Artificial intelligence for end-to-end analytics in magnetic resonance scanning
For data analytics in magnetic resonance (MR) scanning, the scanning configuration information and the resulting raw data are directly used to determine the analytics or clinical decision. Artificial intelligence provides a value for a clinical finding characteristic of the patient based on the raw data from scanning and the controls used to scan, allowing the value to be based on all of the information content of the scan results. Reconstruction is not needed, allowing for simpler hardware, such as hardware with less homogeneous B0 and/or B1 fields than the norm and/or non-linear gradients.
1 . A method of data analytics for magnetic resonance (MR) scanning, the method comprising:
scanning, by a MR scanner, a patient using a first configuration of the MR scanner, the scanning resulting in raw data;
generating, by a machine-learned model, a value for an analytic of the patient in response to input of information of the first configuration and the raw data to the machine-learned model, wherein the machine-learned model comprises a neural network of transformer layers, a first of the transformer layers receiving tokens as input, each of the tokens being measurements from different pulses of the scanning by the MR scanner as the raw data and each of the tokens including the information as a descriptor of the first configuration of the MR scanner, the value generated without image reconstruction; and
displaying, by a display, the value for the analytic.
2 . The method of claim 1 wherein scanning further comprises scanning with a non-uniform main magnetic field, non-homogeneous first pulses, and non-linear gradients.
3 . The method of claim 1 wherein the information of the first configuration comprises first pulses, and wherein generating further comprises generating in response to input of the first pulses to the machine-learned model.
4 . The method of claim 1 wherein the information of the first configuration comprises scanner controls comprising k-space coordinates, repetition time, time-to-echo, and flip angle, and wherein generating further comprises generating in response to input of the k-space coordinates, the repetition time, the time-to-echo, and the flip angle to the machine-learned model.
5 . The method of claim 1 wherein the machine-learned model further comprises a sequence of transformers comprising attention layers, wherein the information is used as position encoding within the sequence.
6 . The method of claim 1 wherein the machine-learned model is pretrained with self-supervision at a plurality of other scanners.
7 . The method of claim 1 wherein generating further comprises generating the value for cancer or not as the analytic.
8 . The method of claim 1 wherein generating further comprises generating the value for no further scanning or further scanning as the analytic.
9 . The method of claim 1 wherein generating further comprises generating without MR image reconstruction as an input to the machine-learned model and without performance of the MR image reconstruction by or within the machine-learned model.
10 . A method of data analytics for magnetic resonance (MR) scanning, the method comprising:
scanning a patient using a first configuration of an MR scanner, the scanning resulting in raw data;
generating a value for a parameter as a global characterization of the patient in response to input of information of the first configuration and the raw data to a machine-learned model, wherein the machine-learned model was pre-trained with self-supervision at a plurality of other scanners; and
displaying the value for the parameter.
11 . The method of claim 10 wherein generating further comprises generating without MR image reconstruction from the raw data.
12 . The method of claim 10 wherein the machine-learned model comprises a neural network of transformer layers, a first of the transformer layers receiving tokens as input, each of the tokens being measurements from different pulses of the MR scanning as the raw data and each of the tokens including the information as a descriptor of settings of the MR scanner.
13 . A system for magnetic resonance (MR) scanning, the system comprising:
a MR scanner configured by settings of controls to scan a region of a patient, the scan providing raw data;
a processor configured to determine a global analytic for the patient by application of a machine-learned network to the raw data without reconstruction from the raw data, wherein the machine-learned network comprises a neural network of a sequence of attention layers where an input of the neural network receives tokens, each token being the raw data from one or more coils in combination with information about the settings of the controls for a transmit pulse; and
a display configured to output the global analytic.