Method and system for black blood T2 weighted fast spin echo imaging under free breathing condition

Abstract

The invention relates to the technical field of medical image diagnosis, in particular to a black blood T2 weighted fast spin echo imaging method and system under a free breathing condition. According to the black blood imaging method based on reverse dual inversion recovery (RDIR) and single excitation fast spin echo (SS-FSE), artificial intelligence assisted compressed sensing (ACS) is introduced to improve the reconstruction quality. The slice selective reversal pulse is moved forward to the previous cardiac cycle, so that myocardial signal loss is effectively reduced; scanning time can be remarkably shortened and motion artifacts can be reduced by adopting single excitation. And in combination with the ACS, a high-resolution T2 weighted black blood image can be obtained under high and under-sampling conditions, and the method is particularly suitable for edema evaluation of children under free breathing.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image diagnosis, and particularly to a method and system for performing black-blood T2-weighted fast spin echo imaging under free-breathing conditions. Background Art

[0002] Black-blood T2-weighted fast spin echo imaging (Double Inversion Recovery Fast Spin Echo, abbreviated as DB-FSE) is a commonly used sequence for evaluating myocardial edema, and has important diagnostic value especially in acute myocarditis, ischemia-reperfusion injury, and other pathological conditions that cause an increase in abnormal water content in the myocardium. Its basic principle is to apply a Double Inversion Recovery (DIR) pulse during the cardiac cycle to suppress blood signals and thus obtain a "black-blood" contrast, so as to have a good contrast between the myocardium and the blood pool, thereby facilitating the observation of internal inflammation or edema in the myocardium.

[0003] However, traditional DB-FSE is extremely sensitive to motion (especially cardiac motion and respiratory motion), requiring the patient to hold their breath strictly during the scan, and the heart rate should not be too fast, which is often difficult to meet for pediatric patients. Pediatric patients generally have challenges such as a relatively high heart rate and low cooperation, resulting in the following problems often occurring in cardiac MRI images under free-breathing conditions:

[0004] 1. Signal Loss Artifacts: Due to slice mismatch of the DIR preparation pulse, myocardial spin signals are erroneously suppressed to a certain extent; the rapid motion of the heart will also exacerbate this problem.

[0005] 2. Motion Artifacts: In multi-shot acquisitions, breathing and heartbeats occur simultaneously. If the acquisition period is long, severe ghost artifacts will be generated, affecting the accuracy of myocardial boundary and lesion assessment.

[0006] 3. Long scan time: Multi-shot DB-FSE usually requires 3 to 5 minutes or even longer to obtain sufficient resolution and signal-to-noise ratio, and it is difficult for pediatric patients to remain quiet breathing or hold their breath during this period.

[0007] In order to overcome the above problems, the Reverse Double Inversion Recovery (RDIR) technique was proposed to shorten the time mismatch between the DIR preparation pulse and the fast spin echo (FSE) echo train, thereby significantly reducing the myocardial signal loss caused by slice mismatch. Previously, the application effect of RDIR technology in adult cardiac imaging has been reported in the literature, but the imaging performance in pediatric patients with high heart rate and free breathing is still lacking in-depth research.

[0008] On the other hand, single-shot fast spin echo technology can complete the acquisition of the entire image within one cardiac cycle, greatly shortening the scanning time and reducing respiratory motion artifacts. However, traditional single-shot DB-FSE (such as HASTE sequence) is susceptible to the influence of longer echo trains (Echo Train) under high-resolution requirements, resulting in a decrease in image signal-to-noise ratio and limited resolution, which makes it difficult to meet the needs of observing the fine structure of the myocardium in children. With the emergence of artificial intelligence-assisted compressed sensing (ACS) technology, image quality comparable to conventional acquisition can still be obtained under conditions of higher undersampling factors. Existing studies have verified the feasibility of ACS in lumbar spine, knee joint, abdomen and black blood heart imaging. However, AI-assisted single-shot imaging technology still lacks a mature solution for the assessment of black blood myocardial edema in children in a free-breathing state.

[0009] Based on this, how to combine RDIR technology with single-shot acquisition and use deep learning-assisted reconstruction methods to solve the artifact and noise problems caused by high undersampling has become an important issue in optimizing black blood T2 imaging of myocardium in children with free breathing. If high-resolution, motion-robust black blood T2 images can be obtained in a relatively short acquisition time, the diagnostic efficiency and accuracy of myocarditis and other myocardial edema diseases in children will be greatly improved, while reducing the need for sedation or anesthesia for children and alleviating the burden on children and their parents. Summary of the invention

[0010] In order to solve the above-mentioned technical problems, the purpose of the present invention is to provide a method for performing black blood T2-weighted fast spin echo imaging under free breathing conditions. The method combines the reverse double inversion recovery (RDIR) and AI-assisted compressed sensing (ACS) single-shot DB-FSE sequence to achieve accurate assessment of myocardial edema in children under free breathing conditions.

[0011] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0012] A method for performing black blood T2-weighted fast spin echo imaging under free breathing conditions, the method at least including the following steps:

[0013] 1) Pulse sequence design: Set a reverse dual inversion recovery (RDIR) preparation pulse before each imaging cardiac cycle to achieve inversion of blood signal and minimize signal loss caused by slice mismatch;

[0014] 2) Single-shot acquisition: After RDIR preparation, use a single-shot fast spin echo (SS-RDIR) sequence to obtain at least one frame of k-space data covering the left ventricle;

[0015] 3) Data undersampling: Perform undersampling of the k-space data during the single-shot acquisition process in a compressed sensing (CS) manner to shorten the scanning time;

[0016] 4) AI-assisted reconstruction: Based on the artificial intelligence-assisted compressed sensing (ACS) technology, perform an image reconstruction process combining deep learning and iterative reconstruction on the undersampled k-space data to obtain a high-resolution T2-weighted black blood image with reduced motion artifacts.

[0017] Preferably, the reverse dual inversion recovery RDIR preparation pulse includes:

[0018] a) Apply a slice-selective inversion pulse in the mid-diastole of the previous cardiac cycle;

[0019] b) Apply a non-selective inversion pulse in the next cardiac cycle after the slice-selective inversion pulse;

[0020] c) Match the time interval between the non-selective inversion pulse and the fast spin echo (FSE) echo train to minimize myocardial signal loss caused by slice mismatch.

[0021] Preferably, in the single-shot fast spin echo (SS-RDIR) acquisition, the number of k-space lines collected for each slice is between 46 and 52, so as to complete the sampling of the entire image within one cardiac cycle.

[0022] Preferably, the reconstruction steps of the artificial intelligence-assisted compressed sensing (ACS) technology include:

[0023] a) Perform preliminary artifact removal and noise reduction processing on the undersampled k-space data through a trained deep neural network to obtain a preliminary reconstructed image;

[0024] b) Use this preliminary reconstructed image as a prior constraint and combine it with a conventional compressed sensing iterative algorithm for further iterative optimization;

[0025] c) Output the final high-fidelity T2-weighted black blood image;

[0026] More preferably, the deep neural network used in the artificial intelligence-assisted reconstruction process is a convolutional neural network (CNN) or a hybrid network containing a Transformer structure, which is used to reduce artifacts and retain image details at high acceleration factors.

[0027] Preferably, the method further includes:

[0028] a) Dynamically adjusting the trigger time between the RDIR preparation pulse and the single-shot FSE echo train according to the real-time heart rate of the child;

[0029] b) Combining at least one signal averaging or registration to enhance the overall visibility and contrast of the myocardium;

[0030] And / or, the method further includes:

[0031] a) Using electrocardiogram or pulse wave-based triggering to determine the acquisition window;

[0032] b) Controlling the entire scanning time to be completed within 30 to 60 seconds to minimize artifacts caused by respiratory inconsistency;

[0033] And / or, the method further includes performing motion correction on the data frames obtained by multiple averaging before image reconstruction to align the myocardial positions before averaging, thereby further reducing motion artifacts and improving image quality.

[0034] Furthermore, the present invention also provides a magnetic resonance imaging system for implementing the imaging method, and the system includes:

[0035] 1) A magnet and gradient coil assembly for generating a uniform static magnetic field and performing pulse gradients at a magnetic field strength of 3T;

[0036] 2) A radio frequency transmitting and receiving unit including at least one multi-channel radio frequency coil for transmitting RDIR preparation pulses and receiving signals;

[0037] 3) A sequence control module programmed to execute a pulse sequence including reverse dual inversion recovery (RDIR) and single-shot fast spin echo (SS-RDIR);

[0038] 4) A data acquisition and storage module for acquiring and storing undersampled k-space data;

[0039] 5) An artificial intelligence-assisted compressed sensing reconstruction module configured to reconstruct the undersampled k-space data using a pre-trained deep learning model and output a final T2-weighted black blood image;

[0040] 6) A display and evaluation module for displaying the reconstructed black blood image and providing a subsequent diagnostic analysis interface.

[0041] Preferably, the artificial intelligence-assisted compressed sensing reconstruction module includes:

[0042] a) A neural network processing unit that pre-loads a deep learning model and is used to generate a preliminary reconstructed image and correct motion artifacts;

[0043] b) An iterative optimization unit that combines the preliminary reconstruction result with the sparsity constraint of compressed sensing and performs multiple iterative updates to obtain a final high-resolution myocardial T2-weighted black blood image;

[0044] c) A parameter scheduling unit that dynamically adjusts the trigger timing and undersampling trajectory of RDIR preparation and single-shot acquisition according to the heart rate and breathing pattern of the child.

[0045] Furthermore, the present invention also provides an application of the method in designing and evaluating software for myocardial edema.

[0046] Furthermore, the present invention also provides a computer-readable storage medium on which a computer program or instruction is stored, and when the computer program or instruction is executed by a processor, the method is implemented.

[0047] Furthermore, the present invention also provides a computer program product that includes a computer program or instruction, and when the computer program or instruction is executed by a processor, the method is implemented.

[0048] Due to the adoption of the above technical solution, the present invention has the following technical effects:

[0049] 1. Significantly shorten the scanning time: By adopting single-shot acquisition (Single-Shot FSE) and combining it with artificial intelligence-assisted compressed sensing (AI-assisted Compressed Sensing, ACS) technology, the present invention can complete the data acquisition of the entire image within a short cardiac cycle. Compared with the traditional multi-shot method, the scanning time is significantly shortened, effectively improving the examination efficiency and reducing the restlessness and discomfort of child patients caused by long-time scanning.

[0050] 2. Reduce motion artifacts: The time position of the RDIR (Reverse Double Inversion Recovery) preparation pulse in the cardiac cycle is optimized, reducing signal loss caused by slice mismatch; while single-shot acquisition significantly reduces the ghosting generated by respiratory or cardiac displacement between multiple shots. The solution of the present invention is particularly suitable for children with high heart rates and free breathing, and can significantly suppress artifacts caused by respiratory movement, heart beating, etc., thus ensuring the accurate presentation of myocardial tissue structure.

[0051] 3. Obtain higher image resolution and signal-to-noise ratio: Although single-shot acquisition inherently has limitations such as a long echo train and easy resolution degradation, through artificial intelligence-assisted compressed sensing reconstruction (ACS), good signal-to-noise ratio and tissue contrast can be maintained at a relatively high undersampling factor. Therefore, while shortening the acquisition time, the present invention can still clearly display myocardial edema and other subtle pathological features in the image.

[0052] 4. Enhance the clinical diagnostic value: Diseases such as pediatric myocarditis often require rapid and accurate assessment of myocardial edema. The present invention can provide high-quality black-blood T2 images in the shortest scanning time, helping doctors quickly and intuitively judge the degree of myocardial inflammation or other lesion sites, improving the diagnostic efficiency and accuracy; at the same time, reducing or avoiding the need for pediatric sedation and anesthesia, and reducing the risks and costs of imaging examinations.

[0053] 5. Strong adaptability and scalability: The RDIR combined with single-shot technology in the present invention can flexibly adjust the triggering and undersampling strategies according to the real-time heart rate and respiratory pattern of children. Coupled with the continuous iteration and upgrade of the AI reconstruction algorithm, the present invention has strong adaptability and is also easy to expand and apply to other cardiovascular examination fields that require rapid and high-resolution black-blood imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 . Schematic diagram of the DB-FSE sequence; wherein: (A) Conventional multi-shot DB-FSE (MS-DIR) based on standard DIR; (B) Multi-shot DB-FSE (MS-RDIR) based on RDIR; (C) Single-shot DB-FSE (SS-RDIR) based on RDIR; ECG: electrocardiogram; RF: radio frequency; IRns: non-selective inversion recovery; IRsel: slice-selective inversion recovery; FatSat: fat suppression.

[0055] Figure 2 . Quantitative comparison of MS-DIR, MS-RDIR, and SS-RDIR in 47 subjects. MS-DIR: Multi-shot DB-FSE based on DIR; MS-RDIR: Multi-shot DB-FSE based on RDIR; SS-RDIR: Single-shot DB-FSE based on RDIR. The symbol \ / \\ indicates P<0.05 / 0.01.

[0056] Figure 3. Qualitative comparison of MS-DIR, MS-RDIR, and SS-RDIR among 47 subjects, with a scoring range of 1 (worst) - 5 (best). MS-DIR: Multi-shot DB-FSE based on DIR; MS-RDIR: Multi-shot DB-FSE based on RDIR; SS-RDIR: Single-shot DB-FSE based on RDIR. The symbol \ / \\ indicates P < 0.05 / 0.01.

[0057] Figure 4 . Comparison of cardiac images of MS-DIR and MS-RDIR in two patients. Signal loss (yellow arrow) was observable in MS-DIR. MS-RDIR produced fewer signal loss artifacts than MS-DIR. MS-DIR: Multi-shot DB-FSE based on DIR; MS-RDIR: Multi-shot DB-FSE based on RDIR; SS-RDIR: Single-shot DB-FSE based on RDIR.

[0058] Figure 5 . Comparison of cardiac images of MS-RDIR and SS-RDIR in two patients. Motion-related ghost artifacts (green arrow) were observable in MS-RDIR. SS-RDIR significantly suppressed the ghost artifacts. MS-DIR: Multi-shot DB-FSE based on DIR; MS-RDIR: Multi-shot DB-FSE based on RDIR; SS-RDIR: Single-shot DB-FSE based on RDIR.

[0059] Figure 6 . Two-slice images of MS-DIR, MS-RDIR, SS-RDIR, LGE, and T2 maps in a patient with [disease name]. Motion-related ghost artifacts (green arrow) were observable in both MS-DIR and MS-RDIR. Consistent with the LGE image and T2 Figure 1 maps, SS-RDIR showed more obvious local T2 elevation (blue arrow) than MS-DIR and MS-RDIR. MS-DIR: Multi-shot DB-FSE based on DIR; MS-RDIR: Multi-shot DB-FSE based on RDIR; SS-RDIR: Single-shot DB-FSE based on RDIR. Detailed implementation manners

[0060] Combined with the embodiments of the present invention below, the technical solutions in the embodiments will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] Embodiment 1

[0062] I. Hardware Environment and Preliminary Preparation

[0063] 1. Magnetic Resonance System and Coil

[0064] Select a 3T magnetic resonance scanner (such as uMR790, United Imaging Healthcare) and cooperate with a 24-channel body coil or a dedicated cardiac coil to balance a large coverage area and a sufficiently high signal acquisition efficiency.

[0065] Enable the system to support electrocardiogram (ECG) / pulse triggering, as well as functional modules with parallel imaging and compressed sensing scanning capabilities.

[0066] 2. Preparation of the Child to be Examined

[0067] Place the child supine on the examination table and let them breathe freely without the need for special breath-holding. For younger or uncooperative children, psychological comfort can be provided with the accompaniment of parents or caregivers to minimize motion interference as much as possible.

[0068] Continuously acquire electrocardiogram or pulse wave to achieve real-time synchronization and triggering of the cardiac cycle.

[0069] 3. Setting of Scanning Parameters

[0070] Field of View (FOV): 360×260mm 2 Or adjust appropriately according to the child's body size;

[0071] Slice Thickness: Approximately 6mm;

[0072] Voxel Size: 1.4×1.76mm 2 ;

[0073] Bandwidth (Hz / pixel): A higher bandwidth is recommended, such as 700Hz / pixel, to shorten the echo interval;

[0074] Echo Train Length (ETL): Approximately 46 - 52 under single-shot excitation;

[0075] Flip Angle: 120°;

[0076] Combined Acceleration Factor: 3 - 4;

[0077] Triggering Method: ECG triggering, or select pulse triggering for children with high heart rates.

[0078] II. Setting of the RDIR Preparation Pulse

[0079] 1. Basic Process of Reverse Dual Inversion Recovery (RDIR)

[0080] First, a slice-selective inversion pulse (IRsel) is applied in the mid-diastole of the previous cardiac cycle (Cycle N); then, a non-selective inversion pulse (IRns) is applied at the beginning of diastole or early systole of the next cardiac cycle (Cycle N+1), and the fat saturation module and fast spin echo (FSE) acquisition window are entered.

[0081] 2. Determination of the time interval (TI)

[0082] For the relatively fast heart rate of children under 3T magnetic field (such as the R-R interval being 600 - 800 ms), the interval between IRsel and IRns should be set in the range of 400 - 600 ms to ensure that the blood signal approaches zero longitudinal magnetization before the start of readout, so as to achieve a good black blood effect; clinically, it can be adjusted according to the heart rate of the child to avoid incomplete blood suppression or myocardial signal decline caused by too long or too short TI.

[0083] III. Single-shot FSE acquisition

[0084] 1. Single-Shot mode

[0085] Within one cardiac cycle, a complete set of k-space data is covered by a long echo train (about 46 - 52 echoes). The number of cardiac cycles required for each slice acquisition is about 6 (including the preparation and readout processes). If multi-slice coverage is needed, RDIR is repeatedly applied in the subsequent several cardiac cycles and single-shot acquisition of the next slice is performed. Usually, 5 - 11 slices can cover the main myocardial regions.

[0086] 2. Undersampled trajectory

[0087] The k-space is randomly or semi-randomly undersampled using compressed sensing. Non-uniform density (Poisson-disk) sampling can be used to ensure a higher sampling density in the central low-frequency part to retain the main tissue contrast information, while the peripheral high-frequency region is sparsely sampled to increase the acceleration ratio. Parallel imaging and compressed sensing can be used in combination, and the comprehensive acceleration factor can be set to 3 - 4, so that sufficient spatio-temporal resolution can still be obtained in the single-shot mode.

[0088] IV. AI-assisted compressed sensing (ACS) image reconstruction

[0089] 1. Reconstruction algorithm structure

[0090] Pre-trained deep neural network: The undersampled data is input into the network, and the network preliminarily removes high-frequency artifacts and reduces noise; iterative optimization: The network output is used as a prior, combined with the sparsity and data consistency in the compressed sensing algorithm, and 10 - 15 iterations (can be adjusted according to actual needs) are performed, and finally a high-fidelity black blood image is output.

[0091] 2. Processing Hardware Support

[0092] Reconstruction is usually carried out on a local workstation or cloud server equipped with a high-performance GPU, ensuring that the reconstruction of multi-slice images for a single scan can be completed within dozens of seconds to 1 minute, meeting the timeliness requirements of clinical diagnosis.

[0093] 3. Visualization of Reconstruction Results

[0094] The generated black-blood T2 images can be directly displayed on a conventional DICOM workstation or imported into post-processing software for ROI (region of interest) measurement (such as myocardial signal, blood pool signal, and statistics of myocardial edema regions).

[0095] The following process obtains undersampled k-space data (k undersampled ) from an MRI scanning system, and the corresponding sampling trajectory (or sampling mask) is known. Among them, the network model can be pre-trained offline in advance or fine-tuned online.

[0096] 1. Data Preprocessing

[0097] 1) Reading Undersampled k-Space

[0098] Receive the k-space data (k undersampled ) collected by the MRI host or front-end; determine the sampling mask (mask), where Ω represents the set of positions for retained sampling.

[0099] 2) Normalization and Filtering (Optional)

[0100] Perform amplitude normalization on the original k-space data to ensure relatively consistent data distribution under different scanning conditions; perform simple noise filtering (removing radio frequency interference) on the high-frequency region to reduce extreme artifact interference.

[0101] 3) Fast Inverse Fourier Transform (Zero-filling or Preliminary Reconstruction)

[0102] For the input of the subsequent neural network, first perform an inverse Fourier transform (with zero-filling) on the undersampled k-space to obtain a preliminary image x0. x0 usually has obvious artifacts and noise, but can be used as a reference input for the neural network.

[0103] 2. Preliminary Reconstruction by Neural Network

[0104] 1) Deep Network Structure

[0105] Networks such as U-Net, ResNet, or networks containing Transformer blocks can be used, and an attention mechanism can also be incorporated among them to enhance the ability to capture local details and global information.

[0106] Model input: The initial reconstructed image x0 or the merged multi-channel information (such as the complex magnitude image after multi-channel Coil synthesis).

[0107] Model output: The "artifact-removed" or "artifact-suppressed" image x processed by the network DL .

[0108] 2) Network inference process

[0109] Input x0 into the network. Inside the network, through operations such as convolutional layers (or feature extraction layers integrating self-attention mechanisms), skip connections, and activation functions (ReLU / LeakyReLU), etc., the separation of artifact components and the smoothing of noise are achieved, while trying to retain image details and texture structures as much as possible; the output image x DL usually has a smoother background than x0 and significantly reduced artifacts, but there may still be a certain degree of detail error.

[0110] 3) Network parameters and offline training

[0111] It is necessary to prepare a training dataset in advance, including fully sampled or standard-quality images x full and their corresponding undersampled images x under , and perform offline iterative training by minimizing the reconstruction error or perceptual loss functions (such as L1 / L2 Loss, SSIM, or PerceptualLoss); after training is completed, solidify the model parameters and deploy them in the MRI system to achieve online inference (the inference time usually only takes hundreds of milliseconds to several seconds).

[0112] 3. Iterative compressive sensing optimization

[0113] 1) Data consistency constraint

[0114] Regard the neural network output image xD as the prior image and ensure its consistency with the original sampling data in the iterative optimization. Specifically:

[0115]

[0116] F Ω represents the Fourier transform operator only at the sampled positions Ω;

[0117] ||F Ω {x}-k undesampled || 2 is the data consistency term to ensure matching with the measured values in the sampled area;

[0118] R(x,x DL ) is the regularization term or penalty function, for example, considering x DLAs a priori, it guides x to be closer to the output image of the neural network; λ is the regularization weight, which is used to control the balance between data consistency and prior constraints.

[0119] 2) Sparsity / low-rank regularization (optional)

[0120] In this iterative framework, traditional CS regularization can also be applied in parallel or alternately, such as constraining the sparsity (Total Variation, TV) or low-rank (Low-rank) of the image to further remove residual artifacts:

[0121]

[0122] in represents a certain sparse transform (such as wavelet transform, TV operator), and α and β are adjustment coefficients.

[0123] 3) Iterative update mechanism

[0124] Use gradient descent or variational optimization algorithm (such as FISTA) to update x for several rounds (such as 10 to 15 rounds);

[0125] After each round of update, the temporary solution x n+1 It can be input into the neural network again for re-correction (i.e., network-iteration cycle), thus achieving a deep fusion of neural network and CS iteration.

[0126] 4) Termination conditions

[0127] When the number of iterations reaches the preset upper limit or the update amplitude of the data consistency item is less than a certain threshold, the final reconstructed image x is output. final .x final It can be regarded as the best estimation result under under-sampling conditions, which not only conforms to the measured data but also makes full use of the neural network prior and sparsity constraints.

[0128] 4. Output reconstruction results

[0129] x final The images are converted into a medical image format (DICOM) and displayed on an MRI console or workstation; adaptive filtering, wavelet post-processing or local enhancement may be performed on the images again to highlight the myocardial lesion area.

[0130] Test example

[0131] 1. Materials and Methods

[0132] 1. Pulse sequence

[0133] 1.1 DIR-based multi-shot DB-FSE

[0134] Figure 1A shows a conventional multi-shot DB-FSE sequence diagram based on standard DIR (MS-DIR). In standard DIR FSE, a non-selective inversion pulse (IRns) and a slice-selective inversion pulse (IRsel) are sequentially executed in the early systolic phase of the cardiac cycle, while the fat suppression module and FSE readout are sequentially performed in the late diastolic phase. During the multi-shot acquisition process, the entire acquisition process spans multiple heartbeats, with a total of 4-5 shots acquired. The acquisition interval is once every two heartbeats to allow signal recovery.

[0135] 1.2 Multi-shot DB-FSE based on RDIR

[0136] Figure 1 B shows a multi-shot DB-FSE sequence diagram based on RDIR (MS-RDIR). In RDIR, IRsel is executed in the cardiac cycle before IRns to eliminate the temporal mismatch between IRsel and the FSE echo train, thereby minimizing the risk of slice misregistration. The total acquisition time of MS-RDIR is the same as that of MS-DIR.

[0137] 1.3 Single-shot DB-FSE based on RDIR

[0138] Figure 1 C shows a single-shot DB-FSE sequence diagram based on RDIR (SS-RDIR). During the single-shot acquisition process, 46-52 k-space lines are acquired for each slice, reducing the total number of acquisition shots to 1. To maintain image quality under high acceleration conditions, an artificial intelligence-assisted compressed sensing (ACS) reconstruction method is applied to reconstruct the SS-RDIR images. Briefly, the ACS method reduces artifacts and noise in the image reconstructed from highly undersampled k-space by training a deep neural network. Then, the network output is added as a constraint condition to the conventional compressed sensing (CS) algorithm for image estimation. Finally, the result of the CS algorithm is used as the final reconstructed image.

[0139] 2. Patient population

[0140] This invention has obtained the approval of the Institutional Ethics Committee of Children's Hospital of Fudan University. All patients or their legal guardians have signed the informed consent form. From January 2024 to April 2024, a total of 47 pediatric patients (age range, 0.5 - 15 years) were included in this invention.

[0141] 3. Image acquisition

[0142] All patients were imaged using a 3T magnetic resonance scanner (uMR790, United Imaging Healthcare, Shanghai, China) and a 24-channel body coil at the Children's Hospital of Fudan University. All patients were scanned in the free-breathing state with three sequences (MS-DIR, MS-RDIR, and SS-RDIR). Different patients were acquired 5 to 11 slices according to the protocol to cover the entire left ventricle. To reduce motion artifacts, all methods used 3 times of average acquisition. The specific parameters of the three sequences are shown in Table 1.

[0143] Table 1 List of Scanning Parameters

[0144] Parameter MS-DIR MS-RDIR SS-RDIR <![CDATA[Field of View (FOV), mm 2 > 360×260 / 320 360×260 / 320 360×260 Voxel size (mm×mm) 1.4×1.76 1.4×1.76 1.4×1.76 Slice thickness (mm) 6.0 6.0 6.0 Sampling bandwidth (Hz / pixel) 400 400 700 Echo train length (ETL) 25 25 46-52 Echo spacing (ms) 5.62 5.62 3.84 Echo time (TE, ms) 67.4 67.4 84.5-96 Flip angle (°) 120° 120° 120° Number of averages (NEX) 3 3 3 Combined acceleration factor 1.8 1.8 3 Number of excitations (shots) 4-5 4-5 1 Number of acquired slices 5-11 5-11 5-11 Scan time per slice (heartbeats) 24 - 30 cardiac cycles 24 - 30 cardiac cycles 6 cardiac cycles Total scan time (heartbeats) 120 - 330 cardiac cycles 120 - 330 cardiac cycles 30 - 66 cardiac cycles

[0145] Note: "Cardiac cycle" can be converted to the corresponding number of seconds or minutes according to the actual heart rate; "Combined acceleration factor" refers to the overall acceleration multiple obtained by comprehensively combining techniques such as parallel imaging and compressed sensing.

[0146] 4. Image Analysis

[0147] To compare the performance of MS-DIR, MS-RDIR, and SS-RDIR, we evaluated multiple quantitative metrics, including myocardial signal-to-noise ratio (SNR), contrast-to-noise ratio between myocardium and blood (CNR), and contrast ratio between myocardium and blood (CR). Two readers (Y.E. and Z.C., with 3 years and 4 years of cardiac MRI experience respectively) carefully delineated the regions of interest (ROIs) of the left ventricular wall and blood pool. SNR was calculated based on the mean signal of the left ventricular wall and the standard deviation of the blood pool ROI, and the blood pool ROI was placed in the middle of the blood pool to exclude the bright edge region caused by slow flow. The formula for calculating the CNR between myocardium and blood is: (mean signal of the left ventricular wall - mean signal of the blood pool) / standard deviation of the blood pool. CR was calculated by the ratio of the mean signal of the left ventricular wall to the mean signal of the blood pool. The quantitative analysis results of the two readers were averaged to generate the final result.

[0148] We also performed a qualitative score on the performance of each imaging method. Two magnetic resonance physicists and radiologists (C.H. and X.H.) with more than 10 years of cardiovascular MRI experience independently performed a blind evaluation on the patient images. The evaluation criteria included myocardial visibility, ghost artifacts, and overall image quality, all scored on a 5-point scale (1: Unable to diagnose; 2: Poor; 3: Fair; 4: Good; 5: Excellent). The scores of the two readers were averaged to obtain the final score.

[0149] 5. Statistical Analysis

[0150] Statistical analysis was completed using Matlab (version 2022b) and IBM SPSS Statistics (version 26.0). Analysis of variance (ANOVA) and post hoc tests were used to evaluate the differences in quantitative parameters among the three different imaging methods. The Kruskal-Wallis test was used to evaluate the differences in qualitative comparisons among the three methods. When performing multiple comparisons, the Bonferroni correction was used to adjust the P-values. Inter-observer agreement was evaluated by the intraclass correlation coefficient (ICC). A P < 0.05 was considered statistically significant.

[0151] II. Results

[0152] All 47 patients successfully completed the examinations of the three sequences. The specific characteristics of these patients are shown in Table 2.

[0153] Table 2 Patient characteristics

[0154]

[0155] Figure 2 The quantitative evaluation results of MS-DIR, MS-RDIR, and SS-RDIR in 47 patients are presented. The total acquisition time of SS-RDIR (62.1 ± 24.0 seconds) was significantly shorter than that of MS-DIR (216.6 ± 74.2 seconds; P < 0.01) and MS-RDIR (229.8 ± 79.2 seconds; P < 0.01). There were no significant differences among MS-DIR, MS-RDIR, and SS-RDIR in terms of signal-to-noise ratio (SNR, P = 0.07) and contrast-to-noise ratio (CNR, P = 0.09); however, the contrast ratio (CR) of MS-RDIR was 11.3 ± 3.8, lower than that of MS-DIR (13.7 ± 5.1, P = 0.07), and significantly lower than that of SS-RDIR (13.9 ± 5.9, P = 0.04). There was no significant difference in CR between MS-DIR and SS-RDIR.

[0156] Figure 3The qualitative comparison results of MS-DIR, MS-RDIR, and SS-RDIR images of 47 patients are shown, with scores ranging from 1 (worst) to 5 (best). In terms of myocardial visibility, MS-RDIR had the highest score, but there was no significant difference among MS-DIR, MS-RDIR, and SS-RDIR (3.98 ± 0.42 vs 4.38 ± 0.35 vs 3.96 ± 0.44, P = 0.053). In terms of ghost artifacts, there was a significant difference among the three techniques (P < 0.01). Pairwise analysis showed that the score of SS-RDIR was significantly higher than that of MS-DIR (4.94 ± 0.11 vs 3.77 ± 0.29, P < 0.01) and MS-RDIR (3.78 ± 0.34, P < 0.01), while there was no significant difference in ghost artifacts between MS-DIR and MS-RDIR. In terms of overall quality, there was a significant difference among the three techniques (P = 0.03). The score of MS-DIR was similar to that of MS-RDIR (3.61 ± 0.40 vs 3.88 ± 0.41, P = 0.25) and was significantly lower than that of SS-RDIR (3.96 ± 0.52, P = 0.02). In addition, there was no significant difference in the overall quality score between MS-RDIR and SS-RDIR.

[0157] Table 3 lists the intraclass correlation coefficients (ICCs) for different comparisons. Both qualitative and quantitative comparisons showed good agreement between the two readers (ICC 0.00–0.20, poor agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, good agreement; greater than 0.81, excellent agreement).

[0158] Table 3. ICC Agreement Scores

[0159]

[0160] Note: The intraclass correlation coefficient (ICC) reflects the degree of agreement in scoring / measurement among different observers or different measurements; SNR: signal-to-noise ratio; CNR: contrast-to-noise ratio; CR: contrast ratio.

[0161] Figure 4 MS-DIR and MS-RDIR images of two patients are shown. The MS-DIR images were significantly disturbed by signal loss artifacts (yellow arrows), especially in the second and fourth frames of the second patient; these signal loss artifacts were significantly reduced in MS-RDIR.

[0162] Figure 5MS-RDIR and SS-RDIR images of two patients are shown. Myocardial motion-related ghost artifacts (green arrows) were observed in MS-RDIR; these artifacts were significantly suppressed in SS-RDIR.

[0163] Figure 6 Two slices of MS-DIR, MS-RDIR, SS-RDIR, LGE images, and T2 maps of a 14-year-old girl with myocarditis are shown. Due to artifact interference (green arrows), SS-RDIR showed more obvious local T2 elevation (blue arrows) than MS-DIR and MS-RDIR, which was consistent with the results of LGE images and T2 maps.

[0164] III. Discussion

[0165] The present invention is based on the combination of RDIR black blood preparation and deep learning-assisted reconstruction for improving free-breathing DB-FSE in pediatric imaging. Compared with conventional DIR multiple-excitation DB-FSE, the combination of RDIR and single-excitation acquisition with ACS reconstruction significantly reduces signal loss and motion ghosts, thus improving the overall image quality. The study shows that the SS-RDIR method has the potential to achieve high-resolution and motion-robust edema assessment under free breathing.

[0166] Children usually have a higher heart rate than adults and are to some extent unable to control their breathing. Therefore, in practice, free breathing combined with multiple averaging is often used in pediatric imaging. However, although DB-FSE is widely used, it is very sensitive to cardiac and respiratory motion, resulting in severe signal loss artifacts and ghost artifacts under free breathing. From the results, our method has the following advantages: First, compared with DIR, the RDIR technique can reduce the risk of signal loss in pediatric imaging. This can be seen from the comparison of myocardial visibility scores between MS-DIR and MS-RDIR. Reducing signal loss is crucial for the diagnosis of edema because local signal loss may make normal intensity regions appear over-enhanced, which may lead to false positive results if judged only based on image features. Second, single-excitation acquisition can significantly reduce ghost artifacts caused by inter-excitation respiratory motion. This can be seen from the substantial improvement in ghost scores of SS-RDIR compared with the other two methods. The reduction of signal loss and ghost artifacts also makes the overall image quality of the proposed method better than that of MS-DIR. Third, compared with MS-DIR and MS-RDIR (3 - 5 minutes), the scanning time of SS-RDIR is the shortest (30 - 60 seconds). The substantial reduction in scanning time significantly improves patient comfort and imaging efficiency.

[0167] In summary, RDIR reduces signal loss artifacts in free-breathing DB-FSE imaging, while deep learning-based accelerated single-shot acquisition reduces ghost artifacts. The proposed framework combining RDIR and single-shot accelerated acquisition shows an overall improvement in image quality even under conditions of high heart rate and free breathing. Further technological development is needed to shorten the echo train length. The application of these techniques is expected to enable artifact-free black-blood imaging in free-breathing pediatric patients.

[0168] The foregoing is a description of embodiments of the present invention. By the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for performing black-blood T2-weighted fast spin echo imaging under free breathing conditions, the method comprising at least the following steps: 1) Pulse sequence design: A reverse dual inversion recovery (RDIR) preparation pulse is set before each imaging cardiac cycle to achieve inversion of blood signals and minimize signal loss caused by slice misregistration; 2) Single-shot acquisition: After the RDIR preparation, at least one frame of k-space data covering the left ventricle is obtained using a single-shot fast spin echo (SS-RDIR) sequence; 3) Data undersampling: The k-space data during the single-shot acquisition process is undersampled in a compressed sensing (CS) manner to shorten the scanning time; 4) AI-assisted reconstruction: Based on the artificial intelligence-assisted compressed sensing (ACS) technology, an image reconstruction process combining deep learning and iterative reconstruction is performed on the undersampled k-space data to obtain a high-resolution T2-weighted black-blood image with reduced motion artifacts.

2. The method according to claim 1, wherein The reverse dual inversion recovery RDIR preparation pulse includes: a) Applying a slice-selective inversion pulse in the mid-diastole of the previous cardiac cycle; b) Applying a non-selective inversion pulse in the next cardiac cycle after the slice-selective inversion pulse; c) Matching the time interval between the non-selective inversion pulse and the fast spin echo (FSE) echo train to minimize myocardial signal loss caused by slice misregistration.

3. The method according to claim 1, characterized in that, In the single-shot fast spin echo (SS-RDIR) acquisition, the number of k-space lines acquired for each slice is between 46 and 52, so as to complete the sampling of the entire image within one cardiac cycle.

4. The method according to claim 1, characterized in that, The reconstruction steps of the artificial intelligence-assisted compressed sensing (ACS) technology include: a) Performing preliminary artifact removal and noise reduction on the undersampled k-space data through a trained deep neural network to obtain a preliminary reconstructed image; b) Using this preliminary reconstructed image as a prior constraint and combining it with a conventional compressed sensing iterative algorithm for further iterative optimization; c) Outputting the final high-fidelity T2-weighted black-blood image; Further preferably, the deep neural network used in the artificial intelligence-assisted reconstruction process is a convolutional neural network (CNN) or a hybrid network containing a Transformer structure, which is used to reduce artifacts and retain image details at a high acceleration factor.

5. The method according to claim 1, wherein The method further includes: a) Dynamically adjusting the trigger time between the RDIR preparation pulse and the single-shot FSE echo train according to the real-time heart rate of the child; b) Combining at least one signal averaging or registration to enhance the overall visibility and contrast of the myocardium; And / or, the method further includes: a) Using electrocardiogram- or pulse wave-based triggering to determine the acquisition window; b) Controlling the entire scanning time to be completed within 30 to 60 seconds to minimize artifacts caused by respiratory inconsistency; And / or, the method further includes performing motion correction on the data frames obtained by multiple averaging before image reconstruction to align the myocardial position before averaging, thereby further reducing motion artifacts and improving image quality.

6. A magnetic resonance imaging system for implementing the imaging method according to any one of claims 1 to 5, the system comprising: 1) A magnet and gradient coil assembly for generating a uniform static magnetic field and performing pulsed gradients at a magnetic field strength of 3T; 2) A radio frequency transmitting and receiving unit including at least one multi-channel radio frequency coil for transmitting RDIR preparation pulses and receiving signals; 3) A sequence control module programmed to execute pulse sequences including reverse dual inversion recovery (RDIR) and single-shot fast spin echo (SS-RDIR); 4) A data acquisition and storage module for acquiring and storing undersampled k-space data; 5) An artificial intelligence-assisted compressed sensing reconstruction module configured to reconstruct undersampled k-space data using a pre-trained deep learning model and output a final T2-weighted black blood image; 6) A display and evaluation module for displaying the reconstructed black blood image and providing an interface for subsequent diagnostic analysis.

7. The magnetic resonance imaging system according to claim 6, characterized in that, The artificial intelligence-assisted compressed sensing reconstruction module includes: a) A neural network processing unit pre-loaded with a deep learning model for generating a preliminary reconstructed image and correcting motion artifacts; b) An iterative optimization unit that combines the preliminary reconstruction result with the sparsity constraint of compressed sensing and performs multiple iterative updates to obtain a final high-resolution myocardial T2-weighted black blood image; c) A parameter scheduling unit that dynamically adjusts the trigger timing and undersampling trajectory of RDIR preparation and single-shot acquisition according to the heart rate and respiratory pattern of the child.

8. Use of the method according to any one of claims 1-5 in the design and evaluation of software for myocardial edema.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instruction, when executed by a processor, implements the method according to any one of claims 1-5.

10. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instruction, when executed by a processor, implements the method according to any one of claims 1-5.

Images