Improving the quality of magnetic resonance scans affective by under-sampling for multiple MRI contrasts

Abstract

An AI-based MRI system (100) reconstructs images of different types of contrasts (like T1, T2 MRI) using several components. The MRI scanner (102) generates scanner measurements (102a), acquires under-sampled data (104a), and provides MRI contrast information (104b). A training data preparation module (106) prepares training data consisting of under-sampled and fully sampled data (104a). A partitioning module (108) partitions this data into support and query datasets for each contrast. The kernel modulation weight prediction network (110), controlled by a processor (112), uses the MRI contrast information (104b) to predict kernel modulation weights. The two-level optimization module (118) iteratively refines the weights of the kernel modulation weight prediction network through inner and outer level corrections using the support and query training data. The base reconstruction CNN module (114), also controlled by the processor (112), applies CNN kernel modulation based on the predicted weights to generate a new contrast-specific K-space (116). Finally, the processor (112) combines the scanner measurements (102a) and the new contrast-specific K-space (116) to reconstruct the final image (120).

Description

[0001] IMPROVING THE QUALITY OF MAGNETIC RESONANCE SCANS AFFECTIVE BY UNDER- SAMPLING FOR MULTIPLE MRI CONTRASTS

[0002] TECHNICAL FIELD

[0003] The present invention relates to a multi-contrast reconstruction method that can be used for reconstructing multiple different MR images of various contrasts, one at a time, leading to a system that scales to different contrasts as well as improves the quality of the degraded images for radiological evaluation.

[0004] BACKGROUND

[0005] Background description includes information that may be useful in understanding the present invention. It is not an admission that any of the information provided herein is prior art or relevant to the presently disclosed invention, or that any publication specifically or implicitly referenced is prior art.

[0006] Multi-contrast MRI (MC-MRI) involves the acquisition of multiple images of different contrast and intensity levels and helps to visualize the different anatomical structures and the functional processes for a given patient’s anatomy. Obtaining multi -contrast magnetic resonance images (MC-MRI) from a single subject is valuable for enhancing the accuracy and thoroughness of radiological assessments. Nevertheless, traditional MRI techniques inherently demand extended acquisition times, causing patient discomfort and introducing motion artifacts and system imperfections that obscure important anatomical details relevant to medical analysis. Consequently, there has been significant interest in developing accelerated methods capable of reconstructing high- quality, artifact-free images using fewer frequency-domain (k-space) MRI measurements, thereby reducing scan time.

[0007] Referring to Figure 1, recently deep learning (DL) methods have shown promising results in recovering images of various contrast. Existing deep-learning techniques address the reconstruction of only a single type of contrast. However, as MRI is diversified by various complementary protocols of the subject, training a model separately for each MRI contrast is practically infeasible for deployment because of large computational requirements and cannot be scaled to multiple acquisition settings. Methods like test-time training can be used to train the DL model at test time; however, this requires large computational resources and machine learning expertise for clinicians to train deep learning models.

[0008] Combining all contrast data and training a single DL model is not efficient as it does not consider the contrast and intensity levels specific to each protocol and learns common knowledge that can be shared across contrasts. So, the current methods are limited because of the lack of two significant attributes: 1) contrast invariant information that is characterized by the common knowledge shared across images of different MRI contrasts and 2) additional information describing the type of contrast of the MRI images and use this information to learn common knowledge across contrasts.

[0009] Therefore, there is a need for a device or a method that imbibes the above attributes to realize an adjustable model capable of reconstructing multiple MRI contrasts, including the contrasts that are deviated in acquisition settings and are not seen during training.

[0010] SUMMARY OF THE INVENTION

[0011] The following presents a simplified summary of the subject matter in order to provide a basic understanding of some aspects of subject matter embodiments. This summary is not an extensive overview of the subject matter. It is not intended to identify key / critical elements of the embodiments or to delineate the scope of the subject matter. Its sole purpose is to present some concepts of the subject matter in a simplified form as a prelude to the more detailed description that is presented later.

[0012] The existing system for solving the problem of multi-contrast restoration consists of training a model for every degradation. Moreover, such solutions consist of only a correction mechanism without an evaluation mechanism. Therefore, an Al-based MRI system for reconstructing images disclosed here addresses the need for a device and associated method that imbibes the above attributes to realize an adjustable model capable of reconstructing multiple MRI contrasts, including the contrasts that are deviated in acquisition settings and are not seen during training. The Al-based MRI system involves a method for reconstructing images, where the method comprising acquiring scanner measurements or the under- sampled data and the MRI contrast information using an MRI scanner.

[0013] The method involves preparing training data from the acquired under-sampled and the corresponding fully-sampled data using a training data preparation module, followed by partitioning the prepared training data into support and query datasets for each MRI contrast using a partitioning module. Kernel modulation weights are then predicted based on the MRI contrast information using a kernel modulation weight prediction network controlled by at least one processor. These predicted weights are used to apply convolutional neural network (CNN) kernel modulation to generate a new MRI contrast- specific K-space using a base reconstruction CNN module controlled by the processor. Finally, the scanner measurements and the new contrast-specific K-space are combined to reconstruct the final image using the processor. The method comprises performing correction and evaluation iteratively to refine the kernel modulation weights using a two-level optimization module controlled by the processor. The two-level optimization module performs an inner level correction of the weights of the kernel modulation network using the support training data, an evaluation using query training data, and an outer level correction of these weights using the query training data. This iterative process ensures the optimization of the kernel modulation network’s weights for accurate image reconstruction to adapt to multiple contrasts in MRI.

[0014] In an embodiment, predicting kernel modulation weights enables an Al system to function similarly to attentional modulation in the human brain. The predicted modulation weights are used to modulate the parameters of a main neural network, enhancing the system's ability to focus on relevant MRI contrasts and improve the accuracy of image reconstruction.

[0015] In an embodiment, the kernel modulation weights enable the base reconstruction convolutional neural network (CNN) module to adapt to varying MRI contrasts. These contrasts include, but are not limited to, Tl, FLAIR, PD, and T2 contrast data. This adaptability ensures that the CNN module can effectively handle different types of MRI data for accurate image reconstruction.

[0016] In an embodiment, the inner level correction step involves faster adaptation to each MRI contrast data using the support data partition. This step allows the kernel modulation weight prediction network to quickly adjust its parameters for each specific contrast, ensuring efficient and accurate image reconstruction during the training phase. In an embodiment, the outer level correction step comprises updating the kernel modulation weight prediction network based on the average error computed from the query data across all MRI contrasts. This update process ensures that the network continuously improves its predictions, leading to more accurate and reliable kernel modulation weights for image reconstruction.

[0017] In an embodiment, a final trained base reconstruction convolutional neural network (CNN) module utilizes the optimized kernel modulation weights for reconstructing MRI images as the final image. The optimized weights ensure that the final CNN module can accurately reconstruct high-quality images from the under-sampled data and MRI contrast information, providing reliable images for clinical use.

[0018] In short, the invention that uses a single -headed multi-task learning (each contrast posed as a task) network for multi-contrast MRI is different in two aspects: 1) A single model is designed to effectively improve the quality of scans of multiple contrasts and 2) Evaluation mechanism is included that boosts the performance of the overall process even in the scenario of unseen deviated contrasts. The invention efficiently utilizes the amounts of under- sampling to provide additional information for improving the scan quality. The following are the features of the embodiment that are different from existing solutions:

[0019] 1. A process of analyzing multiple MRI contrasts by using a multi-task learning method that provides a single model for improving the quality of MRI scans affected by various amounts of under- sampling degradations.

[0020] 2. In an embodiment of the process in point 1, the process outputs a clean counterpart of the under- sampling artifact affected MRI scan. 3. In an embodiment of the process in point 1, the process incorporates the nuances of contrast information into the training method in determining the artifact-free image for a given poor-quality MRI scan.

[0021] 4. In an embodiment of the process in point 1, where multiple contrast data is used for intelligently updating the model during the training to enhance the overall accuracy for multi-contrast reconstruction.

[0022] The prior art involves multi-contrast image reconstruction techniques tailored for specific image intensity and contrast ranges. The new invention enhances multi-contrast MRI by accommodating intensity levels that deviate from the training data.

[0023] Previously, two training methods were used: training the entire network using stochastic gradient descent or meta-learning. However, meta-learning for the entire network is memory intensive. In the Al-based MRI system disclosed here, only the lightweight kernel modulation neural network weights are trained using meta-learning, while the rest are trained using stochastic gradient descent. The kernel modulation network predicts modulation coefficients, which are then multiplied with the weights of each convolution layer of the main convolutional neural network which performs image reconstruction . This approach has two advantages: 1. The main neural network quickly adapts to different unseen MRI contrasts without requiring retraining. 2. The method is both computation- and storage-efficient.

[0024] The Al-based MRI system disclosed here offers significant business value, particularly for the healthcare industry, including scanner manufacturers like GE, Philips, and Siemens, in radiology applications. The method can scan a wide range of anatomies and diverse protocols. This results in increased throughput and adoption of MRI through faster Al-driven workflows, significantly reducing the time required for radiologists and clinicians in the imaging process. While Al-driven workflows typically need retraining for different contrast ranges, this method using Al-based MRI system can reduce this retraining effort dramatically — from months to just one or two days.

[0025] BRIEF DESCRIPTION OF FIGURES The foregoing and further objects, features and advantages of the present subject matter will become apparent from the following description of exemplary embodiments with reference to the accompanying drawings, wherein like numerals are used to represent like elements.

[0026] It is to be noted, however, that the appended drawings along with the reference numerals illustrate only typical embodiments of the present subject matter, and are therefore, not to be considered for limiting its scope, for the subject matter may admit to other equally effective embodiments.

[0027] Figure 1 exemplarily illustrates an MRI contrast data source and a learning pipeline, in the background context of the present invention.

[0028] Figure 2 exemplarily illustrates a schematic drawing that shows regarding the manner in which MRI imaging and deep learning models can be minimized along with increase in adoption.

[0029] Figure 3A exemplarily illustrates a schematic system drawing that shows an Al-based MRI system for reconstructing images, as an example embodiment of the present disclosure.

[0030] Figure 3B exemplarily illustrates a Flowchart of the learning process, as an example embodiment of the present invention, as an example embodiment of the present disclosure.

[0031] Figure 4A exemplarily illustrates a pictorial representation of the task-specific inner loop adaptation that becomes faster in few gradient steps, as an example embodiment of the present disclosure. This process is used to obtain contrast- specific weights of the kernel modulation network shown as <b_Tl for Tl- contrast specific weights, T2 for T2 contrast- specific weights, similarly for <I» FLAIR and PD. Figure 4B exemplarily illustrates a pictorial representation showing meta weight update and outer loop optimization of the kernel modulation network wieghts providing metainitialization of the kernel modulation network, as an example embodiment of the present disclosure.

[0032] DETAILED DESCRIPTION

[0033] Illustrative examples of the subject matter will now be disclosed. In the interest of clarity, not all features of an actual implementation are described in this specification. It will be appreciated that in the development of any such actual implementation, numerous implementation-specific decisions may be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which will vary from one implementation to another. Moreover, it will be appreciated that such a development effort, even if complex and time-consuming, would be a routine undertaking for those of ordinary skill in the art having the benefit of this disclosure.

[0034] Exemplary embodiments now will be described with reference to the accompanying drawings. The disclosure may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey its scope to those skilled in the art. The terminology used in the detailed description of the particular exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting. In the drawings, like numbers refer to like elements.

[0035] It is to be noted, however, that the reference numerals used herein illustrate only typical embodiments of the present subject matter, and are therefore, not to be considered for limiting its scope, for the subject matter may admit to other equally effective embodiments. The specification may refer to “an”, “one” or “some” embodiment(s) in several locations. This does not necessarily imply that each such reference is to the same embodiment(s), or that the feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments.

[0036] As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless expressly stated otherwise. It will be further understood that the terms “includes”, “comprises”, “including” and / or “comprising” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It will be understood that when an element is referred to as being “connected” or “coupled” to another element, it can be directly connected or coupled to the other element or intervening elements may be present. Furthermore, “connected” or “coupled” as used herein may include operatively connected or coupled. As used herein, the term “and / or” includes any and all combinations and arrangements of one or more of the associated listed items.

[0037] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure pertains. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0038] The figures depict a simplified structure only showing some elements and functional entities, all being logical units whose implementation may differ from what is shown. The connections shown are logical connections; the actual physical connections may be different. It is apparent to a person skilled in the art that the structure may also comprise other functions and structures. Also, all logical units described and depicted in the figures include the software and / or hardware components required for the unit to function. Further, each unit may comprise within itself one or more components which are implicitly understood. These components may be operatively coupled to each other and be configured to communicate with each other to perform the function of the said unit.

[0039] Figure 2 exemplarily illustrates a schematic drawing that shows the manner in which MRI imaging and deep learning models can be minimized along with increase in adoption. MRI is a diverse modality with multiple contrasts, anatomies and protocols specific to each subject. However, Al-based approaches or deep learning methods operate only for a specific contrast or anatomy dataset. This creates a gap between the diverse nature of MRI and the task-specific nature of deep learning methods. We need to minimize this gap to make MRI workflow with heterogeneous contrast acquisition not faster with high quality but also with improved throughput. This enables better adoption of MRI and scale its benefits of non-invasiveness in the healthcare industry

[0040] DEFINITIONS:

[0041] The term “MRI scanner” is a medical imaging device that uses magnetic fields and radio waves to produce detailed images of internal body structures. Here, the MRI scanner generates scanner measurements and acquires under-sampled data with embedded MRI contrast information.

[0042] The term “Training data preparation module” is a system or software component that processes and organizes data for use in training machine learning models. In context the Training data preparation module prepares training data from the acquired under-sampled data.

[0043] The term “Partitioning module” is a component that divides data into distinct subsets or partitions for further processing or analysis. In context the Partitioning module partitions the prepared training data into support and query datasets for each contrast.

[0044] The term “Kernel modulation weight prediction network” is a neural network designed to predict parameters (weights) that modulate the behavior of kernels in computational models. In context, the Kernel modulation weight prediction network predicts kernel modulation weights based on MRI contrast information.

[0045] The term “Processor” is the central unit in a computing system that executes instructions and performs calculations. In context, the processor controls the kernel modulation weight prediction network and two -level optimization module to perform image reconstruction.

[0046] The term “Two-level optimization module” is a system that applies two distinct levels of optimization processes to refine a model or parameters. In context the Two- level optimization module refines the kernel modulation weights through inner and outer level corrections using training data.

[0047] The term “Base reconstruction convolutional neural network (CNN) module” is a type of deep learning model that applies convolutional operations to process and analyze data, particularly images. In context the Base reconstruction convolutional neural network (CNN) module applies CNN kernel modulation based on predicted weights to generate a new contrast-specific K-space.

[0048] The term “New contrast- specific K-space” is the frequency domain representation of MRI data, tailored to specific imaging conditions or parameters. In context the New contrast- specific K-space is generated by the CNN module based on predicted kernel modulation weights.

[0049] The term “Final image” is the end result of a reconstruction or imaging process, displaying the processed and analyzed data. In context the final image is the reconstructed MRI image resulting from combining scanner measurements and the new contrast- specific K-space.

[0050] The term “T1 Contrast” refers to imaging based on T1 relaxation time, which highlights tissues with shorter T1 relaxation times, often used to differentiate between fat and water. In context, T1 Contrast provides specific information to help the kernel modulation weight prediction network in generating accurate image reconstructions.

[0051] The term “T2 Contrast” refers to imaging based on T2 relaxation time, emphasizing tissues with longer T2 relaxation times, useful for identifying fluid-filled structures. In context, T2 Contrast is used to adjust the kernel modulation weights to accurately reconstruct images showing tissues with longer T2 relaxation times. FLAIR (Fluid-Attenuated Inversion Recovery): The term “FLAIR” refers to a specialized MRI technique that suppresses fluid signals (such as cerebrospinal fluid) to enhance the visibility of other structures or abnormalities. In context, FLAIR provides unique contrast information that helps in modulating kernel weights to accurately reconstruct images where fluid suppression is required.

[0052] PD (Proton Density): The term “PD Contrast” refers to imaging based on the density of hydrogen protons in tissues, useful for distinguishing between different types of tissue based on proton density. In context, PD Contrast is used to adapt the kernel modulation weights for accurate imaging of tissues with varying proton densities.

[0053] IR (Inversion Recovery): The term “IR Contrast” refers to a technique that uses inversion pulses to enhance the contrast between different tissue types, often used to suppress certain signals or highlight others. In context, IR Contrast helps in modulating the CNN kernel to accurately reconstruct images by enhancing contrast between different tissue types.

[0054] Figure 3A exemplarily illustrates a schematic system drawing that shows an Al-based MRI system (100) for reconstructing images, as an example embodiment of the present disclosure. The Al-based MRI system (100) for reconstructing images includes several interconnected components to achieve high-quality MRI imaging. Central to this system is the MRI scanner (102), which acquires scanner measurements (102a). This scanner acquires corresponding under-sampled data (104a) and embeds MRI contrast information (104b). The under-sampled data (104a) are then forwarded to a training data preparation module (106) connected to the MRI scanner (102). This module is responsible for preparing training data from the acquired under-sampled data (104a) and corresponding fully-sampled data, setting the stage for further processing and partitioning.

[0055] The prepared training data are passed on to a partitioning module (108) that communicates with the training data preparation module (106). This partitioning module (108) is designed to divide the prepared training data into support and query datasets for each contrast. This partitioning is essential for training and optimizing the subsequent network components. Following this, the kernel modulation weight prediction network (110) comes into function. This kernel modulation weight prediction network (110), in communication with the MRI scanner (102) and controlled by at least one processor (112), receives the MRI contrast information (104b) and predicts the necessary kernel modulation weights.

[0056] The next crucial component is the base reconstruction convolutional neural network (CNN) module (114), which is also controlled by the processor (112). This CNN module (114) applies CNN kernel modulation based on the predicted kernel modulation weights. By doing so, it generates a new contrast- specific K-space (116) which gives the final reconstructed MRI image. The under-sampled image reconstruction is obtained by the interplay between kernel modulation network and the reconstruction network. This reconstruction process is adaptable to multiple MRI contrast in a single training.

[0057] In an embodiment, to further enhance the system’s performance, the training process includes a two-level optimization module (118) controlled by the processor (112). This two-level optimization module (118) performs correction and evaluation iteratively to refine the kernel modulation weights. The two-level optimization module (118) operates in two distinct stages: an inner level correction using the support training data and an evaluation using the query training data, followed by an outer level correction using the query training data. This iterative process ensures the continuous refinement of the kernel modulation weights for better image reconstruction.

[0058] In an embodiment, the kernel modulation weight prediction network (110) plays a pivotal role in enabling the Al system to function similarly to attentional modulation in the human brain. By utilizing the predicted modulation weights, the kernel modulation weight prediction network (110) modulates the parameters of a main neural network, enhancing its ability to adapt to different imaging contexts. This analogy to human brain function underlines the sophisticated nature of the Al-based MRI system in managing and processing diverse MRI data. The kernel modulation weights allow the base reconstruction convolutional neural network (CNN) module (114) to adapt to various MRI contrasts, including but not limited to Tl, FLAIR, PD, and T2 contrast data. This adaptability is crucial for handling the diverse range of MRI imaging needs, ensuring that the system can provide high-quality reconstructions across different types of MRI scans. This flexibility significantly enhances the utility and effectiveness of the MRI system in clinical settings.

[0059] The two-level optimization module (118) further refines the kernel modulation weights through its inner loop correction step, which involves faster adaptation to each contrast data using a support data partition. The outer level correction step comprises updating the kernel modulation weight prediction network (110) based on the average error computed from the query data partition across all contrasts. This dual-level optimization ensures that the kernel modulation weight prediction network (110) continuously learns and improves, minimizing errors and enhancing overall image quality. Ultimately, the final trained base reconstruction convolutional neural network (CNN) module (114), labeled as 0_new, utilizes the optimized kernel modulation weights to reconstruct MRI images, resulting in the final high-quality reconstructed image (120).

[0060] The invention is a new process that can potentially be integrated into the MRI workflow. The invention is used for improving the quality of multi-contrast MRI scans. The invention provides a new use and improves an existing method. The invention considers variations occurring in the form of intensity ranges and under- sampling artifact amounts which are significant. Moreover, common knowledge present across various artifacts is modeled so that the proposed invention performs reasonably even in the presence of hitherto unseen artifacts.

[0061] Figure 3B exemplarily illustrates a Flowchart of the learning process, as an example embodiment of the present invention, as an example embodiment of the present disclosure. The flow chart illustrates an Al-based MRI system designed for reconstructing images through a detailed multi-step process. Initially, an MRI scanner (102) generates multiple contrasts (104b) (e.g., MRI Contrast 1, MRI Contrast 2, up to MRI Contrast n). These contrasts (104b) are fed into a kernel modulation weight prediction network (110) that extracts contrast-specific information (202). This information is then processed by an adjustment deep learning (DL) model (2204), which outputs a contrast- specific model (206). The contrast- specific model (206) generates adjustment parameters (208) that are specific to each contrast type. These adjustment parameters (208), along with the raw MRI contrasts (104b), are fed into a reconstruction DL model (210). The outputs from the reconstruction DL model (210) are then evaluated and corrected by an evaluation and correction mechanism (212). The final output (214) of this process is the DL model's (210) reconstructed MRI images, which are refined and optimized based on the contrast- specific adjustments and corrections. This Al-based MRI system (100) aims to enhance the quality and accuracy of the reconstructed MRI images by leveraging contrast-specific information and advanced deep learning techniques. The Al-based MRI system (100) incorporates the following to improve the quality of degraded MRI scans caused by artifacts during acquisition:

[0062] 1) Multi-contrast dataset(s): Various anatomical MRI contrasts like Tl, T2, FLAIR, PD, and IR are considered for the process. The contrast and brightness of the images are different and are predominately determined by the properties of the tissue. All images are affected by the under- sampling of measurements, a type of degradation that is induced by the acquisition settings. The long acquisition duration is shortened at the cost of image quality.

[0063] 2) Data pre-processing: Standardizing the various artifact datasets using by cropping the Region of interest (ROI) defined by the anatomy.

[0064] 3) Reconstruction DL model: A parameterized neural network architecture that facilitates a functional mapping from the input degraded image to the corresponding clean image at its output. The model consists of convolutional layers, batch norm, and nonlinear activation function. The parameters of these modules are appropriately adjusted to effectively improve the quality of the input image.

[0065] 4) MRI contrast- adjustment DL model: This is another neural network that takes information about the range of intensity levels of each contrast and drives the reconstruction DL model to provide the contrast output. This neural network adjusts the parameters of the reconstruction DL model according to the contrast type. 5) Correction mechanism and evaluation mechanism: The correction mechanism of multi-contrast MRI DL model involves two main steps: Local Optimization: In this step, the contrast adjustment model is exposed to a specific MRI contrast image, also called an MRI context or task, and learns to perform well on it. The contrast adjustment model's parameters are updated through backpropagation using the data from the task. Global optimization: In this step, the reconstruction model and the contrast adjustment model are presented with multiple tasks. The models use the knowledge gained from the local optimization to adapt and learn quickly on these new tasks. The global optimization aims to update the common parameters of the models (which determine how the contrast adjustment model learns in the local optimization step) based on how well the model performed on the various tasks. By iteratively updating the common parameters of the models through the global optimization and updating the models' parameters through the local optimization step, the multi-task learning model becomes better at quickly adapting to new tasks, making it a powerful approach for multiple MRI contrast scenarios.

[0066] 6) Image quality assessment measure: A metric that is used to quantify the image quality of the output provided by the model.

[0067] Referring to Figures 4A and 4B, Figure 4A exemplarily illustrates a pictorial representation of the task- specific inner loop adaptation that becomes faster in few gradient steps and Figure 4B exemplarily illustrates a pictorial representation showing meta weight update and outer loop optimization providing meta-initialization, as example embodiments of the present disclosure. The kernel modulation weight prediction network (110) plays a crucial role in the modulation weight prediction process. It communicates with the training process or rule that governs the learning of the parameters of the modulation weight prediction network (110). By predicting modulation weights and using these weights to modulate (i.e., multiply) the parameters of the main neural network, this process enables the Al system to function similarly to attentional modulation in the brain. The kernel modulation weights allow the main reconstruction network to adapt to widely varying MRI contrasts, such as Tl, FLAIR, PD, and T2 contrast data. The training process begins by dividing the dataset of each contrast data into two partitions: support and query. Initially, the modulation network (110), which is randomly initialized, is adapted to each contrast using the support data partition. This creates a contrast- specific model through a step known as the inner loop or fast adaptation to each contrast. Each arrow in the figure below represents an inner loop correction step. Following this, the query data is passed through each contrast- specific model, and the error with respect to the query data is computed. The average error over all contrast data is then determined by aggregating the individual error values. A final correction step, known as meta-weight update or outer loop optimization, is performed. The neural network shown below, labeled as 0_new, represents the final trained modulation neural network.

[0068] The Al-based MRI system (100) is extensively usable in the clinical settings, especially in the multi- contrast MRI workflow to reconstruct images of multiple contrasts quickly. Improving the quality of an artifact image can potentially result in 1) better inference on the treatment procedures, and 2) reducing the time spent by doctors on analyzing the artifact scan; thus, efficiently promoting a better healthcare system. Benefits of the invention include:

[0069] • No deterioration of the decision-making process in the scenario of degradations arising from under-sampling.

[0070] • Long duration MRI acquisition process can be shortened without any loss in the image quality. It results in better MRI scanner utilization by increasing the throughput of the number of patients scanned.

[0071] • Few samples of multiple contrasts (around 50 images) is sufficient to train the system.

[0072] It will be understood that each block of the block diagrams can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0073] In the drawings and specification, there have been disclosed exemplary embodiments of the invention. Although specific terms are employed, they are used in a generic and descriptive sense only and not for purposes of limitation of the scope of the invention.

[0074] The foregoing description, for purposes of explanation, used specific nomenclature to provide a thorough understanding of the disclosure. However, it will be apparent to one skilled in the art that the specific details are not required in order to practice the systems and methods described herein. The foregoing descriptions of specific examples are presented for purposes of illustration and description. They are not intended to be exhaustive of or to limit this disclosure to the precise forms described. Obviously, many modifications and variations are possible in view of the above teachings. The examples are shown and described in order to best explain the principles of this disclosure and practical applications, to thereby enable others skilled in the art to best utilize this disclosure and various examples with various modifications as are suited to the particular use contemplated.

Claims

We Claim:

1. An Al-based MRI system (100) for reconstructing images, comprising: an MRI scanner (102) to acquire scanner measurements (102a) and acquire corresponding under- sampled data (104a) and embed MRI contrast information (104b); a training data preparation module (106) connected to the MRI scanner (102), to prepare training data from the acquired under- sampled data and corresponding fully sampled data (104a); a partitioning module (108) in communication with the training data preparation module (106) to partition the prepared training data into support and query datasets for each contrast; a kernel modulation weight prediction network (110) in communication with the MRI scanner (102) and controlled by at least one processor (112) to receive the MRI contrast information (104b) and predict kernel modulation weights; a two-level optimization module (118) in communication with the kernel modulation weight prediction network (110) and controlled by the processor (112) to perform correction and evaluation iteratively to refine the kernel modulation weights, wherein the two-level optimization module (118) performs:• an inner level correction of the weights of the kernel modulation weight prediction network (110) using the support training data, along with an evaluation using query training data, and• an outer level correction of these weights using the query training data, and a base reconstruction convolutional neural network (CNN) module (114) controlled by the processor (112) applies CNN kernel modulation based on the predicted kernel modulation weights, wherein the CNN module (114) generates a new contrast- specific K-space (116) that provides the final reconstructed MRI image (120).

2. The system as claimed in claim 1, wherein the kernel modulation weight prediction network (110) enables an Al system to function similarly to an attentionalmodulation in human brain by using the predicted modulation weights to modulate the parameters of a main neural network.

3. The system as claimed in claim 1, wherein the kernel modulation weights enable the base reconstruction convolutional neural network (CNN) module (114) to adapt to varying MRI contrasts, including but not limited to: Tl, FLAIR, PD, and T2 contrast data.

4. The system as claimed in claim 1, wherein the inner loop correction step involves faster adaptation to each contrast data using a support data partition.

5. The system as claimed in claim 1, wherein the outer level correction step comprises updating the kernel modulation weight prediction network (110) based on the average error computed from the query data across all contrasts.

6. The system as claimed in claim 1, wherein a final trained base reconstruction convolutional neural network (CNN) module (114), labelled as 0_new, utilizes the optimized kernel modulation weights for reconstructing MRI images as the final image (120).

7. The system as claimed in claim 1, wherein the under-sampled image reconstruction is obtained by interplay between kernel modulation weight prediction network (110) and the base reconstruction convolutional neural network (CNN) module (114) , wherein this reconstruction process is adaptable to multiple MRI contrast in a single training.

8. A method for reconstructing images using an Al-based MRI system (100), comprising: acquiring scanner measurements (102a) and acquiring corresponding undersampled data (104a) and MRI contrast information (104b) using an MRI scanner (102);preparing training data from the acquired under- sampled corresponding fully- sampled data (104a) using a training data preparation module (106); partitioning the prepared training data into support and query datasets for each MRI contrast using a partitioning module (108); predicting kernel modulation weights based on the MRI contrast information (104b) using a kernel modulation weight prediction network (110) controlled by at least one processor (112); performing correction and evaluation iteratively to refine the kernel modulation weights using a two-level optimization module (118) controlled by the processor (112), wherein the two-level optimization module (118) performs: o an inner level correction of the weights of the kernel modulation weight prediction network (110) using the support training data, along with an evaluation using query training data, and o an outer level correction of these weights using the query training data, and applying convolutional neural network (CNN) kernel modulation based on the predicted kernel modulation weights to generate a new contrast- specific K-space (116) using a base reconstruction CNN module (114) controlled by the processor (112) that provides the final reconstructed MRI image (120).

9. The method as claimed in claim 8, wherein predicting kernel modulation weights enables an Al system to function similarly to attentional modulation in the human brain by using the predicted modulation weights to modulate the parameters of a main neural network.

10. The method as claimed in claim 8, wherein the kernel modulation weights enable the base reconstruction convolutional neural network (CNN) module (114) to adapt to varying MRI contrasts, including but not limited to Tl, FLAIR, PD, and T2 contrast data.

11. The method as claimed in claim 8, wherein the inner level correction step involves faster adaptation to each MRI contrast data using the support data partition.

12. The method as claimed in claim 8, wherein the outer level correction step comprises updating the kernel modulation weight prediction network (110) based on the average error computed from the query data across all MRI contrasts.

13. The method as claimed in claim 8, wherein a final trained base reconstruction convolutional neural network (CNN) module (114) utilizes the optimized kernel modulation weights for reconstructing MRI images as the final image (120).

14. The method as claimed in claim 8, wherein the under-sampled image reconstruction is obtained by interplay between kernel modulation weight prediction network (110) and the base reconstruction convolutional neural network (CNN) module (114), wherein this reconstruction process is adaptable to multiple MRI contrast in a single training.

Images