the imaging system. The prior model based on the invertible network provides a closed-form expression of the prior probability, resulting in a more versatile or accurate probability prediction.

Image correction using an invertable network

For reconstruction in medical imaging using a scan protocol with repetition, a machine learning model is trained for reconstruction of an image for each repetition. Rather than using a loss for that repetition in training, the loss based on an aggregation of images reconstructed from multiple repetitions is used to train the machine learning model. This loss for reconstruction of one repetition based on aggregation of reconstructions for multiple repetitions is based on deep set-based deep learning. The resulting machine-learned model may better reconstruct an image from a given repetition and/or a combined image from multiple repetitions than a model learned from a loss per repetition.

Aggregation based on deep set learning in magnetic resonance reconstruction

A computer-implemented method includes, based on scan data defining an input image, determining a reconstructed image using a reconstruction algorithm, and executing a data consistency operation for enforcing consistency between the input image and the reconstructed image. The data consistency operation includes using a norm ball projection that takes into account the available noise level information in order to automatically adjust the balance between the network prediction and the input measurements.

Noise adaptive data consistency in deep learning image reconstruction via norm ball projection

For reconstruction in medical imaging, such as reconstruction in MR imaging, an iterative, hierarchal network for regularization may decrease computational complexity. The machine-learned network of the regularizer is unrolled or made iterative. For each iteration, nested U-blocks form a hierarchy so that some of the down-sampling and up-sampling of some U-blocks begin and end with lower resolution data or features, reducing computational complexity.

For reconstruction of an image in MRI, unsupervised training (i.e., data-driven) based on a scan of a given patient is used to reconstruct model parameters, such as estimating values of a contrast model and a motion model based on fit of images generated by the models for different readouts and times. The models and the estimated values from the scan-specific unsupervised training are then used to generate the patient image for that scan. This may avoid artifacts from binning different readouts together while allowing for scan sequences using multiple readouts.

Reconstruction in magnetic resonance imaging with image representations as implicit functions in time

For medical imaging such as MRI, machine training is used to train a network for segmentation using both the imaging data and protocol data (e.g., meta-data). The network is trained to segment based, in part, on the configuration and/or scanner, not just the imaging data, allowing the trained network to adapt to the way each image is acquired. In one embodiment, the network architecture includes one or more blocks that receive both types of data as input and output both types of data, preserving relevant features for adaptation through at least part of the trained network.

Protocol-aware tissue segmentation in medical imaging

A system is provided for MRI coil sensitivity estimation and reconstruction At least two cascades of regularization networks are serially connected such that the output of a cascade is used as input of a following cascade, at least two deepsets coil sensitivity map networks are serially connected such that the output of a deepsets coil sensitivity map network is used as input of a following deepsets coil sensitivity map network (CR), and wherein the outputs of the deepsets coil sensitivity map networks are also used as inputs for the cascades.

Growing community of inventors

Princeton, NJ, United States of America

Mahmoud Mostapha