Adaptive sampling for parallel imaging

Abstract

A method using one or more processors includes receiving a signal associated with a first value including a first portion of magnetic resonance imaging (MRI) data during an MRI scan. The method also includes generating, using at least one model and the signal, an image including a second portion of the MRI data. The method includes generating a map and determining at least one uncertainty value based at least partially on the map. The method includes comparing the at least one uncertainty value to a threshold and adjusting the first value to a second value responsive to the at least one uncertainty value being above the threshold. The method includes resuming the MRI scan at the second value or resuming the MRI scan at the first value responsive to the at least one uncertainty value being at or below the threshold.

Description

Ref. No.: LAT01-07PCT (046434-0973)ADAPTIVE SAMPLING FOR PARALLEL IMAGINGCROSS-REFERENCE TO RELATED APPLICATION

[0001] The present application claims priority to Greek Patent Application No. 20250100032, filed on January 17, 2025, the entire disclosure of which is incorporated herein by reference for all purposes.TECHNICAL FIELD

[0002] The present disclosure relates generally to adaptive sampling for parallel imaging.BACKGROUND

[0003] Magnetic resonance (MR) imaging (MRI), despite significant improvements over the past decade, has scan times longer than competing imaging modalities such as computed tomography (CT) or X-ray. MRI also does not have real-time adaptability and still relies on users (e.g., radiologist, technician) to ensure image quality. Conventional methods lack ways to determine whether an MRI is clinically usable until the scan is complete. Responsive to the user determining that the scan is not clinically usable, the scan may need to be repeated, thereby increasing scan times and cost due to suboptimal image acquisition settings.

[0004] In addition, the majority of MRI scans employ parallel imaging which uses multiple receiver coils of the MRI machine to reduce image acquisition time as well as reconstruction algorithms to reconstruct a full MR image from undersampled data. Conventional parallel imaging protocols employ a predetermined acceleration factor, typically set at two for non-learning-based methods used for reconstruction (e.g., Fourier transform reconstruction), or up to four if a learningbased method is used (e.g., neural network). For certain exams, an acceleration factor of six or eight could still yield diagnostic image quality when combined with a deep learning reconstruction. Despite this, a conservative, predetermined lower acceleration factor is used because it is currently not possible to predict the diagnostic accuracy of the reconstructed image while setting up a protocol at the scanner, thus unnecessarily prolonging scan times.14932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)SUMMARY

[0005] One aspect of the present disclosure is directed towards a method. The method includes receiving, by one or more processors, a signal associated with a first value comprising a first portion of magnetic resonance imaging (MRI) data during an MRI scan. The method also includes generating, by the one or more processors using at least one model and the signal, an image including a second portion of the MRI data. The method includes generating, by the one or more processors using the at least one model, a map corresponding to the image and determining, by the one or more processors using the at least one model, at least one uncertainty value based at least partially on the map. The method includes comparing, by the one or more processors, the at least one uncertainty value to a threshold and adjusting, by the one or more processors, the first value to a second value responsive to the at least one uncertainty value being above the threshold. The method also includes resuming, by the one or more processors, the MRI scan at the second value responsive to adjusting the first value to the second value and resuming, by the one or more processors, the MRI scan at the first value responsive to the at least one uncertainty value being at or below the threshold.

[0006] Another aspect of the present disclosure is directed towards another method. The method includes performing, by one or more processors, a scan using a magnetic resonance imaging (MRI) machine at a first acceleration factor and receiving, by the one or more processors, a first signal associated with the first acceleration factor. The method also includes generating, by the one or more processors using a reconstruction model, an image based on the first signal and generating, by the one or more processors using at least one neural network, an uncertainty map corresponding to the image. The method includes determining, by the one or more processors, at least one uncertainty value based on the uncertainty map and comparing, by the one or more processors, the at least one uncertainty value to a threshold. The method includes adjusting, by the one or more processors, the overall acceleration factor to a second acceleration factor responsive to the at least one uncertainty value being above the threshold. The method includes performing, by the one or more processors, the scan to acquire additional signal responsive to effectively adjusting the overall acceleration factor to a second acceleration factor and keeping as is the scan at the first acceleration factor responsive to the at least one uncertainty value being at or below the threshold.24932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0007] Another aspect of the present disclosure is directed towards a system. The system includes one or more processors and at least memory. The one or more processors are configured to receive a first signal associated with a first parameter, the first signal received from a medical imaging system during performance of an imaging scan. The one or more processors are configured to generate, using a first model, an image based on the first signal. The one or more processors are configured to generate, using a second model, an uncertainty map corresponding to the image and determine, based on the uncertainty map, at least one characteristic of the uncertainty map.

[0008] It should be appreciated that all combinations of the foregoing concepts and additional concepts discussed in greater detail below (provided such concepts are not mutually inconsistent) are contemplated as being part of the subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are contemplated as being part of the subject matter disclosed herein.BRIEF DESCRIPTION OF THE FIGURES

[0009] The foregoing and other features of the present disclosure will become more fully apparent from the following description and appended claims, taken in conjunction with the accompanying drawings. Understanding that these drawings depict only several implementations in accordance with the disclosure and are therefore not to be considered limiting of its scope, the disclosure will be described with additional specificity and detail through use of the accompanying drawings.

[0010] FIG. 1 is a block diagram of an example system for adaptive sampling, according to some implementations.

[0011] FIG. 2 is a flow diagram of an example method for adaptive sampling, according to some implementations.

[0012] FIG. 3 is a block diagram of an example uncertainty generator, according to some implementations.34932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0013] FIG. 4 is a block diagram of an example model of an example uncertainty generator, according to some implementations.

[0014] FIG. 5 is a block diagram of an example uncertainty estimator of an example uncertainty generator, according to some implementations.

[0015] FIG. 6 is a table of mean and standard deviation of a structural similarity index measure (SSIM) and uncertainty of different acceleration factors, contrasts, and anatomies, according to some implementations.

[0016] FIG. 7 is a table of mean and standard deviations of SSIM and uncertainty of different acceleration factors, coefficients, and anatomies, according to some implementations.

[0017] FIG. 8 are diagrams of a comparison between quantile regression (QR) and residual of the magnitude (ResM) generated uncertainty estimates with a true reconstruction error, according to some implementations.

[0018] FIG. 9 is a chart of a comparison between absolute error, QR-based uncertainty, and ResM-based uncertainty for healthy and abnormal scans, according to some implementations.

[0019] FIG. 10 is a chart of comparison between uncertainty maps generated by an example uncertainty generator for an abnormal scan at different acceleration factors, according to some implementations.

[0020] FIG. 11 is a chart of comparison between uncertainty maps with thresholding for healthy and abnormal cases of different acceleration factors, according to some implementations.

[0021] FIG. 12 is a flow diagram of an example method for acquiring uncertainty maps, according to some implementations.

[0022] FIG. 13 illustrates a computer system for use with certain implementations.

[0023] Reference is made to the accompanying drawings throughout the following detailed description. In the drawings, similar symbols typically identify similar components, unless context dictates otherwise. The illustrative implementations described in the detailed description,44932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)drawings, and claims are not meant to be limiting. Other implementations may be utilized, and other changes may be made, without departing from the spirit or scope of the subject matter presented here. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the figures, can be arranged, substituted, combined, and designed in a wide variety of different configurations, all of which are explicitly contemplated and made part of this disclosure.DETAILED DESCRIPTION

[0024] Magnetic resonance imaging (MRI) is a medical imaging technique used to visualize internal structures in a body. An MRI system uses a magnet to generate a strong, uniform magnetic field, gradient coils enable spatial encoding by varying a strength of the magnetic field, and radiofrequency (RF) coils to both transmit and receive RF pulses. Both the gradient coils and the RF coils are directed towards different parts of the body. Once the MRI system is turned on, the magnet generates a strong magnetic field to align protons in the body and RF pulses transmitted by the RF coils cause the protons to move out of alignment. Once the RF pulse is stopped, the protons return to their original alignment, the RF coils receive the RF waves from the protons, and a computer processes the signals to create detailed contrast images. Contrast images highlight different tissues or structures in the body based off of radiofrequency signal intensity.

[0025] Parallel imaging (PI) is an MRI technique that employs multiple RF coils to reduce scan times and minimize artifacts (e.g., motion artifacts) in the contrast images while maintaining a high spatial resolution. In traditional MRI scanning, every line of k-space data (e.g., spatial frequency representation) is acquired sequentially to obtain a full set of k-space data (e.g., corresponding to a complete contrast image). In contrast, PI leverages the different spatial sensitivity of each RF coil, and collects fewer lines of k-space data, thereby reducing scan time. The missing k-space data is then reconstructed using a reconstruction algorithm to obtain the full contrast image. PI protocols may use a predetermined acceleration factor, which refers to an extent at which scan time is reduced. For example, an acceleration factor of 2 indicates that the scan time is reduced by half compared to traditional MRI scanning. Higher acceleration factors may lead to increased noise and image artifacts. However, the acceleration factor is conventionally54932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)predetermined in clinical protocols, even when the acceleration factor could be increased while still maintaining image quality.

[0026] A majority of MRI systems can acquire (e.g., detect) fully sampled RF signals (k-space) induced by nuclear spins of nuclei (e.g., hydrogen nuclei) using one or more receive antennas-coils (e.g., RF coils) positioned to capture spatially encoded signals of the subject (e.g., patient) being scanned. The RF coils convert the time-varying magnetic fields of the nuclei into electrical signals. The captured signals are digitized and processed through the inverse discrete Fourier transform or similar computational techniques to reconstruct spatially resolved images of the imaged volume, providing detailed anatomical or functional information.

[0027] MRI provides unparalleled soft tissue contrast and plays a central role in modern disease diagnostics. However, the diagnostic value of MRI images is counterbalanced by inherently long acquisition times, which limit patient throughput, increase susceptibility to motion artifacts, reduce patient comfort, and raise operational costs. Accelerated MRI reconstruction techniques, such as PI and compressed sensing (CS) were developed to mitigate these limitations by reconstructing images from undersampled k-space data. For example, the acquisition of the k-space signal can be undersampled to reduce scan time and the MR image can be reconstructed by employing multiple RF coils and image reconstruction techniques using PI, CS (e.g., leveraging sparsity in the k-space signal), or a combination of the aforementioned techniques, among others. The undersampling factor (e.g., acceleration factor) is predetermined during scan time and is typically set to two for traditional PI methods, or up to four for learning-based PI methods rooted on CS principles.

[0028] Although PI and learning-based reconstruction models have accelerated scan times for traditional MRI, the scan can still be time-consuming, which can reduce the subject’s comfort during the exam (e.g., confining structure of the MRI scanner), allow the appearance of motion artifacts in the reconstructed image (e.g., due to the unavoidable patient’s movement during the long exam), and overall increase the cost of MRI exams.

[0029] For certain exams, a higher acceleration factor can still yield diagnostic image quality when combined with a deep learning reconstruction (e.g., for healthy patients, where there are no findings such as pathologies). Despite this, a conservative, predetermined acceleration factor is64932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)used since conventional methods lack capabilities to predict the diagnostic accuracy of the reconstructed image while setting up an imaging protocol at the scanner (e.g., MRI scanner). This is due to the technician operating the scanner assuming that the reconstructed image will always be diagnostic (with the exemption of cases corrupted with motion artifacts, where the scan is repeated) and second, in response to a larger acceleration factor being selected, validation of the image having diagnostic quality or not may be difficult since there is no existing evaluation metric for such a task (e.g., the ground-truth image is unavailable).

[0030] Despite the potential for faster scanning times, such acceleration levels are not routinely implemented in clinical MRI protocols. The primary limiting factor is the absence of an automatic mechanism to assess the reliability of reconstructed images when the fully sampled reference is unavailable. Without a quantitative measure of confidence, increasing the acceleration factor beyond standard values is risky, as undersampling may lead to hallucinations or to pathological feature suppression, thereby compromising diagnostic accuracy. Consequently, current clinical practice favors conservative, fixed acceleration factors, typically set to 2 for non-learning-based reconstructions and to 4 for learning-based ones. To enable higher accelerations, such as 6 or greater, while preserving diagnostic validity, reconstruction error may be estimated directly from undersampled data. Uncertainty quantification provides a principled framework toward direct estimation of the reconstruction error from undersampled data by estimating confidence intervals for reconstructed images and effectively producing an error map when the ground-truth reference image is unavailable.

[0031] Attempts to quantify uncertainty in MRI reconstruction were based on probabilistic formulations of deep reconstruction networks. For example, a variational autoencoder framework that learns to predict pixel-wise residual magnitudes for a pre-trained reconstruction network was used to quantify the uncertainty. However, such an approach remains heuristic and lacks statistical guarantees on the predicted uncertainty levels. Another example includes a Bayesian heteroscedastic uncertainty framework for neuroimage enhancement. Such methods do not yield calibrated or distribution-free confidence intervals. Other methods may include a Bayesian variational formulation of the learned variational network, in which the parameters of the regularizer are modeled as random variables drawn from a learned multivariate Gaussian distribution. Sampling these parameters enables estimation of pixel-wise variance maps that reflect74932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)regions of high uncertainty. While this framework offers a theoretically grounded means of visualizing uncertainty, such method incurs substantial computational overhead and does not ensure calibrated or interpretable confidence intervals. Other methods may reformulate MRI reconstruction as a pixel-classification problem, where uncertainty is derived from the variance of a softmax output distribution. Such approaches are heuristic and lack statistical calibration or formal coverage guarantees.

[0032] Other conventional methods may use the magnitude of the residual (ResM) as an uncertainty estimator. However, ResM yields symmetric confidence bounds that limit interpretability and can inflate the estimated uncertainty range. Moreover, the residuals learned during training are often smaller than those encountered during testing, leading to unreliable uncertainty estimates that correlate poorly with the true reconstruction error.

[0033] Embodiments described herein relate generally to dynamically adjusting an undersampling factor (e.g., acceleration factor) based on real-time uncertainty estimates generated during MRI scans. As such, systems and methods of the present disclosure reduce scan time and mitigates risks of motion artifacts captured in scans associated with prolonged scan times.

[0034] The systems and methods of the present disclosure as described further herein include a computer system that reconstructs an MR image using learning-based image reconstruction methods and determines uncertainty estimates for that image using undersampled MR signals. The uncertainty estimates may be computed using statistical techniques (e.g., conformal prediction). Conformal prediction can be used to provide statistical guarantees on the uncertainty intervals, which can herein be referred to also as confidence intervals. Uncertainty quantification methods such as quantile regression offer a principled way to estimate conditional quantiles of the reconstruction error. These quantiles may not ensure valid coverage alone. Conformal prediction addresses this discrepancy by using a held-out calibration set to compute a scaling factor that adjusts the quantiles sot the derived uncertainty intervals so achieve the desired coverage level (e.g., 90%) in finite samples. This combination leverages the structural modeling capacity of heuristic notions of uncertainty while ensuring that the final uncertainty bounds are statistically valid and calibrated for all acceleration factors.84932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0035] The computer system may include an interface device that communicates with a measurement device, a processor that executes program instructions, and memory. During operation, a computer system may acquire undersampled MR signals associated with a sample from the measurement device or memory. Then, the computer system may reconstruct the associated sample’s image along with an uncertainty estimate using a learning -based image reconstruction method and a statistical technique, respectively. The uncertainty estimate may be processed by the computer system, through at least one algorithm, to evaluate the clinical usability (e.g., validity) of the associated image. In response to the algorithm deeming the image non-clinically valid (e.g., determined by a threshold value, which is pre-determined based on radiologist evaluations on existing data), the computer system may acquire additional undersampled MR signal and combine the signal with the previous undersampled MR signal associated with the same sample from the same measurement device or memory and repeat the aforementioned process until the associated sample’s image is deemed clinically valid (e.g., based on the at least one algorithm) or a predetermined overall acceleration factor is reached. The computational overhead of the image reconstruction and uncertainty estimation may be negligible compared to the overall scan time. The at least one algorithm may be designed based on radiologist input and be tailored to specific MRI exams. For example, learning-based reconstruction methods are able to accurately reconstruct healthy tissues but may struggle to capture all features around pathologies for higher acceleration factors (e.g., 8). As another example, the algorithm may be one or more algorithms for specific MRI exams, such as a first algorithm for knee MRI scans and a second algorithm for elbow MRI scans. The algorithm may look for areas of the highest uncertainty in the reconstructed sample’s image. In response to a weighted average uncertainty of these areas being higher than a threshold value, the reconstructed sample’s image may be deemed not clinically valid (e.g., diagnostically reliable). In response to the weight average uncertainty being below the threshold, the reconstructed sample’s image is deemed clinically valid. The threshold value may be determined based on a multiple radiologists scoring process (e.g., Likert scoring system on existing data).

[0036] By estimating the clinical usability of the subject’s reconstructed image using higher acceleration factors, the systems and methods of the present disclosure may reduce the MRI scan time for measuring MR signals. For example, in response to the algorithm determining that the image is clinically valid at an acceleration factor which is higher than the conventional94932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)predetermined acceleration factor, then unnecessary measurements made by the measurement device may be mitigated (e.g., one may reconstruct a clinically valid image using one fifth of the MR signal, comparing to one fourth that is usually used for learning-based reconstruction methods). The systems and methods of the present disclosure may achieve patient-specific limits for the possible acceleration factor in particular MRI examinations, which may be higher than the standard predetermined acceleration factor, optimizing both the scan time, the diagnostic quality of the image, the exam’s cost, and may improve the overall patient’s experience.

[0037] A second issue addressed by the innovation is the lack of real-time adaptability in conventional MRI methods, which still rely on an MR technologist to ensure proper image quality. Currently, there is no way to determine whether an MRI is clinically usable until the scan is complete. At that point, the technologist decides whether the image quality is poor, and the scan needs to be repeated (e.g., with a smaller acceleration factor). This manual procedure can increase scan time due to suboptimal image acquisition settings. In addition, the technologist might not be able to assess the quality of subtle pathologies in the image without the radiologist’s inputs, which usually come a few hours after the scan time. Thus, the MRI scan may need to be performed again, increasing costs and time for obtaining MR images. The systems and methods of the present disclosure resolve these issues by automatically performing real-time error prediction during the scan which mitigates the need for human intervention. This automation streamlines the workflow and ensures the collected images meet diagnostic standards (otherwise the uncertainty estimation will always be above the threshold value).

[0038] The systems and methods of the present disclosure include a statistically rigorous uncertainty estimation framework based on conformal QE that can be incorporated into an image reconstruction method which was validated on brain MRI reconstructions using a pre-trained end-to-end variational network Apart from qualitatively matching the spatial distribution of reconstruction errors, the QR-based uncertainty maps correlate strongly with the true error, achieving a Pearson correlation coefficient of, for example but not limited to, 0.91 at 4-fold acceleration, compared to 0.69 with ResM. The systems and methods enable reliable assessment of reconstruction accuracy without knowledge of the ground-truth across multiple acceleration factors. The framework is designed to estimate the reliability of reconstructed images for any acceleration factor. The systems and methods of the present disclosure directly estimate the104932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)conditional quantiles of the reconstructed image distribution, thereby producing statistically valid and spatially resolved uncertainty intervals.

[0039] The systems and methods of the present disclosure further include a framework for pixel-wise uncertainty quantification in parallel MRI reconstruction, enabling automatic identification of unreliable regions without the ground-truth reference image. The systems and methods integrate conformal quantile regression with image reconstruction methods, such as an end-to-end Variational Network to estimate statistically rigorous pixel-wise uncertainty intervals.A model to estimate the intervals was trained and evaluated on Cartesian undersampled brain and knee data obtained from a training dataset using acceleration factors ranging from 2 to 10. Quantitative experiments demonstrate strong agreement between the predicted uncertainty maps and the true reconstruction error. For example, the corresponding Pearson correlation coefficient was higher than 90% at four-fold accelerations and above, whereas the correlation coefficient dropped to less than 70% when the uncertainty was computed with a simpler notion (e.g., magnitude of the residual (ResM)) for brain reconstructions. Qualitative examples further show the uncertainty maps capture both the magnitude and spatial distribution of reconstruction errors across acceleration factors, with regions of elevated uncertainty aligning with pathologies and common MRI artifacts. The systems and methods of the present disclosure enable evaluation of reconstruction quality when fully-sampled ground-truth reference images are unavailable, thereby facilitating adaptive MRI acquisition protocols that dynamically balance scan time and diagnostic reliability. The systems and methods of the present disclosure further enable usage of heuristic notions of uncertainty to adjust the acceleration factor during acquisition. By adjusting the amount of samples based on the certainty, the systems and methods enable real-time, uncertainty-aware optimization of scan efficiency, leading to reduction in scan time while preserving diagnostic reliability.

[0040] FIG. 1 is a block diagram of an example system 100 for reconstructing a MR image and estimating at least one uncertainty value for the MR image. The system 100 includes at least one MRI system 102 (e. g. , MRI machine, apparatus) . The MRI system 102 includes a patient table to position the patient for an MRI scan and uses a magnet as well as gradient coils and RF coils to spatially encode RF signals to excite (e.g., move) nuclei and detect signals emitted by the excited nuclei as the nuclei return to an aligned state. The MRI system 102 can detect and receive signals114932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)emited by the nuclei as the nuclei move from an unaligned state back to the aligned state. The signals may be received by the RF coils. In some implementations, the MRI system 102 is a medical imaging system 102. For example, the medical imaging system 102 can be at least one of an MRI system or a computed tomography (CT) system, among others.

[0041] The system 100 includes a computing system 104 (e.g., controller) communicatively coupled to the MRI system 102. The computing system 104 includes one or more processors 106 and at least one memory 108, which can be implemented as one or more processing circuits. The processor 106 may be a general purpose or specific purpose processor, an application specific integrated circuit (ASIC), one or more field programmable gate arrays (FPGAs), a group of processing components, or other suitable processing components. The processor 106 may be configured to execute computer code or instructions stored in memory (e.g., fuzzy logic, etc.) or received from other computer readable media (e.g., CDROM, network storage, a remote server, etc.) to perform one or more of the processes described herein. The memory 108 may include one or more data storage devices (e.g., memory units, memory devices, computer -readable storage media, etc.) configured to store data, computer code, executable instructions, or other forms of computer-readable information. The memory 108 may include random access memory (RAM), read-only memory (ROM), hard drive storage, temporary storage, non-volatile memory, flash memory, optical memory, or any other suitable memory for storing software objects and / or computer instructions. The memory 108 may include database components, object code components, script components, or any other type of information structure for supporting the various activities and information structures described in the present disclosure. The memory 108 may be communicably connected to the processor 106 and may include computer code for executing (e.g., by processor 106) one or more of the processes described herein. The memory 108 can include various modules (e.g., circuits, engines) for completing processes described herein. The one or more processors 106 and memory 108 may include various distributed components that may be communicatively coupled by wired or wireless connections; for example, various portions of the computing system 104 may be implemented using one or more client devices remote from one or more server devices. The computing system 104 can include any one or more rules, heuristics, logic, code, functions, machine learning models, neural networks, algorithms, or various combinations thereof to implement one or more components of the computing system 104, such as signal receiver 110, image generator 112, and uncertainty generator 114. The computing124932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)system 104 and / or various components thereof can execute various operations described herein and / or combinations thereof as one or more tasks. For example, the signal receiver can cause the processor 106 to execute a signal receiving task and the image generator 112 can cause the processor 106 to execute an image generating task.

[0042] The computing system 104 controls the MRI system 102, and provides scanning protocols, such as the acceleration factor, parts of the body to scan, etc. For example, the computing system 104 sets the MRI system 102 to a first acceleration factor to enable PI of the MRI scan. The scanning protocol changes based on a type of MRI scan being performed, such as a scan of a knee or scan of an elbow. The computing system 104 can select and adjust the scanning protocol in response to user (e.g., imaging technician) input. The computing system 104 also sets scan times of the MRI system 102. Each parameter of the scanning protocol may be adjusted by a user using the computing system 104. The computing system 104 may dynamically adjust the first acceleration factor during the MRI scan, as described further herein.

[0043] The computing system 104 includes at least one signal receiver 110. The signal receiver 110 receives the signal detected and collected by the RF coils of the MRI system 102 in response to the nuclei moving to their aligned state. The signal receiver 110 receives signals from multiple RF coils in parallel (e.g., simultaneously) when the MRI system 102 is performing PI. As such, the signal receiver 110 can associate (e.g., correspond) each received signal to its respective RF signal. The signal can be an undersampled k-space signal (e.g., acquiring fewer k-space data points than needed for a full MR image). The signal may be undersampled in response to the MRI system 102 operating at the first acceleration factor. The signal receiver 110 receives signals from the MRI system 102 during the MRI scan process. For example, the signal receiver 110 receives the signals from the MRI system 102 in real-time as the RF receiver coils of the MRI system 102 receive the signals. The signal receiver 110 can associate the signals with the acceleration factor at which the MRI system 102 is emitting and receiving the signals.

[0044] The computing system 104 includes at least one image generator 112. The image generator 112 receives the signals from the signal receiver 110 and reconstructs (e.g., outputs, generates) the MR image based on the undersampled k-space signals. The image generator 112 includes at least one reconstruction model (e.g., algorithm, machine learning model) to generate134932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)the MR image based on the undersampled k-space signals. The reconstruction model may be a PI reconstruction model and may be a machine learning model. The reconstruction model may apply at least Fourier transforms to convert the signals into an image. In some implementations, the reconstruction model is a deep learning model, such as a convolutional neural network (CNN) or any other network, and predicts missing k-space data to reconstruct the MR image. The reconstruction model can also combine the received signals and reconstruct the MR image using the spatial sensitivity profiles of each corresponding RF coil.

[0045] The computing system 104 includes at least one uncertainty generator 114. The uncertainty generator 114 receives the image from the image generator 112 and generates an uncertainty (e.g., confidence intervals) map with dimensions equal to the image. The uncertainty generator 114 includes at least one neural network to generate the uncertainty map. The uncertainty generator 114 may determine the uncertainty map by determining (e.g., calculating, generating) uncertainty values such as a signal-to-noise ratio (SNR), spatial resolution, number of artifacts, degree of undersampling, pixelwise quantile regression, or signal decay, among others, per region (e.g., pixels) of the image. The uncertainty map thus includes uncertainty value (e.g., region uncertainty value) per region of the image. The uncertainty map identifies a confidence that regions of the map are clinically valid (e.g., interpretable, accurate). For example, regions of the image with pathologies may have a higher uncertainty than regions of the image containing healthy tissue. As another example, the uncertainty map may identify regions of the image that contain artifacts or regions of low signal (e.g., as determined by the scanning protocol) with minimal diagnostic interest (e.g., edges of a skull) with higher uncertainty.

[0046] Following generation of the uncertainty map, the uncertainty generator 114 determines areas of the map with higher uncertainties than others. For example, the uncertainty generator 114 determines areas of the map with higher clusters of uncertainty, such as areas of the image with pathology. The areas of the image with pathology may contain multiple regions of higher uncertainty compared to areas of the image with healthy tissue. The uncertainty generator 114 can segment the map by identifying the areas of the image with higher uncertainties. Once these areas are determined, the uncertainty generator 114 determines an average uncertainty value of each of the areas by calculating a weighted average of the detected uncertainties values in the regions of each area. For example, the uncertainty generator 114 determines the uncertainty values for each144932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)region in the area and then determines the weighted average based on all of the uncertainty values. Once the average uncertainty value is calculated, the uncertainty generator 114 compares the uncertainty value to a threshold. Given multiple areas with higher uncertainty, the uncertainty generator 114 determines the average uncertainty value for each area and compares each of the average uncertainty values to the threshold. The threshold is predetermined and is associated with an amount of uncertainty that the image can contain while still being clinically valid (e.g., radiologist can accurately interpret the image).

[0047] Responsive to the uncertainty value (e.g., any of the uncertainty values) being at or below the threshold, the computing system 104 controls the MRI system 102 to maintain (e.g., resume performing) the MRI scan at the first acceleration factor. The uncertainty value being at or below the threshold indicates that the first acceleration factor and the degree of undersampling is sufficient for generating clinically valid images. Thus, no change to the MRI scan is needed. Responsive to the uncertainty value (e.g., any of the uncertainty values) being above the threshold, the uncertainty generator 114 instructs the computing system 104 to change the first acceleration factor to a second acceleration factor. The second acceleration factor is lower than the first acceleration factor. For example, the first acceleration factor may be 8 and the second acceleration factor may be 4. The uncertainty value being above the threshold indicates that the uncertainty is too high and images generated by the image generator 112 from the undersampled signal are not clinically valid. The signal receiver 110 then begins receiving signals at the lower acceleration factor, and the image generator 112 combines the signals from the first acceleration factor and the second acceleration factor to generate the MR image. For example, the signals at the second acceleration factor supplement information from the signals at the first acceleration factor. The uncertainty generator 114 can then repeat generating the uncertainty value until the uncertainty value is at or below the threshold. For example, responsive to the uncertainty generator 114 determining that the uncertainty value is still above the threshold responsive to the MRI system 102 operating at the second acceleration factor, the uncertainty generator 114 may instruct the computing system 104 to change the MRI system 102 to operate at a third acceleration factor lower than the second acceleration factor to obtain more detailed signals.

[0048] In some implementations, the computing system 104 has a minimum acceleration factor. For example, once the signal receiver 110 determines that the signals received are at the154932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)minimum acceleration factor, the signal receiver provides the signals to the image generator 112 to generate the images, but the image is not provided to the uncertainty generator 114. The signals provided at the minimum acceleration factor indicate that the images produced by the image generator 112 will be clinically valid. For example, the signal receiver 110 can compare the acceleration factor of the signals received to a second threshold. Responsive to the acceleration factor being equal to the second threshold, the signal receiver 110 indicates to the image generator 112 to not provide the images to the uncertainty generator 114. The uncertainty generator 114 may adjust the acceleration factor to be the minimum acceleration factor (e.g., 2) following one or more rounds of generating the uncertainty map and values at higher acceleration factors (e.g., 4, 6, 8, 10).

[0049] In some implementations, the uncertainty generator 114 determines which acceleration factor to adjust the MRI system 102 to based on the uncertainty value. For example, different ranges of uncertainty values may be associated with each acceleration factor, and the uncertainty generator 114 can instruct the computing system 104 to change the first acceleration factor to a second acceleration factor based on which range the uncertainty value falls into. In other implementations, the uncertainty generator 114 sequentially adjusts the first acceleration factor in response to determining that the uncertainty value is above the threshold. For example, the uncertainty generator 114 adjusts the acceleration factor from 8 to 6. In some implementations, the acceleration factor is increased. The uncertainty generator 114 adjusts the acceleration factor of the MRI system 102 during a scan. For example, the signal receiver 110 receives the first signal associated with a first acceleration factor and the second signal associated with a second acceleration factor during one scan performed by the MRI system 102. The computing system 104 can thus adapt scanning protocols of the MRI system 102 in real-time depending on results of the image reconstruction generated from the signals provided by the MRI system 102.

[0050] In some implementations, the uncertainty generator 114 includes a plurality of neural networks. Each of the plurality of neural networks is associated with at least one of a type of MRI scan or an acceleration factor. For example, a first neural network is associated with an acceleration factor of 8 and a second neural network is associated with an acceleration factor of 6. In this case, prior to generating the uncertainty map, the signal receiver 110 and / or the image generator 112 provides the acceleration factor to the uncertainty generator 114, and the uncertainty generator 114164932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)determines the type of MRI scan being performed based on, for example, the scanning protocol. The uncertainty generator 114 then selects a neural network for the image output by the image generator 112 to be input into based on at least one of the acceleration factors or the type of MRI scan.

[0051] The system 100 may be implemented as described further herein. A neural network of the system 100 (e.g., the image generator 112) can receive an input of the undersampled k-space signal and output a reconstructed image. The network can be supervised, or self-supervised, and can be trained with an existing dataset (e.g., of MR images) The network can achieve a diagnostic accuracy (e.g., produce clinically valid images) for four-fold accelerations (e.g., acceleration factor of 4), and in some cases, is clinically valid at six-fold and eight-fold accelerations. The network can be trained using a loss function including at least two terms. The first term can be a distance measure (e.g., pixel-wise or perceptual) between the ground-truth (e.g., diagnostic) image (x) and the reconstructed image (x) (e.g., structural similarity index measure (SSIM)). The second term can be a heuristic notion of uncertainty, for example, magnitude of the residual or pixelwise quantile regression. The loss function can be expressed as the average of the following:

[0052] [u(x) - SSIM(x.x)]2,

[0053] where, u(x) is a map of the upper (and lower) confidence interval length (the uncertainty) of the MR image. A number of different model versions (e.g., networks) are trained corresponding to different acceleration factors (e.g., 10-fold, 8-fold, 6-fold, 4-fold).

[0054] In some implementations, the estimated uncertainty map may be provided as an additional input to a neural network to enhance reconstruction performance. For example, the uncertainty map can be concatenated with intermediate feature representations, used to modulate convolutional activations, or incorporated into attention or gating mechanisms. By supplying uncertainty information directly to the model, the network can adapt processing to regions with high estimated error, focusing reconstruction resources on challenging anatomical structures or artifact-prone areas.

[0055] In some implementations, the estimated uncertainty map can also be used as a weighting mask when computing the loss between the reconstructed image and the reference174932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)image. For example, the uncertainty values can modulate pixel- wise contributions to metrics such as SSIM, MSE, or perceptual losses. Regions with higher predicted uncertainty can receive increased weighting during training, allowing the network to focus more on areas that are difficult to reconstruct or prone to error. This uncertainty-weighted loss formulation enables targeted learning in clinically relevant regions while reducing the influence of well-reconstructed areas. In addition, the uncertainty map can be incorporated directly into the loss definition itself to address recognized limitations of SSIM and related metrics in both image evaluation and neural -network training.

[0056] In some implementations, the uncertainty can also serve as a standalone output used for anomaly detection or triage. Because the uncertainty values reflect the model’s predicted reconstruction error, the map can indicate regions likely to contain pathologies, other abnormalities, or unexpected anatomical patterns. Unlike conventional anomaly-detection methods that typically degrade with increased undersampling, the uncertainty-map-based detector can become more sensitive to anomalies as the acquired data become sparser, thereby enabling use of the uncertainty maps as an automated flagging mechanism or prioritization tool in accelerated MRI workflows.

[0057] The confidence interval length is a map with dimensions equal to the image. Since the network can achieve good (e.g., clinically valid) reconstruction for healthy tissues, larger confidence intervals in regions with pathologies can be expected (e.g., since pathologies are sparse in the training data). In addition, large uncertainty in regions of low signal may exist that have minimal diagnostic interest (e.g., towards the edges of a skull).

[0058] At least one algorithm (e.g., the uncertainty generator 114) can detect a few small regions (number of areas will be determined through a cross-validation study for each type of MRI exam) that have high uncertainty (e.g., areas of pathologies compared to areas with healthy tissue). The area of each region should capture pathologies and artifacts and avoid areas of the image with lower diagnostic value (e.g., healthy tissue areas). To update (e.g., train) the algorithm, a weighted average of the uncertainties in the detected regions can be computed for each image in a predetermined test dataset (e.g., where the ground-truth data is also available) for all versions of the model (e.g., for each acceleration factor, type of MRI exam, etc.) A radiologist study with184932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)multiple board-certified radiologists can be performed to determine which minimum uncertainty value leads to clinically valid results for diagnostic accuracy. The minimum uncertainty value can be set as a threshold value for the algorithm.

[0059] The uncertainty map output by the uncertainty generator 114 can indicate or otherwise include at least one characteristic. The at least one characteristic can include at least one of the uncertainty map being at or below, or above a threshold, a detected anomaly (e.g., using the uncertainty), or a severity of a condition of an individual corresponding to the uncertainty map. Based on the characteristic, the computing system 104 can determine to at least one of adjust an acceleration factor of the MRI system 102, generate a notification regarding the anomaly, or order the individual according to the severity, such as in a line of individuals in triage. For example, the computing system 104 can determine an anomaly (e.g., pathologies, etc.) based on a region of the uncertainty map having an uncertainty greater than a threshold, as described herein. The presence of such anomalies can indicate a higher severity of the condition of the individual, causing the computing system 104 to order the individual at a greater priority in triage.

[0060] During scan time of a new subject (e.g., in vivo testing of the model), a, for example, 10-fold undersampled k-space signal is collected and is given as an input to the model (e.g., reconstruction model). The model reconstructs the image and estimates the uncertainty value using the algorithm. In response to the uncertainty value being above the threshold, an additional signal equal to, for example, 1 / 40 of the total signal is collected (e.g., the initially acquired 1 / 10 undersampled signal and the new additional signal combines to be overall an 8-fold undersampled signal). The process is repeated with the 8-fold undersampled signal as an input and so on until the estimated weighted average uncertainty is below the threshold value, or a minimum acceleration factor is reached (e.g., which is known to be clinically valid). The model is optimized and runs rapidly on a graphical processing unit, therefore the uncertainty estimation and image reconstruction will be thousands of times faster than the actual scanning time, and thus will not introduce any additional time burden, even if a four -fold acceleration factor (e.g., current standard acceleration) has to be used to achieve a clinical reconstruction.

[0061] Different versions of the model may be used for different MRI exams to ensure optimal performance as the uncertainty estimates will be evaluated from different radiologists each time.194932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0062] The incorporation of the additional signal can be easily implemented with available data that emulate the MR measurements. For a practical validation in the scanner, sampling protocols may need to be altered.

[0063] The systems and methods of the present disclosure for image reconstruction represent a significant departure from traditional MRI protocols, where undersampling rates are predetermined and fixed across all scans. By dynamically adjusting the undersampling factor based on real-time uncertainty estimates, the systems and methods of the present disclosure address the limitations of current practices, including the inability to assess diagnostic image quality during the scan itself. The introduction of learning-based reconstruction combined with statistical uncertainty quantification offers a reliable framework for ensuring clinical usability while optimizing scan efficiency.

[0064] This approach is particularly valuable in scenarios where reducing scan time is critical, such as pediatric or uncooperative patients, or in high-throughput clinical settings. It also mitigates the risks of motion artifacts associated with prolonged scans. The integration of radiologists’ expertise into the determination of uncertainty thresholds ensures the robustness of the proposed system and aligns it with clinical diagnostic standards.

[0065] The systems and methods of the present disclosure can also be used for other imaging modalities, such as CT scans. For example, the algorithm can be trained on CT imaging data and generate uncertainty maps on CT images. The systems and methods can be further integrated with advanced machine learning techniques, such as federated learning, to facilitate the deployment of the systems and methods across multiple institutions while respecting patient privacy. Generalized, uncertainty quantification models tailored to MRI can also be developed to facilitate generation of the uncertainty maps.

[0066] Ultimately, the systems and methods align with the broader goal of precision medicine, enabling patient-specific imaging protocols that maximize diagnostic quality while minimizing discomfort and cost. By reducing scan time without compromising image quality, the systems and methods have the potential to redefine the patient experience and broaden the accessibility of MRI as a diagnostic tool.204932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0067] FIG. 2 is a flow diagram of an example method 200 for adaptive sampling for PI, according to some implementations of the present disclosure. The method 200 can be performed using various systems described herein. Various steps in the method 200 may be repeated, omitted, performed in various orders, or otherwise modified. Various steps in the method 200 may be run concurrently, in parallel, or individually. The method 200 may be implemented by at least one or more processors.

[0068] The method 200 includes, at block 202, receiving a signal associated with a first value. The signal may include a first portion of MRI data and may be received during an MRI scan. In some implementations, the first value is a first acceleration factor. In some implementations, prior to block 202, the method 200 includes performing a scan using an MRI machine at a first acceleration factor. The method 200 includes, at block 204, generating an image. The image is an MR image and is generated using at least one model and the signal. The at least one model may include a reconstruction model, and the image includes a second portion of the MRI data. The second portion of the MRI data can include the first portion. For example, the first portion of the MRI data may be an undersampled dataset while the second portion is generated by the at least one model and represents the fully sampled dataset.

[0069] The method 200 includes, at block 206, generating a map corresponding to the image. The map is generated using the at least one model which may include at least one neural network. The at least one model can include a reconstruction model to generate the image and at least one neural network to generate the map. The map may be an uncertainty map. The method 200 includes, at block 208, determining an uncertainty value based at least partially on the map. The uncertainty value is determined using the at least one model.

[0070] In some implementations, the map includes a plurality of regions associated with regions of the image, each of the plurality of regions associated with a region uncertainty value. The region uncertainty value can be determined by at least one of a magnitude of a residual or a pixelwise quantile regression. The at least one uncertainty value can be an average uncertainty value and to determine the at least one uncertainty value, the method can further include segmenting the plurality of regions into a plurality of areas based on respective region uncertainty values. The method can also include combining, in each of the plurality of areas, the region214932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)uncertainty values to determine an area uncertainty value and determining, by the one or more processors, the average uncertainty value for at least a first area of the plurality of areas with a highest area uncertainty value.

[0071] The method 200 includes, at block 210, comparing the uncertainty value to a threshold. At block 212, responsive to the value being above the threshold, adjusting the first value to a second value. The second value can be a second acceleration factor. At block 214, the MRI scan resumes at the second value responsive to adjusting the first value to the second value. At block 216, the MRI scan is resumed at the first value responsive to the uncertainty value being at or below the threshold.

[0072] In some implementations, the signal is a first signal, the image is a first image and the uncertainty value is a first uncertainty value. The method can further include receiving a second signal associated with the second value including the first portion and a third portion of the MRI data. The method can also include combining second signal and the first signal into a third signal and generating, using the at least one model and the third signal, a second image including a fourth portion of the MRI data. The fourth portion can include the first portion and the third portion. In some implementations, the map is a first map and the threshold is a first threshold. The method can further include comparing the second value to a second threshold corresponding to a third value, the third value less than the first value. The method also includes resuming the MRI scan at the second value responsive to the second value being equal to the second threshold and generating, using the at least one model, a second map corresponding to the second image responsive to the second value being greater than the second threshold.

[0073] In some implementations, the method further includes determining a distance value based on the image and a training image corresponding to the image, the training image comprised in a training dataset. The method also includes determining a loss value as a function of the map and the distance value and updating the at least one model based on the loss value. The at least one model can include a plurality of neural networks, each neural network corresponding to at least one of a type of MRI scan or an acceleration factor. For example, a first neural network corresponds to an acceleration factor of 8 and a second neural network corresponds to an acceleration factor of 4.224932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0074] In some implementations, the method 200 does not include generating the uncertainty value and, instead, compares the uncertainty map to the threshold. In some implementations, the first signal is received from a medical imaging system that may include an MRI system, but can also include other imaging modalities, such as CT. The at least one neural network can be updated using federated learning. The threshold can correspond to a minimum uncertainty value. The minimum uncertainty value corresponds to a minimum uncertainty value for radiologists to accurately interpret the image and provide a diagnosis based on the image.

[0075] In some implementations, to generate the uncertainty map, as shown in at least FIG. 3, the uncertainty generator 114 includes a model 302, such as an end-to-end variational network. The model 302 can be pre-trained according to a pinball loss for QR on a training dataset including images captured with 2-, 4-, 6-, 8-, and 10-fold acceleration factors. Following training, the weights of the model 302 can be frozen. The uncertainty generator 114 can attach an uncertainty estimator 304 to the model 302. The uncertainty estimator 304 can include two convolutional neural networks (e.g., U-Net, etc.) that predict the lower and upper confidence interval quantiles of the reconstruction, denoted as x and ft(x) , respectively. The uncertainty estimator 304 can ensure consistent dynamic range and stable uncertainty estimation across all acceleration factors.

[0076] At least one of the model 302 or the uncertainty estimator 304 is updated on a training dataset. The training dataset includes images of at least the brain and knee captured with various acceleration factors. The training dataset can be divided into subsets based on whether the image depicts a brain or a knee. The subsets can be randomly divided into subsets for validation and calibration. In some implementations, such as for the knee subsets, the subsets can be divided for validation, calibration, and testing.

[0077] Cartesian undersampling was used for the training of at least one of the brain or knee. For the subsets depicting the brain, 16%, 8%, 5.3%, 4%, or 3% of the central k-space was retained as autocalibration signal (ACS) lines and the remaining k-space was uniformly sampled to achieve acceleration factors of 2, 4, 6, 8, and 10, respectively. At least one of the model 302 or the uncertainty estimator 304 were trained and tested using fixed undersampling masks, resulting in multiple separate models for reconstruction. For the subsets depicting the knee, 8% of the central234932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)k-space was retained as ACS, and the remaining k-space was uniformly sampled to achieve an acceleration factor of 4.

[0078] The input of the model 302 can be undersampled multi-coil k-space (e.g., Cartesian undersampling, etc.) 306 and the output is the reconstruction 308 (e.g., reconstructed MRI image). FIG. 4 depicts an example of the model 302. In each cascade 310, each undersampled k-space 306 can transform to images using the inverse fast Fourier transform (iFFT), and the resulting images are weighted with the corresponding coil sensitivities 402 and are combined into one using a reduce operator. The resulting image can be processed through a convolutional neural network 312. The output can expand (e.g., be separated) into the individual coil images, which are transformed to k-space using the FFT. Data consistency can follow, or proceed in parallel with the transformation to the k-space. The final image (e.g., the reconstruction 308) may be obtained using the root sum of squares (RSS) on (e.g., by combining) the individual coil images obtained from the 12thcascade 310.

[0079] A total of 12 cascades 310 can be used in the model 302. In each cascade 310, the model 302 can include at least one convolutional neural network 312 using 32 feature channels, 4 layers of average pooling and transpose convolutions, each with a kernel size, stride, and padding of 2, 2, and 0, respectively. The convolution layers can have a kernel size of 3 with both padding and stride set to 1. A rectifier activation functions can be used with a negative slope of 0.2. The convolutional neural network 312 for the sensitivity estimation can have the same parameters as other convolutional neural networks 312 except the number of channels which was reduced to 8. The input tensors to all convolutional neural networks 312 were normalized to ensure that each channel had a mean of 0 and a standard deviation of 1. A batch size of 1 was used and trained on a high-performance computing cluster using four graphics processing units (GPUs), such as NVIDIA Al 00 Tensor Core GPUs, each with 80 GB of memory. The loss function can be the SSIMto update the model 302. Training was performed using an optimizer with a learning rate of 0.0003 for 210,000 iteration steps. A warm-up ramp was applied during the first 7,500 steps, followed by cosine annealing after 140,000 steps. The overall number of trainable parameters can be 93.6 million and the training times can be approximately two and a half days for both the brain and knee subsets.244932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0080] During training of the uncertainty estimator 304, the weights of the model 302 are frozen, and the reconstruction 308 generated by the model 302 is the input to the convolutional neural networks 314 of the uncertainty estimator 304. The convolutional neural networks 314 of the uncertainty estimator 304 can have the same parameters as the convolutional neural networks 312 of the cascades 310 of the model 302 aside from a number of input and output channels being 1. Each convolutional neural network 314 output passes through a sigmoid activation to scale the values between 0 and 1. Following scaling, the output is multiplied by the model 302’ s reconstruction to match a same intensity range as the reconstructed image 308 generated by the model 302. For example, the output can be added to (e.g., multiplied) or subtracted (e.g., divided) from the reconstructed image 308 to compute the upper 316 and lower bound 318, respectively. The resulting outputs of the uncertainty estimator 304 are used as the lower 316 or the upper quantile interval 318.

[0081] To calibrate the uncertainty estimator 304 following training, the predicted offsets are scaled by the calibration factor X which can be computed using conformal prediction in the calibration set. The uncertainty map 320 output by the uncertainty estimator 304 can be computed by subtracting the calibrated lower 318 from the calibrated upper bound 316. In situations where symmetric upper 316 and lower quantiles 318 are desired, the uncertainty estimator 304 can use a single convolutional neural network 314.

[0082] After calibration, the predicted quantiles 316, 318 define the uncertainty 320. Assuming an a-level uncertainty interval of the reconstructed image (x) 308 generated by the model 302, the upper quantile ii (x) 316 estimates the (1 — cr / 2) conditional quantile, and the lower quantile Z(x) 318 estimates the (cr / 2) conditional quantile. The uncertainty estimator 304 can use a pinball loss for quantile a (e.g., predicted quantile, pixel wise quantiles) which can be:

[0084] Here, x is the reconstruction 308 generated by the model 302, y is the ground-truth image, and qa(%)) is the predicted a -quantile of y given x . Unlike ResM, QR asymmetrically penalizes underestimation and overestimation according to a. Since the upper 316 and lower254932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)quantiles 318 correspond to different conditional quantiles, the upper 316 and lower quantiles 318 can be trained using separate losses. The total loss for the interval prediction can be:

[0086] After training, the lower 318 and upper quantile 316 estimates can be expected to converge asymptotically to the true conditional (cr / 2)-and (1 — cr / 2)-quantiles, respectively.

[0087] Referring now to FIG. 5, training can be performed in a supervised manner using (x, y) pairs. The uncertainty estimator 304 can be trained separately from the model 302, using the same optimizer, learning rate, scheduler, and hyperparameters. During training 502, weights of the model 302 can either be kept frozen and used to generate reconstructions 308, or the reconstructions 308 generated by the model 302 can be precomputed for training and validation data and loaded from memory, a can be set to 90% for training and experimental iterations. The number of trainable parameters can be 15.5 million and the training time can be approximately one day for both knee and brain subsets.

[0088] For comparison, in the ResM-based approach, u(x) = Z(x) is assumed. The loss function 504 used in this case is:

[0089] T(x,y) = (u(x) - |x - y|)2.

[0090] The quantile estimates I and ii 316, 318 may be calibrated to be statistically valid, meaning that the resulting uncertainty intervals may contain at least a fraction (1 a) of the groundtruth pixel values with probability not smaller than (1 - a). Neural networks trained to predict quantiles 316, 318 may produce intervals that are miscalibrated, either too narrow or too wide. To guarantee valid coverage, conformal prediction calibration was applied, as shown in FIG. 5.

[0091] In particular, the predicted offsets were rescaled between the reconstructed image x 308 and the learned quantiles 316, 318 using a data-driven correction factor A. In conventional conformal QR, the quantiles themselves are rescaled. However, since the uncertainty generator 114 scales the convolutional neural network 312 outputs with the images, the quantities to calibrate may be the offsets (x-Z) and (u-x). The pixelwise uncertainty interval for pixel (m, n) of x can be:264932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0092] TA(x(m.n)) = [lb,ub],

[0095] In some implementations, to obtain statistically valid uncertainty intervals, the predicted bounds can be scaled by a calibration factor A. computed on 50% of a validation set. The calibrated pixel-wise uncertainty interval can be:

[0097] The smallest X may be desirable such that T,_ achieves desired coverage. The smallest X can be computed using a held-out calibration dataset of I image pairs {(x\ y1)}where xlare reconstructions from the model 302 and ylare the corresponding ground-truth images. The value of A. may be chosen such that the average fraction of all pixels (m, n) across the calibration set whose ground-truth intensities fall outside respective corresponding intervalsdoes not exceed a. Since this fraction is estimated from a finite calibration set, a conservative upper bound R++ (A) is used following Hoeffding’s inequality. The final calibrated scaling factor is therefore given by:<

[0099] A calibration process 506 may use approximately 6 and 3 hours for the brain subsets and knee subsets, respectively. The calibrated scaling factors for the brain and QR-based uncertainty can be 1.31, 1.54, 1.59, 1.74, and 1.87 for 2*, 4*, 6*, 8*, and 10x acceleration, respectively. For the ResM-based uncertainty, the value at 4 acceleration was 1.02. For the knee subsets, the values for the QR-based uncertainty (e.g., generated by the uncertainty generator 114) were 1.42 and 1.68 for 4 and 6 acceleration, respectively. For the ResM-based uncertainty the value at 4 acceleration was 0.82.274932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0100] To assess the correspondence between the predicted uncertainty 320 by the uncertainty generator 114 and the true reconstruction error, the uncertainty map q 320 and the absolute error map e can be defined as:

[0101] q = ud(x) - Zd(x), e = |x - y|

[0102] The map q 320 quantifies the pixelwise uncertainty width, while e represents the magnitude of the absolute (e.g., true) reconstruction error, q does not have any knowledge of the ground-truth y.

[0103] The correspondence between the predicted uncertainty q 320 and the true error e is quantified using the Pearson and Spearman correlation coefficients. The Pearson correlation coefficient measures the linear association between q and e and is defined as:

[0104]

[0105] where cov denotes the covariance.

[0106] The Spearman rank correlation coefficient quantifies whether pixels with higher predicted uncertainty also tend to have higher reconstruction errors, independent of respective exact numerical scaling. The spearman correlation coefficient is computed by applying the Pearson correlation to the rank-ordered variables, where each pixel value in q and e is replaced by a respective position in the sorted list of all pixel values within the image, yielding Rq and R§, respectively. The coefficient is then computed as:

[0107]

[0108] Here, Rq and R§ denote the rank-ordered maps, in which lower pixel values receive smaller ranks and higher pixel values receive larger ranks.

[0109] In addition to global correlations, region-based Pearson and Spearman correlations were computed to assess whether uncertainty reflects local reconstruction errors. Each image was284932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)divided into 100 non-overlapping patches of approximately equal size, and the correlation metrics were computed within each patch and averaged across all regions.

[0110] Before computing correlation metrics, both q and e were blurred using a Gaussian filter with a standard deviation of c = 2 to suppress pixel-level noise and emphasize spatially coherent patterns.

[0111] FIG. 6 shows quantitative results of the reconstruction output by the uncertainty generator 114 and corresponding uncertainty estimates output by the uncertainty generator 114 across multiple acceleration factors and contrasts. FIG. 6 depicts mean and standard deviation of QR-uncertainty (e.g., output by the uncertainty generator 114) across multiple contrasts and accelerations. The magnitude of the residual (ResM) method is shown at 4x acceleration for comparison. Specifically, FIG. 6 reports the mean and standard deviation of the SSIM between the reconstruction and the ground-truth and the uncertainty maps estimated by the uncertainty generator 114 for acceleration factors of 2, 4, 6, 8, and 10 for each contrast in the brain subset and for acceleration factors of 4 and 6 for the knee subset. For the 4 fold accelerated reconstructions, QR-based and ResM-based uncertainties were compared. Both methods yielded comparable ranges for the uncertainty estimates. All uncertainty maps were normalized by the maximum reconstruction magnitude to express values as percentages. All values were computed per volumetric case.

[0112] FIG. 7 reports Pearson and Spearman correlation coefficients (including region-based) between the predicted uncertainty and the true reconstruction error for the brain subset for 2, 4, 6, 8, and 10 acceleration factors, and the knee subset for 4 and 6 acceleration factors. For the 4-fold accelerated reconstructions, the correlation coefficients were computed between the ResM-based uncertainty and the true reconstruction error. The Pearson and Spearman correlation coefficients are computed after Gaussian blurring (cr = 2) to suppress pixel-level noise. FIG. 7 depicts a correlation between uncertainty and reconstruction error using Pearson and Spearman correlations and the corresponding region-based correlations (e.g., 100 regions). QR shows strong consistency with the error (e.g., Pearson correlation coefficients at 4x is >0.9), while the ResM correlates poorly.294932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0113] FIG. 8 shows the distribution of the per-case Pearson correlations between the predicted uncertainty and the true reconstruction error for the 4-fold acceleration factor undersampled brain (top) and knee (bottom) subsets. Each bar represents the number of cases achieving a given correlation level. The QR-based estimates are consistently higher than the ResM-based ones for both anatomies, indicating improved alignment with he true reconstruction error.

[0114] FIG. 9 illustrates a healthy and three abnormal brains, and compares QR-based (e.g., uncertainty generator 114) and ResM-based uncertainty maps for 4-fold undersampling. Specifically, FIG. 9 depicts comparison between the absolute error (x50), QR-, and ResM-based uncertainty, both normalized with the maximum reconstruction value, for one healthy and three abnormal brains including pathologies. All reconstructions were performed with four-fold undersampling. Unlike the ResM method, QR closely matches the absolute error. In the abnormal cases, QR delineates the lesions, demonstrating superior localization of uncertainty compared with the magnitude of the ResM method. FIG. 9 also depicts blurred versions of the error and the uncertainties. The QR-based uncertainty closely matches the absolute error distribution and in the abnormal cases, the QR-based uncertainty delineates the lesions, demonstrating superior localization of uncertainty compared to ResM.

[0115] FIG. 10 overlays the QR-based uncertainty map on the reconstruction of an abnormal brain, highlighting correspondence with pathological regions. Specifically, FIG. 10 depicts a comparison between reconstructions and QR-based uncertainty maps for 2-, 4-, 6-, 8-, and 10-fold acceleration factors in an abnormal brain from the testing dataset. The uncertainty map highlights a one-to-one correspondence between uncertainty and pathological regions in the reconstruction. For example, the uncertainty highlights the lesion and the posterior susceptibility artifact from a prior craniotomy starting at 4-fold acceleration. The bottom row shows the reconstruction overlaid with the uncertainty map, which can be an uncertainty heatmap, generated by the uncertainty generator 114, illustrating that regions of high uncertainty spatially coincide with the areas of the pathology and artifacts.

[0116] FIG. 11 presents thresholded uncertainty maps comparing a healthy and three abnormal cases with pathologies. The top panels are not thresholded, the middle and the bottom panels depict the uncertainty thresholded using the maximum value from 2 and 4 accelerated cases, respectively.304932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)Specifically, the lower bound is set to 0 for the top panel, and the lower bound is set to the maximum uncertainty for the 2 fold and 4 fold accelerated cases on the middle and bottom panels, respectively. QR-based uncertainty maps for 2, 4, 6, 8, and 10 fold acceleration are shown after thresholding, with the lower bound set to the maximum uncertainty of the 4-fold cases, since the 4-fold cases are considered clinically valid (e.g., true and accurate, etc.). Thresholding highlighted regions of elevated uncertainty at higher accelerations, corresponding to lesions and post-treatment effects. In some implementations, uncertainty maps of abnormal cases may appear brighter and noisier compared to other cases due to a higher level of background noise corrupting the groundtruth, thereby causing the uncertainty generator 114 to increase the uncertainty of the distribution.

[0117] The uncertainty generator 114 can use sigmoid activations to scale convolutional neural network outputs relative to the reconstruction of the model 302, ensuring consistent dynamic range and stable uncertainty estimates across all acceleration factors. FIG. 7 shows that QR achieves high Pearson and Spearman correlations with the true reconstruction error for 4-, 6-, 8-, and 10-fold acceleration, confirming that the uncertainty maps generated by the uncertainty generator 114 reliably reflect the magnitude and spatial distribution of the reconstruction error. This trend is further supported qualitatively in FIG. 9, where QR uncertainty visually tracks the absolute error maps, whereas ResM fails to consistently capture the patterns. The strong agreement between QR uncertainty and reconstruction error suggests that the uncertainty maps generated by the uncertainty generator 114 can assess image reliability for accelerated acquisitions when fully sampled ground-truth reconstructions are unavailable. In addition, regions with a high degree of uncertainty for high acceleration factors can facilitate detection anatomical abnormalities or pathologies, as shown in at least FIG. 11. The uncertainty generator 114 can quantify a reliability of accelerated MRI reconstructions without a ground-truth reference.

[0118] FIGS. 9-10 demonstrate that the uncertainty generator 114 provides more accurate and interpretable estimates that methods using the magnitude of the residual. In particular, QR produces uncertainty maps that more closely resemble an absolute error distribution, indicating that the uncertainty generator 114 reliably captures regions where the reconstruction deviates from the ground-truth, which is also seen in the overlaid images with uncertainty maps of FIG. 11. Such alignment suggests that the uncertainty generator 114 can be used instead of reconstruction error during scan time in situations where fully sampled data is unavailable, such as in cases of reduced314932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)scan time. Furthermore, thresholding the uncertainty maps using ranges from well -validated clinical reconstructions (e.g., 2-fold, 4-fold) enables identification of regions with elevated uncertainty at higher accelerations as shown in FIG. 11, which can facilitate detection of anatomical abnormalities or pathologies.

[0119] FIG. 6 shows that uncertainty values generated by the uncertainty generator 114 spans a similar numerical range to ResM. However, the qualitative comparison shown in at least FIG. 9 demonstrates that QR provides uncertainty maps that better mirror the absolute error distribution. For the healthy brain in FIG. 9, QR highlights a noisy posterior region where the reconstruction error is elevated, whereas ResM produces uniformly high uncertainty around the entire brain, inconsistent with the spatial distribution of error. In the abnormal cases including pathology, the uncertainty generator 114 captures global noise patterns and consistently identifies the lesion location and structure more accurately. In the first abnormal case (Abnormal #1) the ResM-based uncertainty in the lesion is low, where the error is large and vice-versa. The uncertainty map output by the uncertainty generator 114, however, clearly delineates the lesion and follows the error pattern. In the second abnormal case (Abnormal #2), the morpohology of the lesion is distorted by ResM. In the third abnormal case (Abnormal #3), the uncertainty generator 114 captures the boundary of the treated region, which is lost in the ResM estimate.

[0120] FIG. 7 further supports these observations by quantifying the relationship between uncertainty and true reconstruction error. After conformal calibration, the uncertainty maps generated by the uncertainty generator 114 show high Pearson correlations (e.g., roughly 90%) with the true error for 4-fold accelerated brain reconstruction and above, indicating strong agreement between uncertainty and error magnitude. These high correlations show that the calibrated uncertainty values generated by the uncertainty generator 114 scale proportionally with the underlying error across the image. For the ResM-based uncertainty the correlation drops to 69%. In particular, as shown in FIG. 8, the Pearson correlation is higher when the uncertainty generator 114 is used over ResM. For the reconstructed knee images, for example, the ResM-based uncertainty correlates poorly with the true reconstruction error (e.g., Pearson correlation < 25%).

[0121] Pearson correlations are consistently higher than Spearman due to the magnitude alignment being stronger than strict pixelwise rank ordering, which is more sensitive to local324932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)fluctuations in low-error regions. The lower correlations observed at 2-fold acceleration are expected, as the reconstruction error for 2-fold acceleration is small and spatially homogeneous compared to higher acceleration factors, yielding insufficient dynamic range for a meaningful correlation analysis.

[0122] FIG. 10 shows the QR-based uncertainty heatmap overlaid on the reconstruction, revealing clear spatial correspondence between the lesions and regions of high uncertainty. At 2-fold acceleration, the uncertainty is near zero and therefore uninformative, which is expected given the high reconstruction accuracy of the model 302 at low acceleration factors. At higher accelerations, however, the uncertainty increases, especially in the area of the pathology and in the area affected by the posterior susceptibility artifact from the prior craniotomy. Notably, the regions of elevated uncertainty precisely overlap with the true lesion locations, indicating that the uncertainty generator 114 is a trustworthy indicator of reconstruction reliability. Furthermore, the alignment suggests that the uncertainty generator 114 can be used instead of reconstruction error during scan time, when fully sampled data are unavailable.

[0123] Conventional reconstruction methods yield clinically reliable images up to 2-fold acceleration, while unrolled network architectures, such as the end-to-end variational network, maintain clinical reliability for up to 4-fold acceleration. For this reason, as shown in FIG. 11, the maximum uncertainty value from the 4-fold accelerated reconstructions were used as a threshold to evaluate higher-acceleration factor reconstructions. In FIG. 11 , the healthy brain example shows uncertainty below the threshold at 10-fold acceleration, suggesting that certain slices from healthy studies can be reconstructed at higher acceleration factors. For example, in the first abnormal case (Abnormal #1), uncertainty at 6-fold undersampling remains below the threshold as well, indicating that some lesions may still be reconstructed accurately with fewer k-space samples than the conventional 4-fold limit, such as 6-fold or more. Overall, the results of FIG. 11 suggest that a highly accelerated scan-such as a 10-fold accelerated acquisition-can be used as an anomaly detector or pre-scan to determine whether a longer, diagnostic-quality scan should be performed.

[0124] In some implementations, the uncertainty generator 114 integrates the uncertainty with decision rules for adaptive sampling. In some implementations, the predicted uncertainty maps by334932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)the uncertainty generator 114 can be incorporated into a reconstruction pipeline to selectively refine high-uncertainty areas and potentially improve SSIM performance at higher accelerations.

[0125] As shown, the uncertainty generator 114 can quantify the uncertainty of accelerated MRI reconstructions in the absence of ground-truth references. The uncertainty generator 114 demonstrates strong agreement in both magnitude and spatial distribution across contrasts and acceleration factors. The uncertainty generator 114 consistently highlights pathological regions and reconstruction failures that are missed by simpler heuristic notions of uncertainty, such as ResM. Since uncertainty remains low for clinically validated accelerations and increases at higher accelerations, the uncertainty generator 114 can be used to assess reconstruction reliability in real time, thus allowing for highly accelerated scans to be used as rapid anomaly detectors. Real-time access to uncertainty information can enable adaptive sampling in which the MRI scanner acquires additional k-space lines only when the uncertainty exceeds a threshold, thereby optimizing scan efficiency while preserving diagnostic quality.

[0126] FIG. 12 is a flow diagram of an example method 1200 for acquiring uncertainty maps, according to some implementations of the present disclosure. The method 1200 can be performed using various systems described herein. Various steps in the method 1200 may be repeated, omitted, performed in various orders, or otherwise modified. Various steps in the method 1200 may be run concurrently, in parallel, or individually. The method 1200 may be implemented by at least one or more processors.

[0127] The method 1200 includes, at block 1202, receiving a signal associated with a first value. The signal may include a first portion of MRI data and may be received during an MRI scan. In some implementations, the first value is a first acceleration factor. In some implementations, prior to block 1202, the method 1200 includes performing a scan using an MRI machine at a first acceleration factor. The method 1200 includes, at block 1204, generating an image. The image is an MR image and is generated using at least one model and the signal. The at least one model may include a reconstruction model, and the image includes a second portion of the MRI data. The second portion of the MRI data can include the first portion. For example, the first portion of the MRI data may be an under sampled dataset while the second portion is generated by the at least one model and represents the fully sampled dataset.344932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)

[0128] The method 1200 includes, at block 1206, generating a map corresponding to the image. The map is generated using the at least one model which may include at least one neural network. The at least one model can include a reconstruction model to generate the image and at least one neural network to generate the map. The map may be an uncertainty map.

[0129] The method 1200, at block 1208, includes determining, based on the uncertainty map, at least one characteristic of the uncertainty map. The at least one characteristic can include at least one of the uncertainty map being at or below, or above a threshold, a detected anomaly, or a severity of a condition of an individual corresponding to the uncertainty map. The method 1200 can further include the one or more processors being configured to compare the uncertainty map to a threshold, adjust the first parameter to a second parameter responsive to the uncertainty map being above the threshold, operate the medical imaging system to perform the imaging scan at the second parameter responsive to adjusting the first parameter to the second parameter, and operate the medical imaging system to perform the imaging scan at the first parameter responsive to the uncertainty map being at or below the threshold. The first parameter and the second parameter can be acceleration factors.

[0130] In some implementations, the method 1200 further includes one or more processors being configured to compare the uncertainty map to a threshold and detect at least one anomaly responsive to at least a portion of the uncertainty map being above the threshold. In some implementations, to determine the at least one characteristic including the severity of the condition of the individual, the method 1200 can include comparing the uncertainty map to generated uncertainty maps of other individuals. In some implementations, the method 1200 can include one or more processors determining an order of triage based on an individual having a highest uncertainty value of the uncertainty map compared to other individuals. The highest uncertainty value can be at least one of an average, maximum, minimum, or any other value of the uncertainty map. In some implementations, the uncertainty map can be provided to a model to update or otherwise train the model. In such implementations, the method 1200 may or may not include determining the at least one characteristic.

[0131] Moreover, for high acceleration factors, regions of the uncertainty map with high uncertainty values may correspond to pathology. Thus, the uncertainty values can be used for354932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)anomaly detection which varies from typical anomaly detection which relies on high quality reconstruction. The anomaly detection can include detecting pathology. In addition to anomaly detection, the uncertainty maps acquired with rapid under sampled acquisitions can be used for triage. For example, the uncertainty maps can be used to determine a severity of a condition of an individual corresponding to the uncertainty maps, such as presence of pathology or a size of the pathology. The uncertainty maps can thus be used to prioritize individuals with more severe conditions.Definitions.

[0132] As used herein, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, the term “a member” is intended to mean a single member or a combination of members, “a material” is intended to mean one or more materials, or a combination thereof.

[0133] As used herein, the terms “about” and “approximately” generally mean plus or minus 10% of the stated value. For example, about 0.5 would include 0.45 and 0.55, about 10 would include 9 to 11, about 1000 would include 900 to 1100.

[0134] It should be noted that the term “exemplary” as used herein to describe various embodiments is intended to indicate that such embodiments are possible examples, representations, and / or illustrations of possible embodiments (and such term is not intended to connote that such embodiments are necessarily extraordinary or superlative examples).

[0135] As used herein, the terms “coupled,” “connected,” and the like mean the joining of two additional intermediate members being integrally formed as a single unitary body with one another or with the two members or the two members and any additional intermediate members being attached to one another.

[0136] As shown in FIG. 13, a system 1300 includes a computer-accessible medium 1320, a processing arrangement 1310, and a storage arrangement 1340 joined directly or indirectly to one another. Such joining may be stationary (e.g., permanent) or moveable (e.g., removable or releasable). Such joining may be achieved with the two members or the two members and any device such as a hard disk, floppy disk, memory stick, CD-ROM, RAM, ROM, etc., or a collection364932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)thereof) can be provided (e.g., in communication with the processing arrangement 1310). The computer-accessible medium 1320 may be a non-transitory computer-accessible medium. The computer-accessible medium 1320 can contain executable instructions 1330 thereon. In addition, or alternatively, a storage arrangement 1340 can be provided separately from the computer-accessible medium 1320, which can provide the instructions to the processing arrangement 1310 so as to configure the processing arrangement to execute certain exemplary procedures, processes and methods, as described herein, for example. The instructions may include a plurality of sets of instructions.

[0137] The system 1300 may also include a display or output device, an input device such as a keyboard, mouse, touch screen or other input device, and may be connected to additional systems via a logical network. Many of the embodiments described herein may be practiced in a networked environment using logical connections to one or more remote computers having processors. Logical connections may include a local area network (“LAN”) and a wide area network (“WAN”) that are presented here by way of example and not limitation. Such networking environments are commonplace in office-wide or enterprise-wide computer networks, intranets and the Internet and may use a wide variety of different communication protocols. Those skilled in the art can appreciate that such network computing environments can typically encompass many types of computer system configurations, including personal computers, hand-held devices, multiprocessor systems, microprocessor -based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. Embodiments of the invention may also be practiced in distributed computing environments where tasks are performed by local and remote processing devices that are linked (either by hardwired links, wireless links, or by a combination of hardwired or wireless links) through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0138] Various embodiments are described in the general context of method steps, which may be implemented in one embodiment by a program product including computer-executable instructions, such as program code, executed by computers in networked environments. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of program code374932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps.

[0139] Software and web implementations of the present invention could be accomplished with standard programming techniques with rule-based logic and other logic to accomplish the various database searching steps, correlation steps, comparison steps and decision steps. It should also be noted that the words “component” and “module,” as used herein and in the claims, are intended to encompass implementations using one or more lines of software code, and / or hardware implementations, and / or equipment for receiving manual inputs.

[0140] It is important to note that the construction and arrangement of the various exemplary embodiments are illustrative only. Although only a few embodiments have been described in detail in this disclosure, those skilled in the art who review this disclosure will readily appreciate that many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.) without materially departing from the novel teachings and advantages of the subject matter described herein. Other substitutions, modifications, changes and omissions may also be made in the design, operating conditions and arrangement of the various exemplary embodiments without departing from the scope of the present invention.

[0141] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any inventions or of what may be claimed, but rather as descriptions of features specific to particular implementations of particular inventions. Certain features described in this specification in the context of separate implementations can also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation can also be implemented in multiple implementations separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.384932-2355-4177.2

Claims

Ref. No.: LAT01-07PCT (046434-0973)WHAT IS CLAIMED IS:

1. A method, comprising:receiving, by one or more processors, a signal associated with a first value comprising a first portion of magnetic resonance imaging (MRI) data during an MRI scan;generating, by the one or more processors using at least one model and the signal, an image comprising a second portion of the MRI data;generating, by the one or more processors using the at least one model, a map corresponding to the image;determining, by the one or more processors using the at least one model, at least one uncertainty value based at least partially on the map;comparing, by the one or more processors, the at least one uncertainty value to a threshold;adjusting, by the one or more processors, the first value to a second value responsive to the at least one uncertainty value being above the threshold;resuming, by the one or more processors, the MRI scan at the second value responsive to adjusting the first value to the second value; andresuming, by the one or more processors, the MRI scan at the first value responsive to the at least one uncertainty value being at or below the threshold.

2. The method of claim 1 , wherein the map comprises a plurality of regions corresponding to regions of the image, each of the plurality of regions associated with a region uncertainty value.

3. The method of claim 2, wherein the region uncertainty value is determined by at least one of a magnitude of a residual or a pixelwise quantile regression.

4. The method of claim 2, wherein the at least one uncertainty value is an average uncertainty value and to determine the at least one uncertainty value, the method further comprises:segmenting, by the one or more processors, the plurality of regions into a plurality of areas based on respective region uncertainty values;394932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)comparing, by the one or more processors, the plurality of regions to determine at least one region with higher region uncertainty values compared to other regions in the plurality of regions; anddetermining, by the one or more processors, the average uncertainty value for the at least one region with higher region uncertainty values.

5. The method of claim 1 , wherein the first value and the second value are acceleration factors.

6. The method of claim 1, wherein the signal is a first signal, the image is a first image and the at least one uncertainty value is a first uncertainty value, the method further comprising: receiving, by the one or more processors, a second signal associated with the second value comprising the first portion and a third portion of the MRI data;combining, by the one or more processors, the second signal and the first signal into a third signal; andgenerating, by the one or more processors using the at least one model and the third signal, a second image comprising a fourth portion of the MRI data.

7. The method of claim 6, wherein the map is a first map and the threshold is a first threshold, the method further comprising:comparing, by the one or more processors, the second value to a second threshold corresponding to a third value, the third value less than the first value;resuming, by the one or more processors, the MRI scan at the second value responsive to the second value being equal to the second threshold; andgenerating, by the one or more processors using the at least one model, a second map corresponding to the second image responsive to the second value being greater than the second threshold.

8. The method of claim 1, wherein the at least one model comprises a reconstruction model to generate the image and at least one neural network to generate the map.404932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)9. The method of claim 1, further comprising:determining, by the one or more processors, a distance value based on the image and a training image corresponding to the image, the training image comprised in a training dataset; determining, by the one or more processors, a loss value as a function of the map and the distance value; andupdating, by the one or more processors, the at least one model based on the loss value.

10. The method of claim 1, wherein the at least one model comprises a plurality of neural networks, each neural network corresponding to at least one of a type of MRI scan or an acceleration factor.

11. A method, comprising:performing, by one or more processors, a scan using a magnetic resonance imaging (MRI) system at a first acceleration factor;receiving, by the one or more processors, a first signal associated with the first acceleration factor;generating, by the one or more processors using a reconstruction model, an image based on the first signal;generating, by the one or more processors using at least one model, an uncertainty map corresponding to the image;determining, by the one or more processors, at least one uncertainty value based on the uncertainty map;comparing, by the one or more processors, the at least one uncertainty value to a threshold;adjusting, by the one or more processors, the first acceleration factor to a second acceleration factor responsive to the at least one uncertainty value being above the threshold; performing, by the one or more processors, the scan at the second acceleration factor responsive to adjusting the first acceleration factor to the second acceleration factor; and performing, by the one or more processors, the scan at the first acceleration factor responsive to the at least one uncertainty value being at or below the threshold.414932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)12. The method of claim 11, wherein the image is a first image, the method further comprising:receiving, by the one or more processors, a second signal associated with the second acceleration factor;combining, by the one or more processors, the second signal and the first signal into a third signal; andgenerating, by the one or more processors using the reconstruction model, a second image based on the third signal.

13. The method of claim 12, wherein the uncertainty map is a first uncertainty map and the threshold is a first threshold, the method further comprising:comparing, by the one or more processors, the second acceleration factor to a second threshold corresponding to a third acceleration factor, the third acceleration factor less than the first acceleration factor;performing, by the one or more processors, the scan at the second acceleration factor responsive to the second acceleration factor being equal to the second threshold; and generating, by the one or more processors using the at least one model, a second uncertainty map corresponding to the second image responsive to the second acceleration factor being greater than the second threshold.

14. The method of claim 11, further comprising:determining, by the one or more processors, a distance value based on the image and a training image corresponding to the image, the training image comprised in a training dataset; determining, by the one or more processors, a loss value as a function of the uncertainty map and the distance value; andupdating, by the one or more processors, the at least one model based on the loss value.

15. The method of claim 11, wherein the at least one model comprises a plurality of neural networks, each neural network corresponding to at least one of a type of MRI scan or an acceleration factor.424932-2355-4177.2Ref. No.: LAT01-07PCT (046434-0973)16. A system, comprising one or more processors and at least one memory, the one or more processors configured to:receive a first signal associated with a first parameter, the first signal received from a medical imaging system during performance of an imaging scan;generate, using a first model, an image based on the first signal;generate, using a second model, an uncertainty map corresponding to the image; and determine, based on the uncertainty map, at least one characteristic of the uncertainty map.

17. The system of claim 16, wherein the at least one characteristic comprises the uncertainty map being at or below, or above a threshold, the one or more processors further configured to: compare the uncertainty map to the threshold;adjust the first parameter to a second parameter responsive to the uncertainty map being above the threshold;operate the medical imaging system to perform the imaging scan at the second parameter responsive to adjusting the first parameter to the second parameter; andoperate the medical imaging system to perform the imaging scan at the first parameter responsive to the uncertainty map being at or below the threshold.

18. The system of claim 17, wherein the first parameter and the second parameter are acceleration factors.

19. The system of claim 16, wherein the at least one characteristic comprises a detected anomaly, the one or more processors further configured to:compare the uncertainty map to a threshold; anddetect at least one anomaly responsive to at least a portion of the uncertainty map being above the threshold.

20. The system of claim 16, wherein the medical imaging system is at least one of a magnetic resonance imaging (MRI) or computed tomography (CT).434932-2355-4177.2

Images