A method for generating magnetic resonance proton density fat fraction and R2* quantitative maps

Abstract

The present invention relates to the fields of image processing and medicine, and discloses a method for generating magnetic resonance proton density fat fraction and R2* quantitative maps, comprising the following steps: acquiring mDixon sequence image data, DWI sequence image data, and corresponding qDixon sequence map data, and integrating the image data to form a data set; preprocessing the acquired image data; constructing a deep learning model for generating magnetic resonance proton density fat fraction and R2* quantitative maps from the mDixon sequence images and DWI sequence images; using the preprocessed image data to train the established deep learning model to obtain a trained deep learning model; acquiring the mDixon sequence image data and DWI sequence image data of a subject to be tested, and inputting them into the trained deep learning model to generate proton density fat fraction and R2* quantitative maps. The present invention solves the problem of large PDFF deviation in patients with iron overload.

Description

Technical Field

[0001] The present invention relates to the fields of image processing and medicine, and particularly to a method for generating magnetic resonance proton density fat fraction and R2 star quantitative maps. Background Art

[0002] Quantitative Dixon imaging (qDixon) is an important technique for accurately measuring the proton density fat fraction (PDFF) and R2 star (R2*) of the liver. These two parameters are crucial for the diagnosis of hepatic steatosis (i.e., liver fat deposition). By utilizing multiple echo signals, qDixon can precisely separate the fat and water components, thereby providing a non-invasive quantitative assessment of the liver fat content. Due to its high precision and non-invasive nature, qDixon is applied to the clinical diagnosis of liver diseases, especially in the evaluation of diseases such as fatty liver, non-alcoholic fatty liver disease (NAFLD), and cirrhosis.

[0003] However, a major drawback of the qDixon technique is its long scan time, which leads to a long examination time for patients. This not only increases the examination cost and affects the patient experience but also increases the chance of voluntary or involuntary movement during the examination due to the decreased patient tolerance, thus affecting the image quality. In addition, the qDixon method is usually only achievable on high-end MRI devices, which also limits the widespread application of this method in routine clinical practice.

[0004] In contrast, mDixon is a more simplified technique. Compared with qDixon, mDixon requires fewer echo signals, thus significantly shortening the acquisition time and reducing the cost, making this technique more applicable in routine clinical examinations. The advantage of the mDixon method is that it not only reduces the equipment burden but also improves the patient comfort. Especially in cases where rapid examination is required, mDixon can provide an effective alternative.

[0005] However, since mDixon uses two-point echo signals and lacks correction for T2 decay, it may produce deviations in the presence of tissue changes such as iron deposition. Iron deposition, especially in patients with iron overload, is one of the important factors affecting the accuracy of mDixon. Iron deposition can cause a decrease in the T2 value of the liver, thereby affecting the accuracy of water-fat separation and further leading to inaccurate estimation of the PDFF value. Therefore, the application effect of mDixon in patients with iron overload is far less ideal than in the normal population, and obvious deviations may occur in the PDFF estimation. Therefore, there is an urgent need for a PDFF and R2 star quantitative method that can both improve efficiency and maintain high accuracy. Summary of the Invention

[0006] The main object of the present invention is to solve the technical problem that the application effect of mDixon in patients with iron overload is far less ideal than that in the normal population. A method for generating magnetic resonance proton density fat fraction and R2* quantitative maps includes the following steps:

[0007] S1. Obtain mDixon sequence image data, DWI sequence image data, and corresponding qDixon sequence map data, and integrate the image data to form a data set;

[0008] S2. Preprocess the acquired image data, and the data preprocessing includes image normalization, image denoising, and data augmentation;

[0009] S3. Construct a deep learning model for generating magnetic resonance proton density fat fraction and R2* quantitative maps from mDixon sequence images and DWI sequence images;

[0010] S4. Use the image data preprocessed in S2 to train the deep learning model established in S3 to obtain a trained deep learning model;

[0011] S5. Obtain mDixon sequence image data and DWI sequence image data of the subject to be tested, input them into the trained deep learning model, and generate proton density fat fraction and R2* quantitative maps.

[0012] As a preferred technical solution, in step S1, the mDixon sequence image is used to provide separation information of fat and water components. Four images will be obtained through the mDixon sequence image, namely in-phase image (IP), out-of-phase image (OP), water and fat images; the qDixon sequence image data obtains accurate PDFF and R2* quantitative maps as the ground truth during model training and evaluation.

[0013] As a preferred technical solution, in step S2, the specific steps of the image normalization, image denoising, and data augmentation are as follows: uniformly adjust all images to the same resolution; then, apply the Z-Score normalization method to perform pixel value normalization on the images; automatically segment the images to remove background noise and only retain the target area for subsequent analysis; randomly flip, rotate, crop, adjust brightness, change contrast, and add random noise.

[0014] As a preferred technical solution, in step S3, the deep learning model for generating the magnetic resonance proton density fat fraction and R2* quantitative map is based on the U-Net convolutional neural network. The U-Net network structure includes multiple convolutional layers, pooling layers, upsampling layers, and skip connection layers, and the network adopts a multi-branch input and multi-channel output method. As a preferred technical solution, the U-Net model input includes multiple branches, which respectively receive the IP and OP images of the mDixon sequence images and the DWI sequence images. The input images are first subjected to feature extraction through a series of convolutional layers and pooling layers of each branch, and the features of each branch are fused in the middle layer of the network; in the upsampling stage, the multi-scale features of each branch are fused, and the final feature map will be sent to two different convolutional layers, and each layer corresponds to generating a PDFF and an R2* quantitative map.

[0015] As a preferred technical solution, in step S4, the training of the deep learning model for generating the magnetic resonance proton density fat fraction and R2* quantitative map includes the following steps: during the training process, the model is iteratively trained using the training data with labels; the labels include the PDFF and R2* quantitative maps from the qDixon technique, and the model gradually optimizes the network weights by minimizing the loss function between the predicted value and the true label; the cross-validation method is used to verify the generalization ability of the model.

[0016] The second aspect of the present invention provides a system for generating a magnetic resonance proton density fat fraction and R2* quantitative map, including:

[0017] An acquisition module, configured to acquire mDixon sequence image data, DWI sequence image data, and corresponding qDixon sequence map data, and integrate the image data to form a data set;

[0018] A preprocessing module, configured to preprocess the acquired image data, and the data preprocessing includes image normalization, image denoising, and data augmentation;

[0019] A deep learning model construction module, configured to construct a deep learning model for generating a magnetic resonance proton density fat fraction and R2* quantitative map from mDixon sequence images and DWI sequence images;

[0020] A training module, configured to use the image data preprocessed in S2 to train the deep learning model established in S3 to obtain a trained deep learning model;

[0021] A generation module, configured to acquire mDixon sequence image data and DWI sequence image data of a subject to be measured, input the trained deep learning model, and generate a proton density fat fraction and an R2* quantitative map.

[0022] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein instructions are stored in the memory, and the memory and the at least one processor are interconnected by a line; the at least one processor invokes the instructions in the memory to cause the electronic device to execute the method for generating the magnetic resonance proton density fat fraction and R2-star quantitative map as described above.

[0023] A fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored, and when it runs on a computer, it causes the computer to execute the method for generating the magnetic resonance proton density fat fraction and R2-star quantitative map as described above.

[0024] The present invention has the following beneficial effects:

[0025] By using conventional MRI sequences (mDixon, DWI) to generate PDFF and R2-star quantitative maps, the present invention reduces the scanning time by at least 30% compared with the traditional qDixon technique, and avoids the discomfort caused by the patient's long-term exposure to MRI scans.

[0026] By combining multi-modal data including diffusion weighted imaging (DWI) that provides iron deposition information with a deep learning model, the estimation accuracy of PDFF and R2-star parameters can be significantly improved. Compared with the mDixon technique and single-modal input data, the present invention solves the problem of large PDFF deviation in patients with iron overload, and experiments show that the PDFF values predicted by the present invention are highly consistent with the qDixon method.

[0027] The present invention is not only applicable to the estimation of liver fat fraction, but also can be extended to the quantitative imaging of other organs or tissues, and has broad clinical application prospects. By combining with other MRI sequences and quantitative analysis methods, the present invention can provide strong support for the early diagnosis of various diseases and promote the development of precision medicine. Description of the Drawings

[0028] Figure 1 It is the overall flowchart of the method for generating the magnetic resonance proton density fat fraction and R2-star quantitative map based on deep learning of the present invention;

[0029] Figure 2 It is the network architecture diagram of the present invention;

[0030] Figure 3 It is the comparison diagram of the PDFF and R2-star quantitative maps generated by the model of the present invention and the quantitative maps obtained by the qDixon technique;

[0031] Figure 4are the true PDFF data distribution map and the true R2-star data distribution map. A is the true PDFF data distribution in the training set and the test set, and B is the true R2-star data distribution map in the training set and the test set;

[0032] Figure 5 is the regression analysis graph of MBCP-Net PDFF (No DWI) and MBCP-Net PDFF (With DWI) in the group without moderate iron deposition, where qDixon PDFF is used as the reference standard;

[0033] Figure 6 is the same regression analysis graph in the group with moderate to severe iron deposition;

[0034] Figure 7 is the Bland-Altman analysis graph of MBCP-Net PDFF (No DWI) and qDixon PDFF in the two iron deposition groups;

[0035] Figure 8 is the Bland-Altman analysis graph of MBCP-Net PDFF (With DWI) and qDixon PDFF in the two iron deposition groups;

[0036] Figure 9 shows the distribution graphs of qDixon PDFF minus MBCP-Net PDFF (With DWI) and MBCP-Net PDFF (No DWI) respectively in the group with moderate to severe iron deposition. Detailed implementation manners

[0037] The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0038] Proton density fat fraction (PDFF)

[0039] PDFF represents the percentage of fat proton signal in the total proton signal (fat + water) per unit volume of tissue, which is a dimensionless quantitative index (range: 0% - 100%). Through multi - echo gradient echo (multi - echo GRE) or chemical shift encoding (IDEAL / Dixon) sequences, signals at different echo times (TE) or different chemical shifts (frequency difference between fat and water) are acquired. Using the law of signal variation with TE in the fat - water mixture model and combining complex algorithms (such as multi - peak fat spectrum model, R2 star correction), the signal contributions of fat and water are separated, and PDFF is calculated.

[0040] R2 star

[0041] R2* is the transverse relaxation rate, with the unit of second⁻¹ (s⁻¹), which reflects the rate of attenuation of the transverse magnetization signal in the tissue and is directly related to magnetic field inhomogeneity (such as iron deposition, hemorrhage, calcification, etc.).

[0042] qDixon sequence image

[0043] qDixon (Quantitative Dixon) is an improved magnetic resonance imaging technique. Based on the physical model of the traditional Dixon method, through multi - echo gradient echo (multi - echo GRE) sequences and complex post - processing algorithms, proton density fat fraction (PDFF) and R2 (transverse relaxation rate) maps are quantitatively generated.

[0044] mDixon sequence image

[0045] mDixon is an optimized version of the traditional Dixon technique. Through improved scanning sequences and post - processing algorithms, more robust fat - water separation is achieved, and multiple contrast images are generated.

[0046] DWI sequence image

[0047] DWI (Diffusion - Weighted Imaging) is a magnetic resonance imaging technique based on the diffusion motion of water molecules. By applying diffusion - sensitive gradient pulses, the restriction of tissue microstructure (such as cell density, membrane integrity) on the diffusion of water molecules is detected.

[0048] A method for generating magnetic resonance proton density fat fraction and R2 star quantitative maps based on deep learning includes the following steps:

[0049] S1: Use a magnetic resonance device to collect the mDixon sequence, diffusion - weighted imaging DWI sequence, and corresponding qDixon sequence image data of the subject to be measured, and integrate the data to form a dataset;

[0050] S2: Preprocess the collected MRI images. The data preprocessing includes image normalization, image denoising, and data augmentation;

[0051] S3: Construct a deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative maps from mDixon sequence images and DWI sequence images;

[0052] S4: Use the preprocessed magnetic resonance imaging data in S2 to train the network model established in S3 to obtain the trained model weights.

[0053] S5: Use the mDixon sequence and diffusion weighted imaging (DWI) sequence image data collected by multiple magnetic resonance devices, input them into the deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative maps, and use the weights obtained in S4 to generate the corresponding PDFF and R2 star quantitative maps.

[0054] Further, in step S1, the mDixon sequence is used to provide the separation information of fat and water components. Four types of images will be obtained through the mDixon sequence, namely in-phase image (IP), opposed-phase image (OP), water and fat images; while the DWI sequence is used to provide the signal attenuation and tissue change information caused by iron deposition to improve the sensitivity to iron deposition; the qDixon sequence obtains accurate PDFF and R2 star quantitative maps as the ground truth during model training and evaluation.

[0055] Further, in step S2, the specific steps of image normalization, image denoising, and data augmentation are as follows:

[0056] Uniformly adjust all images to the same resolution to ensure a consistent spatial scale between different images. Then, apply the Z-Score normalization method to perform pixel value normalization on the images, thereby eliminating the biases and inconsistencies caused by different devices or scanning conditions, and ensuring the unity of the input data and the stability of model training;

[0057] Automatically segment the images, remove background noise, and only retain the target area for subsequent analysis to ensure the accuracy and focus of model processing;

[0058] Random flipping, rotation, cropping, brightness adjustment, contrast change, and random noise addition, etc. These augmentation means can simulate the changes that may occur under different scanning conditions, and effectively prevent overfitting, improving the performance and adaptability of the model in practical applications.

[0059] Further, in step S3, the deep learning model for generating magnetic resonance proton density fat fraction and R2-star quantitative maps is based on the U-Net convolutional neural network. The U-Net network structure includes multiple convolutional layers, pooling layers, upsampling layers, and skip connection layers. The network adopts a multi-branch input and multi-channel output method to simultaneously process mDixon sequence and DWI sequence image data. The network effectively captures the detailed information in the image through the feature extraction module, and uses the skip connection layer to fuse features of different scales, thereby maintaining the spatial information of the image and enhancing the feature learning ability.

[0060] Among them, the input of the U-Net model includes multiple branches, which respectively receive the IP and OP images from the mDixon sequence and the DWI image. The input images are first subjected to feature extraction through a series of convolutional layers and pooling layers of each branch, and the features of each branch are fused in the middle layer of the network. In the upsampling stage, the multi-scale features of each branch are fused to make full use of the multi-modal information. The final feature map will be sent to two different convolutional layers, and each layer corresponds to generating PDFF and R2-star quantitative maps.

[0061] Further, in step S4, the training of the deep learning model for generating magnetic resonance proton density fat fraction and R2-star quantitative maps includes the following steps:

[0062] During the training process, the model is iteratively trained using the training data with labels. The labels include PDFF and R2-star quantitative maps from the qDixon technique. The model gradually optimizes the network weights by minimizing the loss function between the predicted value and the true label;

[0063] The cross-validation method is used to verify the generalization ability of the model to ensure that the weights obtained by training can achieve consistent and accurate results on different data sets.

[0064] For easy understanding, the specific process of the embodiment of the present invention is described below. Please refer to Figure 1 The first embodiment of the method for generating magnetic resonance proton density fat fraction and R2-star quantitative maps in the embodiment of the present invention includes:

[0065] As Figure 1 shown, this embodiment provides a method for generating magnetic resonance proton density fat fraction and R2-star quantitative maps based on deep learning, including the following steps:

[0066] S1: Use a magnetic resonance device to collect the mDixon sequence and diffusion weighted imaging (DWI) sequence of the subject to be measured, as well as the corresponding qDixon sequence abdominal image data, and integrate the data to form a data set.

[0067] Specifically, in this embodiment, the mDixon sequence, diffusion-weighted imaging DWI (b = 800) sequence, and corresponding qDixon sequence abdominal image data of 301 subjects to be measured were collected by United Imaging magnetic resonance equipment as the training set, and the mDixon sequence, diffusion-weighted imaging DWI (b = 800) sequence, and corresponding qDixon sequence abdominal image data of 135 subjects to be measured were collected by Siemens, Philips, and some United Imaging equipment as the test set.

[0068] S2: Perform data preprocessing on the collected MRI images. The data preprocessing includes image normalization, image denoising, and data augmentation.

[0069] Specifically, in step S2, all images are uniformly adjusted to the same resolution to ensure consistent spatial scales between different images. Then, the Z-Score normalization method is applied to perform pixel value normalization on the images. The nnUNet model is used to accurately segment the abdominal images. By automatically segmenting the abdominal images, the background noise outside the abdomen is effectively removed, and only the target abdominal region is retained. Data augmentation operations such as random flipping, rotation, cropping, brightness adjustment, contrast change, and random noise addition are performed on the data.

[0070] S3: Construct a deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative maps.

[0071] Specifically, in step S3, the network structure is as Figure 2 shown. The deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative maps is based on the 2D U-Net convolutional neural network. The U-Net network structure includes multiple convolutional layers, pooling layers, upsampling layers, and skip connection layers. The network adopts a multi-branch input and multi-channel output method to simultaneously process the mDixon sequence and DWI sequence image data. The network effectively captures the detailed information in the images through the feature extraction module and uses the skip connection layers to fuse features of different scales, thereby maintaining the spatial information of the images and enhancing the feature learning ability.

[0072] S4: In the network model established in S3, train the preprocessed magnetic resonance image data in S2 to obtain the trained weights.

[0073] Specifically, in step S4, during the training process, the model is iteratively trained using the training data with labels. The labels include PDFF and R2 star quantitative maps from the qDixon technique. The model gradually optimizes the network weights by minimizing the loss function between the predicted values and the true labels. The loss function is the L1 norm loss. The 5-fold cross-validation method is used to verify the generalization ability of the model to ensure that the trained weights can achieve consistent and accurate results on different data sets.

[0074] S5: Use the abdominal imaging data of the mDixon sequence and DWI (b = 800) sequence in the test set, input the abdominal imaging data into the deep learning model for generating magnetic resonance proton density fat fraction and R2-star quantitative maps, and use the weights obtained in S4 to generate the corresponding PDFF and R2-star quantitative maps.

[0075] In specific implementation:

[0076] We obtained approval from the ethics committee of Renji Hospital and retrospectively collected 436 cases of abdominal MR imaging cases. T1w Dixon in-phase, opposed-phase, water, and fat four-phase maps, DWI, PDFF, and R2-star sequences were collected for these cases, and the case data were from MR devices of three manufacturers, Siemens, Philips, and UIH.

[0077] The in-phase, opposed-phase, water, and fat four-phase maps and DWI are conventional sequences for liver MR imaging scans.

[0078] The value of PDFF reflects the fat content, while the R2-star value reflects the degree of iron deposition. Steatosis and iron deposition are manifestations of common liver metabolic diseases on imaging. Due to the characteristics of the MR sequence technical principle, these two are usually inseparable in the form of a "twin combination", but special sequences are required to collect the PDFF and R2-star values.

[0079] Adopt a multi-branch conformal prediction network (MBCP-Net)

[0080] The preprocessed in-phase (IP), opposed-phase (OP), and diffusion-weighted imaging (DWI) data are input into the multi-branch conformal prediction network (MBCP-Net). This network can synchronously generate quantitative proton density fat fraction (PDFF) and transverse relaxation rate (R2-star) maps, and output the corresponding upper bound, lower bound, and uncertainty maps of the prediction values.

[0081] This network is constructed based on an improved 2D U-Net framework and has the following core features:

[0082] Multi-branch feature extraction: The IP, OP, and DWI three-modal data are processed through independent branches respectively. Each branch contains cascaded convolutional layers and pooling layers to complete primary feature extraction.

[0083] Cross-modal feature fusion: Deep fusion of the three-branch features is achieved in the middle layer of the network, and multi-modal information is fully integrated through a cross-connection architecture.

[0084] Multi-scale feature enhancement: The skip connection mechanism is adopted in the upsampling stage to dynamically fuse the multi-scale features of different branches and effectively utilize the complementary information of multi-modal data.

[0085] Introduce a conformal prediction mechanism in the output stage:

[0086] Configure three independent convolutional layers for the PDFF and R2-star mapping diagrams respectively to generate predicted values, upper bounds, and lower bounds. The predicted values are strictly constrained within the upper and lower bound intervals, and the interval span is visually presented through the uncertainty map. The conformal quantile regression algorithm is used to dynamically adjust the confidence interval range to ensure statistical reliability.

[0087] The network training adopts a dual-loss collaborative optimization strategy: (1) Quantile regression loss: precisely constrain the confidence interval boundaries of the predicted values. (2) L1 loss: used to generate accurate pixel values.

[0088] Among them, represents the quantile to be optimized, represents the quantile estimator. The upper and lower bounds can be estimated through the conditional quantile. m is the weight coefficient of the PDFF part, and n is the weight coefficient of the R2-star part.

[0089] Experiment

[0090] For the entire dataset, data from 301 patients from UIH were used in the training, validation, and calibration stages. The remaining data were used as the test set, including 54 UIH subjects, 61 subjects from Siemens Healthcare (SIE), and 20 subjects from Philips Healthcare (PHI).

[0091] In the training stage (N = 990 slices) and the validation stage (N = 265 slices), 5 slices were randomly selected from each case as inputs for the MBCP-Net. The model used the Adam optimizer with an initial learning rate of 0.001, trained for 500 epochs, and the batch size was 2. To make the learning process more stable, a cosine learning rate scheduler was used for dynamic adjustment. In the loss function, the weights of m = 0.7 and n = 0.3 balanced the error parts of the PDFF and R2-star. To estimate the 90% uncertainty interval for each pixel point, the model was trained to predict the 95% and 5% quantiles, where α = 10%. During the validation process, the peak signal-to-noise ratio (PSNR) metric was used to monitor the model performance, and the results of the best-performing epoch were saved to achieve optimal quantitative prediction.

[0092] In the calibration stage, the model was fine-tuned through a calibration set of 250 slices to ensure that the predicted uncertainty interval could accurately cover the pixel points outside the expected quantile range. The specific operations included adjusting the model prediction results so that the pixel points outside the prediction boundaries were properly included within the uncertainty interval.

[0093] In addition, to evaluate the role of DWI, we removed the DWI input branch in MBCP-Net while keeping other conditions unchanged for a comparative experiment. At this time, the network only uses IP and OP images as inputs to facilitate our evaluation of the contribution of multimodal information in DWI images to the model performance.

[0094] Automated reliable parameter extraction

[0095] To evaluate the quality of the generated quantitative maps, we developed an automatic parameter extraction method to replace the traditional manual region extraction work and utilized the uncertainty quantification function of MBCP-Net to provide statistical confidence for the extracted parameters. This method abandons the cumbersome operation of manually annotating regions of interest in the traditional method, and through the uncertainty quantification results provided by MBCP-Net, it can provide a reliability assessment for the extracted parameters.

[0096] Specifically, we first use TotalSegmentator to segment the liver region to obtain a liver region mask. Then, we normalize the uncertainty map output by the model to map it to the range of [0,1]. By subtracting the normalized uncertainty value from 1, we obtain a confidence map, thus converting uncertainty into a confidence metric. For each case, we select the regions with a confidence level exceeding 90% to form a confidence mask. Finally, the parameter extraction region (i.e., the reliable liver region) is defined as the intersection of the liver mask and the confidence mask, ensuring that we only consider the highly reliable regions in the liver. Within this region, we extract the average values of the PDFF and R2* maps generated by the model and compare them with the standard reference values of the qDixon method. Through this automated processing flow, we can not only obtain evaluation results with high statistical reliability, but also ensure the good repeatability of the entire process. More importantly, without any manual operations, the research efficiency is greatly improved.

[0097] Statistical analysis

[0098] To evaluate the performance of the proposed method, we mainly analyze two groups: (1) the PDFF estimation values generated by the MBCP-Net model by inputting IP, OP, and DWI images (abbreviated as MBCP-Net PDFF With DWI), and (2) the PDFF estimation values generated by the MBCP-Net model using only IP and OP images as inputs (abbreviated as MBCP-Net PDFF No DWI), with qDixon PDFF as the reference standard.

[0099] When making a quantitative comparison between MBCP-Net PDFF With DWI and MBCP-Net PDFF No DWI, we selected the intersection of two reliable liver regions to ensure that parameter extraction was performed for both methods within the same anatomical region. In addition, according to the reference R2* value, the patients were divided into two categories: no or mild iron deposition (≤62.5) and moderate to severe iron deposition (>62.5). Subsequently, we performed a quantitative comparison between MBCP-Net PDFF With DWI and MBCP-Net PDFF No DWI among different iron deposition groups respectively.

[0100] For the two groups of PDFF estimates, we used paired t-tests and Wilcoxon signed-rank tests to analyze their biases. In addition, we calculated the intraclass correlation coefficient (ICC) to evaluate the consistency between the PDFF estimates and qDixon PDFF. At the same time, linear regression analysis and Bland-Altman analysis were also used to evaluate the consistency between the PDFF estimates and qDixon PDFF.

[0101] All statistical analyses were set at a significance level of p < 0.05. The results were expressed as mean ± standard deviation. Statistical calculations were performed using the SciPy and statsmodels libraries in Python to ensure the reproducibility of the research results.

[0102] Results

[0103] In the mild iron deposition group, both MBCP-Net PDFF (With DWI) and MBCP-Net PDFF (No DWI) demonstrated high consistency and reliability with qDixon PDFF, with the DWI group showing slightly better results. Specifically, the intraclass correlation coefficients (ICC) of PDFF were 0.976 (With DWI) and 0.949 (No DWI) respectively. Further confirmation by linear regression analysis showed that With DWI performed better in maintaining the slope (1.026) and intercept (-0.200), while the No DWI group had a slope of 1.106 and an intercept of -1.201. The Bland-Altman analysis also indicated that the agreement limits of the With DWI group (-3.709 to 3.686) were more compact than those of the No DWI group (-5.188 to 4.792).

[0104] Table 1 shows the comparison between MBCP-Net PDFF With DWI and MBCP-Net PDFF No DWI (using qDixon PDFF as the standard). For the moderate-to-severe iron deposition group, MBCP-Net PDFF (With DWI) showed significant advantages. In this group, the ICC of With DWI reached 0.960, which was better than 0.949 of the No DWI group, reflecting higher consistency. Linear regression analysis revealed that the slope of the With DWI group (0.987, intercept 1.080) was closer to the ideal value, indicating a significant reduction in systematic bias. Bland-Altman analysis further highlighted the advantages of With DWI in the moderate-to-severe iron deposition group, with a bias of 0.940% and more compact agreement limits (-7.242% to 9.121%), while the bias of the No DWI group was 2.224% and the agreement limits were (-6.605% to 11.052%). Paired t-tests and Wilcoxon signed-rank tests showed significant differences between the No DWI and With DWI groups in the moderate-to-severe iron deposition group (P<0.0001).

[0105] Table 1

[0106]

[0107] The consistency of qDixon PDFF with the two methods (PDFF With DWI and PDFF No DWI) in the test sets of different iron deposition groups is shown in Figures 3 - 9 :[[]]END]] Figure 3 is a comparison chart of the PDFF and R2-star quantitative maps generated by the model of the present invention and the quantitative maps obtained by the qDixon technique; Figure 4 are the true PDFF data distribution map and the true R2-star data distribution map. A is the true PDFF data distribution in the training set and the test set, and B is the true R2-star data distribution map in the training set and the test set; Figure 5 is the regression analysis of MBCP-Net PDFF (No DWI) and MBCP-Net PDFF (With DWI) in the group without moderate iron deposition, where qDixon PDFF is used as the reference standard; Figure 6 is the same regression analysis in the moderate-to-severe iron deposition group; Figure 7 is the Bland-Altman analysis of MBCP-Net PDFF (No DWI) and qDixon PDFF in the two iron deposition groups; Figure 8 is the Bland-Altman analysis of MBCP-Net PDFF (With DWI) and qDixon PDFF in the two iron deposition groups;Figure 9 It shows the distribution of qDixon PDFF minus MBCP-Net PDFF (With DWI) and MBCP-Net PDFF (No DWI) respectively in the moderate to severe iron deposition group. Paired t-tests and Wilcoxon signed-rank tests showed significant differences between MBCP-Net PDFF (With DWI) and PDFF (No DWI).

[0108] The method for generating magnetic resonance proton density fat fraction and R2-star quantitative maps in the embodiments of the present invention has been described above. Next, the device for generating magnetic resonance proton density fat fraction and R2-star quantitative maps in the embodiments of the present invention will be described:

[0109] An acquisition module, configured to acquire mDixon sequence image data, DWI sequence image data, and corresponding qDixon sequence map data, and integrate the image data to form a data set;

[0110] A preprocessing module, configured to preprocess the acquired image data, and the data preprocessing includes image normalization, image denoising, and data augmentation;

[0111] A deep learning model construction module, configured to construct a deep learning model for generating magnetic resonance proton density fat fraction and R2-star quantitative maps from mDixon sequence images and DWI sequence images;

[0112] A training module, configured to use the image data preprocessed in S2 to train the deep learning model established in S3 to obtain a trained deep learning model;

[0113] A generation module, configured to acquire mDixon sequence image data and DWI sequence image data of a subject to be tested, input the trained deep learning model, and generate proton density fat fraction and R2-star quantitative maps.

[0114] The embodiments of the present invention also provide an electronic device, which may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) (for example, one or more processors) and a memory, and one or more storage media for storing application programs or data (for example, one or more mass storage devices). Among them, the memory and the storage medium may be transient storage or persistent storage. The program stored in the storage medium may include one or more modules, and each module may include a series of instruction operations on the electronic device. Further, the processor may be configured to communicate with the storage medium and execute a series of instruction operations in the storage medium on the electronic device.

[0115] The electronic device may further include one or more power supplies, one or more wired or wireless network interfaces, one or more input / output interfaces, and / or one or more operating systems, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art can understand that the structure of the electronic device in this embodiment does not constitute a limitation on the electronic device, and it may include more or fewer components, or combine certain components, or have different component arrangements.

[0116] The structure of an electronic device provided by an embodiment of the present invention may vary greatly due to different configurations or performances. It may include one or more processors (central processing units, CPUs) (for example, one or more processors) and a memory, and one or more storage media for storing applications or data (for example, one or more mass storage devices). Among them, the memory and the storage media may be transient storage or persistent storage. The program stored in the storage media may include one or more modules, and each module may include a series of instruction operations on the electronic device. Further, the processor may be configured to communicate with the storage media and execute a series of instruction operations in the storage media on the electronic device.

[0117] The electronic device may further include one or more power supplies, one or more wired or wireless network interfaces, one or more input / output interfaces, and / or one or more operating systems, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art can understand that the structure of the electronic device does not constitute a limitation on the electronic device, and it may include more or fewer components than the foregoing, or combine certain components, or have different component arrangements.

[0118] The present invention also provides a computer-readable storage medium. The computer-readable storage medium may be a non-volatile computer-readable storage medium, or may also be a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a computer, the computer is caused to execute the steps of the foregoing method.

[0119] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, or units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0120] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0121] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map, characterized in that: The following steps are involved: S1, acquiring mDixon sequence image data, DWI sequence image data and qDixon sequence image data, and integrating the image data to form a data set; S2, preprocessing the image data acquired in S1, wherein the preprocessing includes image standardization, image denoising and data amplification; S3, construct a deep learning model to generate magnetic resonance proton density fat fraction and R2 star quantitative map from mDixon sequence images and DWI sequence images; S4, using the image data pre-processed by S2, training the deep learning model established by S3 to obtain a trained deep learning model; S5, obtaining the mDixon sequence image data and DWI sequence image data of the subject, inputting the trained deep learning model, and generating the proton density fat fraction and R2 star quantitative map; In step S1, the mDixon sequence image is used to provide separation information of fat and water components. Four images are obtained through the mDixon sequence image, namely, the in-phase image IP, the anti-phase image OP, and the water and fat images; the qDixon sequence image data obtains accurate magnetic resonance proton density fat fraction and R2 star quantitative map as the benchmark truth value for model training and evaluation.

2. The method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map according to claim 1, characterized in that: In step S2, the specific steps of image standardization, image denoising and data amplification are: Resize all images to the same resolution; Next, the Z-Score normalization method is applied to normalize the pixel values ​​of the image; Automatically segment the image, remove background noise, and retain only the target area for subsequent analysis; Random flipping, rotation, cropping, brightness adjustment, contrast changes, and random noise addition.

3. The method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map according to claim 1, characterized in that: In step S3, the deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative image uses U-Net convolutional neural network as the basic architecture. The U-Net convolutional neural network structure includes multiple convolutional layers, pooling layers, upsampling layers and jump connection layers. The network adopts a multi-branch input and multi-channel output method.

4. The method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map according to claim 3, characterized in that: The U-Net convolutional neural network input includes multiple branches, which respectively receive IP and OP images and DWI sequence images from the mDixon sequence images. The input image is first subjected to feature extraction through a series of convolutional layers and pooling layers of each branch, and the features of each branch are fused in the middle layer of the network; in the upsampling stage, the multi-scale features of each branch are fused, and the final feature map is sent to two different convolutional layers, and each layer generates a magnetic resonance proton density fat fraction and an R2 star quantitative map.

5. The method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map according to claim 1, characterized in that: In step S4, the deep learning model training for generating magnetic resonance proton density fat fraction and R2 star quantitative image includes the following steps: During the training process, the model is iteratively trained using labeled training data; the labels include magnetic resonance proton density fat fraction and R2 star quantitative map from qDixon sequence image data. The model gradually optimizes the network weights by minimizing the loss function between the predicted value and the true label. The cross-validation method was used to verify the generalization ability of the model.

6. A system for generating a magnetic resonance proton density fat fraction and R2 star quantitative map, characterized in that: The system comprises: An acquisition module is used to acquire mDixon sequence image data, DWI sequence image data and qDixon sequence image data, and integrate the image data to form a data set; A preprocessing module, used for preprocessing the acquired image data, wherein the preprocessing includes image standardization, image denoising and data amplification; A deep learning model building module is used to build a deep learning model for generating magnetic resonance proton density fat fraction and R2 star quantitative images from mDixon sequence images and DWI sequence images; A training module is used to train the established deep learning model using the image data obtained by the preprocessing module to obtain a trained deep learning model; A generation module is used to obtain the mDixon sequence image data and DWI sequence image data of the subject, input the trained deep learning model, and generate the proton density fat fraction and R2 star quantitative map; The mDixon sequence image is used to provide separation information of fat and water components. Four types of images will be obtained through the mDixon sequence image, namely, the in-phase image IP, the anti-phase image OP, and the water and fat images; the qDixon sequence image data obtains accurate magnetic resonance proton density fat fraction and R2 star quantitative map as the benchmark truth value during model training and evaluation.

7. An electronic device, comprising a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the electronic device to execute each step of the method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map according to any one of claims 1 to 5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the various steps of the method for generating a magnetic resonance proton density fat fraction and R2 star quantitative map as described in any one of claims 1 to 5 are implemented.

Images