Liver fat automatic quantitative detection method, device and system based on multi-modal magnetic resonance imaging

Abstract

This invention relates to the field of medical image processing technology, and discloses an automated quantitative detection method, device, and system for liver fat based on multimodal magnetic resonance imaging (MRI). The method includes: acquiring multimodal MRI image data of the subject and classifying it to obtain T1-weighted sequence data and PDFF sequence data; inputting the T1-weighted sequence data into a liver segmentation model to obtain an initial liver mask adapted to the T1-weighted sequence; performing cross-modal mask alignment processing on the initial liver mask to obtain a target liver mask adapted to the PDFF sequence data; identifying and removing intrahepatic vascular regions, extracting effective liver parenchyma regions, and constructing a three-dimensional liver fat volume distribution after abnormal voxel removal, calculating the average liver fat fraction, and outputting the result. The automated quantitative detection of liver fat of this invention can achieve accurate segmentation and quantitative modeling of liver fat distribution from MRI, shortening MRI-PDFF processing time and improving repeatability and consistency, thereby enhancing the automation and accuracy of liver fat assessment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to liver fat quantitative analysis technology based on cross-modal MRI fusion and neural networks. Specifically, it relates to an automated quantitative detection system and method for liver fat based on multimodal magnetic resonance imaging. Background Technology

[0002] Metabolic dysfunction-associated fatty liver disease (MASLD) has become one of the most common chronic liver diseases worldwide, with a significantly increased prevalence in obese individuals or those with type 2 diabetes. Its disease spectrum can progress from simple fat deposition to steatohepatitis, a stage involving both inflammation and hepatocellular damage, further increasing the risk of liver fibrosis, cirrhosis, and hepatocellular carcinoma. Therefore, objective, quantitative, and reproducible assessment of hepatic fat deposition is a crucial technological foundation for early screening, treatment monitoring, and disease stratification management. Currently, commonly used clinical assessment methods include ultrasound, elastography, computed tomography (CT), magnetic resonance imaging (MRI), and liver biopsy. Among these, magnetic resonance proton density fat fraction (MRI-PDFF) is widely considered an important reference standard for assessing fat content due to its full liver coverage, non-invasiveness, and good quantitative consistency.

[0003] Despite the high accuracy of MRI-PDFF, its clinical application still relies on manual selection of regions of interest (ROIs) for measurement. This process is time-consuming, dependent on operator experience, and prone to sampling bias and inter-observer variability. Even with multi-slice and multi-point sampling to improve stability, the labor and time costs increase further, making it difficult to meet high-throughput clinical demands. Furthermore, while double-checking in existing methods can improve consistency, it significantly reduces process efficiency, limiting the widespread adoption of PDFF quantitative analysis in routine clinical settings. Overcoming these shortcomings to improve the efficiency and reliability of liver lipid quantification assessment has become an urgent need.

[0004] In recent years, artificial intelligence (AI) methods have been introduced into the field of medical image analysis for automated ROI selection and liver semantic segmentation, showing improvements in efficiency and consistency compared to manual processes. However, existing solutions still face several technical bottlenecks when dealing with quantitative analysis of MRI-PDFF: First, some existing analysis systems are designed at a semi-quantitative or local analysis level, failing to fully leverage the potential of PDFF as a quantitative biomarker; second, PDFF images primarily reflect fat signals and lack clear anatomical structural information, easily leading to blurred liver boundaries, vascular signal interference, and segmentation errors, affecting the accuracy of the analysis results. Furthermore, current AI-assisted MRI-PDFF analysis methods mainly rely on two-dimensional or global statistical indicators, failing to acquire three-dimensional volume data, thus making it difficult to accurately characterize the spatial heterogeneity of fat deposition and affecting the ability to distinguish between diffuse and focal lesions. Summary of the Invention

[0005] In view of the technical problems and defects of the existing technology, the purpose of this invention is to provide an automated quantitative detection system and method for liver fat based on multimodal magnetic resonance imaging. It utilizes T1-weighted dual-echo images with rich structural information to train a pseudo-3D segmentation network, combined with a cross-modal registration framework, to achieve accurate segmentation of the liver region in MRI-PDFF, automatically generate high-fidelity 3D liver fat volume maps, realize steatosis grading and high-accuracy voxel-level quantitative analysis, significantly shorten MRI-PDFF processing time and improve repeatability and consistency, and improve the automation and accuracy of liver fat assessment. It can be applied to efficacy follow-up and clinical auxiliary diagnosis.

[0006] According to a first aspect of the present invention, an automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging is proposed, comprising the following steps:

[0007] S1. Acquire MRI image data of the subject to be tested, and read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data;

[0008] S2. Input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask adapted to the T1-weighted sequence;

[0009] S3. Based on the image size and interlayer structure of the PDFF sequence, perform cross-modal mask alignment processing on the initial liver mask. Through interpolation resampling, scale recognition and displacement discrimination processing, convert it into a target liver mask that adapts to the PDFF sequence data.

[0010] S4. Identify the intrahepatic vascular region based on the target liver mask and extract the effective liver parenchyma region;

[0011] S5. Abnormal voxels within the extracted effective liver parenchyma region are removed, and a three-dimensional liver fat volume distribution is constructed based on the PDFF values ​​of the remaining effective voxels. The average liver fat fraction is then calculated.

[0012] S6. Output the three-dimensional liver fat volume distribution and the average liver fat fraction.

[0013] According to a second aspect of the present invention, an automated quantitative detection device for liver fat based on multimodal magnetic resonance imaging is provided, comprising:

[0014] The MRI image input module is used to acquire MRI image data of the subject under test, and to read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data;

[0015] The liver segmentation module is used to input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask that adapts to the T1-weighted sequence.

[0016] The alignment correction module is used to perform cross-modal mask alignment processing on the initial liver mask based on the image size and interlayer structure of the PDFF sequence. Through interpolation resampling, scale recognition and displacement discrimination processing, it is converted into a target liver mask that adapts to the PDFF sequence data.

[0017] An effective liver parenchyma region extraction module is used to identify intrahepatic vascular regions based on the target liver mask and extract the effective liver parenchyma region.

[0018] A quantitative detection module is used to remove abnormal voxels within the extracted effective liver parenchyma region, construct a three-dimensional liver fat volume distribution based on the PDFF values ​​of the remaining effective voxels, and calculate the average liver fat fraction; and

[0019] The output module is used to output the three-dimensional liver fat volume distribution and the average liver fat fraction.

[0020] According to a third aspect of the present invention, a computer system is provided, comprising:

[0021] One or more processors; and

[0022] Memory stores instructions that can be operated.

[0023] When the instructions are executed by one or more processors, they cause the aforementioned one or more processors to perform operations, including the process of executing the aforementioned automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging.

[0024] The automated quantitative detection system and method for liver fat based on multimodal magnetic resonance imaging (MRI) described in the above embodiments of the present invention integrates structural image information and achieves automated quantitative detection of liver fat through cross-modal registration and three-dimensional voxel-level analysis, completing precise segmentation and quantitative modeling of liver fat distribution from MRI. The method of the present invention can improve the consistency and spatial resolution of MRI-PDFF detection while reducing manual intervention, generating a three-dimensional fat volume representation with anatomical constraints, improving the accuracy and stability of fat quantitative assessment, and changing liver fat assessment from traditional local sampling to precise three-dimensional analysis of the whole liver. This provides stable, repeatable, and clinically interpretable quantitative evidence for early identification, efficacy evaluation, and long-term follow-up of MASLD. Simultaneously, it can be embedded in routine MRI examination procedures to achieve automated joint quantification of liver blood flow and metabolic parameters, thereby providing more accurate, objective, and repeatable diagnostic evidence for clinical practice.

[0025] Compared with existing technologies, the automatic quantitative detection method for liver fat based on multimodal magnetic resonance imaging of the present invention achieves automated and highly consistent quantitative assessment of the entire liver by fusing structure information-driven automatic segmentation, cross-modal precise alignment and voxel-level three-dimensional fat modeling. This reduces manual dependence and sampling bias, improves computational efficiency and repeatability, and enhances the ability to resolve fat spatial heterogeneity. It is suitable for clinical screening, efficacy monitoring and long-term follow-up analysis.

[0026] This invention proposes an automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging (MRI), and the detection system based on this method can be embedded into routine MRI examination procedures. This enables automated combined quantification of liver blood flow and metabolic parameters, providing objective and reproducible imaging biomarkers for clinical application. This facilitates clinical translation and direct application, aiding in liver disease staging, treatment monitoring, and complication risk prediction, and supporting individualized treatment decisions. By reducing reliance on manual labor and analysis time, it can improve the efficiency of radiology departments and promote standardized multi-center applications.

[0027] Meanwhile, this invention introduces a hemodynamic-metabolic multimodal joint analysis framework into the automated liver fat quantification system, organically integrating structural segmentation, cross-modal alignment, voxel-level fat quantification, and high-resolution hemodynamic assessment to achieve multidimensional, automated, and reproducible quantitative characterization of liver status. The method of this invention not only improves the accuracy and stability of single fat quantification assessment but also provides data conditions for quantitative analysis extending to hemodynamic-metabolic coupling mechanisms, offering a richer imaging indicator system for disease stratification, risk prediction, and efficacy monitoring, thereby enhancing clinical application value and technological scalability.

[0028] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0029] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0030] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings.

[0031] Figure 1 This is a flowchart of an automatic quantitative detection method for liver fat based on multimodal magnetic resonance imaging provided in an embodiment of the present invention. It illustrates the overall processing flow from multimodal image acquisition, automatic liver segmentation, cross-modal alignment to three-dimensional fat volume reconstruction and quantitative output.

[0032] Figure 2 For based on Figure 1 The Bland–Altman analysis plot of the average liver fat fraction calculated by the example method and the results of traditional artificial region of interest sampling is used to assess the systematic bias and range of consistency between the two methods in fat quantification results.

[0033] Figure 3 For based on Figure 1 The scatter plot comparing the average liver fat fraction calculated by the example method with the results of traditional manual region of interest sampling, combined with ICC(2,1) consistency analysis, is used to quantitatively evaluate the reliability between automated quantification methods and manual measurements.

[0034] Figure 4 This is a typical case in an embodiment of the present invention. The three-dimensional liver fat volume distribution map of patient GONYESEJ is used to demonstrate the visualization and spatial characteristics of voxel-level fat distribution in three-dimensional space.

[0035] Figure 5 This is a schematic diagram of the structure of the Pseudo-3D U-Net model according to an embodiment of the present invention. Detailed Implementation

[0036] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0037] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0038] {Example 1}

[0039] Combined with appendix Figure 1 As shown, the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging according to an embodiment of the present invention includes the following steps:

[0040] Step S1: Acquire multimodal MRI image data of the object to be tested, and read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data;

[0041] Step S2: Input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask adapted to the T1-weighted sequence;

[0042] Step S3: Based on the image size and interlayer structure of the PDFF sequence, perform cross-modal mask alignment processing on the initial liver mask. Through interpolation resampling, scale recognition and displacement discrimination processing, convert it into a target liver mask that adapts to the PDFF sequence data.

[0043] Step S4: Identify the intrahepatic vascular region based on the target liver mask and extract the effective liver parenchyma region;

[0044] Step S5: Abnormal voxels within the extracted effective liver parenchyma region are removed, and a three-dimensional liver fat volume distribution is constructed based on the PDFF values ​​of the remaining effective voxels. The average liver fat fraction is then calculated.

[0045] Step S6: Output the three-dimensional liver fat volume distribution and the average liver fat fraction.

[0046] As an optional implementation, in step S1, the multimodal MRI image data of the subject to be tested includes at least T1-weighted dual-echo sequence images and proton density fat fraction images (i.e., PDFF sequence images). It should be understood that in various embodiments of the present invention, the subject to be tested is liver fat, so the location corresponding to the MRI image is the abdominal image.

[0047] In step S2, the liver segmentation model used is a pseudo-three-dimensional liver semantic segmentation model obtained by pre-training the Pseudo-3D U-Net network based on the T1-weighted dual-echo image set. The Pseudo-3D U-Net network introduces the context information of the current slice and its neighboring slices to perform semantic segmentation on the liver region in the T1-weighted sequence data, so as to output the initial three-dimensional liver mask.

[0048] In embodiments of the present invention, a pseudo-3D liver semantic segmentation model is constructed by collecting multimodal abdominal magnetic resonance imaging data of the subject, and then iteratively trained based on the Pseudo-3D U-Net network to obtain a liver segmentation model. In the implementation of the present invention, adjacent slices are used as joint inputs to enhance the modeling ability of the continuity of the three-dimensional structure while maintaining computational efficiency.

[0049] As an example, multimodal abdominal magnetic resonance imaging data of the subjects are collected. The multimodal images include at least T1-weighted dual-echo sequence images and proton density fat fraction images. The data are de-identified, quality-screened and divided into queues according to preset standards to form a model training queue and a validation queue. The validation queue performs artificial region of interest sampling of PDFF.

[0050] During training, the T1-weighted dual-echo images of the training queue are used as the model input of Pseudo-3D U-Net, and the liver segmentation mask is used as the output. After the model is trained until convergence, a pseudo-3D liver semantic segmentation network is constructed to realize automatic liver segmentation based on MRI images.

[0051] like Figure 5 The diagram shows the structure of the Pseudo-3D U-Net model. In the embodiment of the present invention, the model uses adjacent slices (n-1, n, n+1) stacked as a six-channel input (each slice contains T1-IP and T1-OP images; the first and last slices are the two nearest adjacent slices stacked), thereby preserving the three-dimensional structural information while avoiding the high computational cost of traditional full three-dimensional networks, effectively improving the segmentation accuracy and efficiency.

[0052] The model training process, as an example, was based on data from 21 labeled cases, each containing 60-72 slices. Dice Loss was used as the objective loss function to optimize the learning process, and a liver semantic segmentation model based on MRI was finally established, exhibiting good segmentation consistency and generalization ability (dice > 0.95).

[0053] Therefore, by inputting a T1-weighted dual-echo sequence image of the object to be detected into a trained and deployed liver segmentation model, a liver segmentation mask output is obtained.

[0054] Combined with appendix Figure 5 The Pseudo-3D U-Net network structure shown is used in the embodiments of the present invention to design a liver segmentation model based on this architecture and obtain a liver segmentation model for actual deployment and use after training.

[0055] Combined with appendix Figure 5 As shown, in a specific embodiment, the overall architecture of the Pseudo-3D U-Net network structure is a symmetrical U-shaped structure, which includes a shrinking path (Encoder) and an expanding path (Decoder).

[0056] The input layer is designed to receive stacked 6-channel 3D contextual feature data, i.e., with a size of H×W×6, containing T1-IP and T1-OP images of layers n-1, n, and n+1, and is mapped to a 16-channel feature map through initial convolution.

[0057] The shrinkage path is designed as follows: it comprises four downsampling stages. Each stage consists of two pseudo-3D convolutional layers (using 3×3 kernels, decoupling the 3×3×3 convolution into a 1×3×3 spatial convolution and a 3×1×1 inter-layer temporal / spatial context convolution) concatenated together. The feature map size is then halved using a 2×2 max-pooling layer with a stride of 2. The number of channels in each stage is 16, 32, 64, and 128, respectively.

[0058] The bottleneck layer is designed with 256 channels and uses two pseudo-3D convolutional layers to extract deep semantic features.

[0059] The expansion path is designed as follows: it consists of four upsampling stages. Each stage first doubles the feature map size using a 2×2 transposed convolution, then concatenates it with the feature map of the corresponding scale in the contraction path along the channel dimension using a skip connection, and finally connects two 3×3 pseudo-3D convolutional layers. The number of channels in each stage decreases to 128, 64, 32, and 16 respectively.

[0060] The output layer is designed as follows: a 1×1 convolutional layer maps the number of channels to 2 (liver target and background), and uses the Softmax function to output the probability that each pixel belongs to the liver region. All hidden layers in the network are connected to a batch normalization layer and a ReLU activation function after convolution.

[0061] The training process of the model can be carried out based on the PyTorch 2.0 framework and the Ubuntu 22.04 operating system software environment.

[0062] Configure the batch size for training, such as setting it to 16, and the total number of training epochs to 500 for iterative training. The training process uses the Adam optimizer with a learning rate of 1×10⁻⁶. -3 .

[0063] The model training design uses Dice Loss as the objective loss function, and the formula for example is expressed as follows:

[0064] ;

[0065] Where M is the total number of voxels in the current slice, p i This represents the probability that the model predicts the current voxel belongs to the liver, g.i This represents the actual liver mask label value (0 or 1). To prevent smoothing terms with a denominator of 0, the value is set to 1 × 10. -5 When the training reaches around the 85th round, the loss function tends to stabilize, the model converges, and the weights at this point are saved as the final liver segmentation model.

[0066] Furthermore, after obtaining the initial 3D liver mask output by the pseudo 3D liver semantic segmentation model, a morphology-based liver denoising algorithm (morphological constraints) is used to remove pseudo regions that do not conform to anatomical rules, thereby obtaining an initial liver mask with consistent structure.

[0067] As an optional example, a morphology-based liver denoising algorithm is configured to be constrained by the anatomical continuity of the liver morphology (i.e., the liver's connected regions are regular and do not naturally break into three or more independent parts): all pixel clusters in the mask are identified and sorted by area, retaining the top two largest clusters and removing the remaining smaller clusters. Specifically, if there are only two or fewer pixel clusters, noise clusters with an area less than 5 pixels are additionally removed. Thus, the morphology-constrained denoising algorithm removes isolated noise while preserving the main structural regions, effectively suppressing segmentation artifacts and ensuring that the liver contour conforms to the anatomical structure.

[0068] Furthermore, in response to the differences in slice thickness, scale, and center position between the T1-weighted image and the proton density fat fraction image, in step S3, cross-modal mask alignment processing is performed on the initial liver mask based on the image size and interlayer structure of the PDFF sequence, thereby generating a liver mask consistent with the PDFF-MRI image space to ensure the consistency between the subsequent fat fraction statistical region and the PDFF image spatial position.

[0069] As an optional implementation, cross-modal mask alignment processing is performed on the initial liver mask, specifically including the following steps:

[0070] First, based on the image size and interlayer structure of the PDFF sequence, the liver initial mask is resampled using an interpolation algorithm. The liver initial mask is then transformed to the matrix size, voxel spacing, and interlayer structure corresponding to the PDFF sequence image, so that it initially corresponds to the size and interlayer structure of the PDFF sequence image, thus completing the interlayer mapping operation.

[0071] Then, the scale recognition algorithm is used to obtain the spatial scaling relationship between the two types of sequences based on the spatial size matching results between the two sequences, determine the scaling coefficient, and complete the scale matching operation.

[0072] Finally, a displacement discrimination algorithm is used to detect whether there is a spatial offset between the two sequences, and displacement correction is performed when a spatial offset exists, thus completing the displacement discrimination and correction.

[0073] As an optional specific example, in order to accurately map the liver region (segmented from T1 images) to each layer of the PDFF image, even if the number of layers and locations scanned by the two are different, we designed the inter-layer mapping operation to adopt an inter-layer mapping method based on interpolation-extrapolation, which specifically includes the following process:

[0074] (1) Locate the reference layer: For each layer of PDFF, find the two nearest and adjacent T1 layers in T1, namely the main layer (MT1_main) and the reference layer (MT1_ref). The main layer is the T1 layer that is closest to the current PDFF layer.

[0075] (2) Interpolation of PDFF layers falling between T1 layers: If the PDFF layer is located between the main layer and the reference layer, its corresponding liver region is obtained by linearly mixing these two layers, as shown in the following formula:

[0076] ;

[0077] Where α represents the relative position ratio of the PDFF layer between the T1 master layer and the reference layer (with a value between 0 and 1), thus generating a smooth transition in the liver region;

[0078] (3) Extrapolation of PDFF layers outside the T1 layer range: If a PDFF layer is located outside the T1 layer range, it is generated by extrapolation through the extension directions defined by the main layer and the reference layer, as shown in the following formula:

[0079] ;

[0080] Here, β is the extrapolation factor. When the PDFF layer is below the main layer, β takes a negative value, but the extrapolation process still applies.

[0081] As an optional implementation, the scale matching operation aims to automatically calculate the scaling ratio between the T1 and PDFF images. We designed three complementary sub-modules to implement this, specifically:

[0082] (1) The AbdomenXScaling module and the AbdomenYScaling module are used to measure the belly width of the two modal images on the X and Y axes, respectively, and the proportion of the most frequently occurring slices is used as the scaling factor; the belly region of the PDFF image is obtained by retaining the largest connected region to exclude external artifacts.

[0083] (2) LiverScaling module: On the liver region aligned across modalities, the optimal scale is selected by evaluating the reward function of different scaling factors. This can handle situations such as arm occlusion or abnormal body shape. The specific design of the reward function is as follows:

[0084] ;

[0085] Among them, MR ij This represents the grayscale value at pixel location (i,j), where (i,j) originates from the mask area. The grayscale reward measures the degree of grayscale matching of the liver region under the current scaling result. When the scaling ratio is unreasonable, the selected area may contain many non-liver tissues, which typically introduce higher pixel response values, thus reducing the grayscale reward. The valid-pixel reward constrains the selection of candidate scaling factors, where Scaling represents the current candidate scaling factor, and α is the weight coefficient of the scaling term, which can be adjusted experimentally. The weight coefficient α of the scaling term can be used to adjust the contribution balance between the grayscale reward and the scaling reward, enabling the model to achieve a better trade-off between the degree of grayscale matching and the effective scaling ratio.

[0086] By maximizing the reward function to guide the algorithm to convergence, the system determines that when the Reward reaches the global maximum value, the corresponding scaling factor is the candidate scaling factor determined by the LiverScaling module.

[0087] Therefore, the system outputs candidate scaling ratios and confidence levels through three sub-modules, and finally selects the scaling ratio with the highest confidence level as the system output.

[0088] As an optional implementation, the displacement discrimination and correction operation is intended to verify whether the coordinate displacement parameters in the DICOM header file are truly effective. A true displacement should allow the mask to more accurately cover the liver region, thereby detecting a higher PDFF fat score within that region. Therefore, in the method of this invention, based on the center position, boundary position, and / or registration results between the initial liver mask and the PDFF sequence image, it is determined whether a displacement offset exists, and translation correction is performed when the displacement offset exceeds a preset threshold.

[0089] In a specific embodiment, the displacement discrimination and correction operation includes the following process:

[0090] First, an abdominal mask is generated based on the T1 image, and its size is adjusted according to the scaling factor to obtain the reference mask;

[0091] Next, the displacement parameters are extracted from the DICOM file and the translated candidate mask is generated;

[0092] Then, compare the average PDFF values ​​measured by the two masks within the coverage area of ​​the nearest PDFF layer: if the PDFF value of the mask decreases after translation, it indicates that the displacement causes the mask to deviate from the liver (the PDFF value of the liver is usually higher than that of the surrounding tissue), thus determining that the parameter is a pseudo displacement and should be discarded; otherwise, it is determined to be a true displacement and adopted.

[0093] The final determined displacement parameters are fed back to the LiverScaling module to re-optimize the scaling factor, thereby achieving accurate adaptive geometric alignment of the cross-modal image.

[0094] Therefore, displacement discrimination and correction effectively avoid misregistration caused by DICOM coordinate system errors, ensuring the spatial accuracy of subsequent quantitative fat analysis.

[0095] Further, in step S4, the intrahepatic vascular region is identified based on the target liver mask, and the effective liver parenchyma region is extracted, specifically including the following steps:

[0096] Based on the target liver mask, a structural image enhancement method based on the Meijering filter is used to identify and extract intrahepatic vascular tissue to distinguish vascular tissue from liver parenchyma, thereby reducing the interference of vascular signals on liver parenchyma fat estimation.

[0097] The intrahepatic vascular tissue is removed from the target liver mask, and the remaining area after removal is used as the effective liver parenchyma area.

[0098] Further, in step S5, abnormal voxels within the extracted effective liver parenchyma region are removed, and a three-dimensional liver fat volume distribution is constructed based on the PDFF values ​​of the remaining effective voxels. The average liver fat fraction is then calculated, including the following processes:

[0099] The effective liver parenchyma fat fractional data is compensated and corrected based on the outlier suppression strategy to obtain voxel-level fat distribution data; the outlier suppression strategy includes removing extreme noise points based on preset thresholds or statistical distributions.

[0100] The corrected fat distribution data were reconstructed in three dimensions to generate a liver fat volume map, and a quantitative index of whole liver fat (based on the proton density fat fraction assessed by MRI-PDFF) was calculated, namely the effective liver parenchyma ratio, which is obtained by the ratio of the number of voxels in the effective liver parenchyma region to the total number of voxels in the target liver mask, and the value is expressed as a percentage.

[0101] {Example 2}

[0102] Based on the above embodiments of the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging, we will describe the implementation process of the above method in more detail with specific examples.

[0103] In this example, the design of an automated quantitative detection system for liver fat based on magnetic resonance imaging (MRI) (hereinafter referred to as the ALFQ system) is proposed, such as... Figure 1 As shown, it mainly includes various functional modules such as data reading and classification, T1 sequence liver segmentation, cross-sequence mask adaptation, intrahepatic vascular identification and effective liver parenchyma region extraction, fat fraction statistical analysis, and result output. It realizes end-to-end automated quantitative detection and analysis, automatically completing the process of MRI raw data import, liver region segmentation, mask cross-sequence matching, effective liver parenchyma screening, three-dimensional liver fat volume distribution construction, and quantitative result output, reducing manual intervention and improving measurement efficiency, consistency, and clinical applicability.

[0104] To verify the accuracy and consistency of the method in the foregoing embodiments of the present invention, system performance was evaluated based on data from 30 independent cohorts of subjects. According to the Bland–Altman analysis results (see...),... Figure 2 The ALFQ system showed only a low systematic bias of 0.85% compared to human measurements; furthermore, the intraclass correlation coefficient (ICC(2,1)) reached 0.993 (see...). Figure 3 This indicates that the ALFQ results of this invention have extremely high consistency with manual measurement results. In summary, the ALFQ system design proposed in this invention possesses good stability, accuracy, and clinical application potential.

[0105] The following describes the specific implementation process of the present invention in detail, taking the construction process of the three-dimensional liver fat volume distribution of patient GONYESEJ as an example.

[0106] Step 1: MRI data input and classification reading.

[0107] The input is the raw MRI image data of the subject, GONYESEJ, including T1-weighted dual-echo sequences and PDFF sequences. The T1-weighted dual-echo sequences contain 144 DICOM files, and the PDFF sequences contain 64 DICOM files. The ALFQ system first automatically reads the input DICOM files and classifies them according to sequence name, image parameters, or preset rules to create separate T1-weighted dual-echo sequence and PDFF sequence datasets, providing standardized input for subsequent segmentation and quantitative analysis.

[0108] Step 2: Liver semantic segmentation based on T1-weighted dual-echo sequences.

[0109] The T1-weighted dual-echo sequence is input into a pre-trained and deployed liver segmentation model based on Pseudo-3D U-Net to perform layer-by-layer semantic segmentation of the liver region, outputting a three-dimensional initial liver mask adapted to the spatial structure of the T1 sequence. As mentioned above, Pseudo-3D U-Net, by introducing contextual information from adjacent slices, can enhance the representation of three-dimensional anatomical continuity while maintaining the computational efficiency of two-dimensional convolution, thereby improving the accuracy of liver boundary recognition and segmentation stability.

[0110] Step 3: Cross-sequence mask conversion and adaptation.

[0111] Since T1-weighted dual-echo sequences and PDFF sequences may differ in spatial resolution, slice thickness, matrix size, or field of view, the liver mask obtained in step two needs to be converted into a target mask suitable for PDFF sequences.

[0112] Specifically, firstly, an interpolation algorithm is used to resample the initial mask to achieve a preliminary correspondence with the size and interlayer structure of the PDFF sequence image. Then, a scale recognition algorithm is used to calculate the spatial scaling relationship between the two types of sequences, determining the optimal scaling factor in this embodiment to be 0.84. Further, a displacement discrimination algorithm is used to detect the spatial offset between the two sequences, and the results show that no additional displacement correction is needed. Based on this, a liver mask strictly corresponding to the PDFF sequence is established to ensure that the subsequent fat fraction statistical region is consistent with the spatial location of the PDFF image.

[0113] Step 4: Extraction of effective liver parenchyma area and quantitative analysis of liver fat.

[0114] After obtaining a liver mask adapted to the PDFF sequence, the system uses structural image enhancement methods to identify and extract intrahepatic vascular regions to distinguish vascular tissue from liver parenchyma. By removing non-target tissues such as blood vessels, the system estimates the effective liver parenchyma proportion to be 88%.

[0115] Subsequently, by combining statistical distribution analysis methods, extreme noise points and abnormal voxel values ​​were further removed to reduce the impact of local artifacts, edge errors and occasional high noise on the fat quantification results.

[0116] Based on the above processing results, the system constructed a three-dimensional liver fat volume distribution map of patient GONYESEJ, as follows: Figure 4 As shown.

[0117] The average liver fat fraction of the patient was calculated to be 10.23% based on the PDFF values ​​of all effective liver parenchyma voxels. Compared with the manual measurement result of 11.05%, the absolute error was only 0.82%, indicating that the quantitative results of the method of the present invention in the test case have high accuracy.

[0118] Step 5: Result generation and file output.

[0119] After completing the quantitative analysis of fat, the system outputs the generated three-dimensional liver fat volume distribution results, the average liver fat score, and the basic information of the test subject. The basic information of the test subject can optionally include name, age, height, weight, BMI, and patient number.

[0120] Preferably, the 3D visualization results are output in PNG image format, and the quantitative analysis results and corresponding patient information are output in Excel file format. All output files are uniformly named and archived according to the patient file number to facilitate subsequent clinical image reading, case management, follow-up comparison and scientific research statistical analysis.

[0121] {Example 3}

[0122] In conjunction with the aforementioned embodiments of the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging, this invention also proposes an automated quantitative detection device for liver fat based on multimodal magnetic resonance imaging to implement this method, comprising the following modules:

[0123] The MRI image input module is used to acquire MRI image data of the subject under test, and to read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data;

[0124] The liver segmentation module is used to input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask that adapts to the T1-weighted sequence.

[0125] The alignment correction module is used to perform cross-modal mask alignment processing on the initial liver mask based on the image size and interlayer structure of the PDFF sequence. Through interpolation resampling, scale recognition and displacement discrimination processing, it is converted into a target liver mask that adapts to the PDFF sequence data.

[0126] An effective liver parenchyma region extraction module is used to identify intrahepatic vascular regions based on the target liver mask and extract the effective liver parenchyma region.

[0127] A quantitative detection module is used to remove abnormal voxels within the extracted effective liver parenchyma region, construct a three-dimensional liver fat volume distribution based on the PDFF values ​​of the remaining effective voxels, and calculate the average liver fat fraction; and

[0128] The output module is used to output the three-dimensional liver fat volume distribution and the average liver fat fraction.

[0129] {Example 4}

[0130] In conjunction with the aforementioned embodiments of the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging, the present invention also proposes a computer system comprising: one or more processors; and a memory storing operable instructions.

[0131] When the aforementioned instructions are executed by one or more processors, the aforementioned one or more processors perform operations, including the process of executing the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging according to the aforementioned embodiments.

[0132] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. An automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging, characterized in that, Includes the following steps: Step S1: Acquire multimodal MRI image data of the object to be tested, and read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data; Step S2: Input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask adapted to the T1-weighted sequence; Step S3: Based on the image size and interlayer structure of the PDFF sequence, perform cross-modal mask alignment processing on the initial liver mask. Through interpolation resampling, scale recognition and displacement discrimination processing, convert it into a target liver mask that adapts to the PDFF sequence data. Step S4: Identify the intrahepatic vascular region based on the target liver mask and extract the effective liver parenchyma region; Step S5: Remove abnormal voxels from the extracted effective liver parenchyma area, construct a three-dimensional liver fat volume distribution based on the PDFF value of the remaining effective voxels, and calculate the average liver fat fraction. as well as Step S6: Output the three-dimensional liver fat volume distribution and the average liver fat fraction.

2. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 1, characterized in that, In step S2, the liver segmentation model is a pseudo-three-dimensional liver semantic segmentation model obtained by training a Pseudo-3D U-Net network based on a T1-weighted dual-echo image set. The Pseudo-3D U-Net network introduces the context information of the current slice and its neighboring slices to perform semantic segmentation on the liver region in the T1-weighted sequence data, so as to output a three-dimensional liver initial mask.

3. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 2, characterized in that, In step S2, after obtaining the initial 3D liver mask output by the pseudo 3D liver semantic segmentation model, the pseudo regions that do not conform to the anatomical rules are removed using a morphological constraint denoising algorithm to obtain an initial liver mask with consistent structure.

4. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 1, characterized in that, In step S3, cross-modal mask alignment processing is performed on the initial liver mask based on the image size and interlayer structure of the PDFF sequence. Includes the following processes: First, based on the image size and interlayer structure of the PDFF sequence, an interpolation algorithm is used to resample the initial liver mask, transforming the initial liver mask to a matrix size, voxel spacing, and interlayer structure corresponding to the PDFF sequence image, so that it initially corresponds to the size and interlayer structure of the PDFF sequence image. Then, the scaling algorithm is used to obtain the spatial scaling relationship between the two types of sequences based on the spatial size matching results between the two sequences, and the scaling factor is determined. Finally, a displacement discrimination algorithm is used to detect whether there is a spatial offset between the two sequences, and displacement correction is performed when a spatial offset exists.

5. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 4, characterized in that, The method for detecting spatial offset between two sequences based on the displacement discrimination algorithm and performing displacement correction when spatial offset exists includes the following process: Based on the center position, boundary position and / or registration results between the initial liver mask and the PDFF sequence image, it is determined whether there is a displacement offset, and translation correction is performed when the displacement offset exceeds a preset threshold.

6. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 4, characterized in that, In step S4, the intrahepatic vascular region is identified based on the target liver mask, and the effective liver parenchyma region is extracted. This specifically includes the following steps: Based on the target liver mask, a structural image enhancement method based on the Meijering filter is used to identify and extract intrahepatic vascular tissue in order to distinguish vascular tissue from liver parenchyma. The intrahepatic vascular tissue is removed from the target liver mask, and the remaining area after removal is used as the effective liver parenchyma area.

7. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 4, characterized in that, In step S5, abnormal voxels within the extracted effective liver parenchyma region are removed, and a three-dimensional liver fat volume distribution is constructed based on the PDFF values ​​of the remaining effective voxels. The average liver fat fraction is then calculated, including the following processes: Based on the outlier suppression strategy, the effective liver parenchyma fat fraction data are compensated and corrected to obtain voxel-level fat distribution data. The corrected fat distribution data is reconstructed in three dimensions to generate a liver fat volume map, and the quantitative index of whole liver fat, namely the effective liver parenchyma ratio, is calculated. This is obtained by the ratio of the number of voxels in the effective liver parenchyma region to the total number of voxels in the target liver mask, and the value is expressed as a percentage.

8. The method for automated quantitative detection of liver fat based on multimodal magnetic resonance imaging according to claim 7, characterized in that, In step S5, the outlier suppression strategy includes removing extreme noise points based on a preset threshold or statistical distribution.

9. An automated quantitative detection device for liver fat based on multimodal magnetic resonance imaging, implementing the method of claim 1, characterized in that, include: The MRI image input module is used to acquire MRI image data of the subject under test, and to read and classify the MRI image data to obtain T1-weighted sequence data and PDFF sequence data; The liver segmentation module is used to input the T1-weighted sequence data into the liver segmentation model to obtain an initial liver mask that adapts to the T1-weighted sequence. The alignment correction module is used to perform cross-modal mask alignment processing on the initial liver mask based on the image size and interlayer structure of the PDFF sequence. Through interpolation resampling, scale recognition and displacement discrimination processing, it is converted into a target liver mask that adapts to the PDFF sequence data. An effective liver parenchyma region extraction module is used to identify intrahepatic vascular regions based on the target liver mask and extract the effective liver parenchyma region. The quantitative detection module is used to remove abnormal voxels in the extracted effective liver parenchyma area, construct a three-dimensional liver fat volume distribution based on the PDFF value of the remaining effective voxels, and calculate the average liver fat fraction. as well as The output module is used to output the three-dimensional liver fat volume distribution and the average liver fat fraction.

10. A computer system, characterized in that, include: One or more processors; as well as Memory stores instructions that can be operated. When the instructions are executed by one or more processors, they cause the aforementioned one or more processors to perform operations, including the process of performing the automated quantitative detection method for liver fat based on multimodal magnetic resonance imaging as described in any one of claims 1-8.

Images