Cardiac magnetic resonance cardiac function assessment method based on multi-task learning

Abstract

The present invention discloses a cardiac magnetic resonance cardiac function assessment method based on multi-task learning, which belongs to the technical field of medical image processing. The present invention labels, enhances and preprocesses cardiac magnetic resonance image data; then uses the pre-processed data to train a multi-task learning model, introduces regional geometric consistency loss to supervise the multi-task training process, and improves the accuracy and generalization performance of the model; then inputs the cardiac magnetic resonance image to be evaluated into the trained multi-task learning model, obtains the left and right ventricular endocardial segmentation results in the cardiac anatomical structure, the left ventricular myocardial segmentation results, and the right ventricular insertion point and left ventricular center point in the cardiac key point position, further calculates the ventricular volume to obtain the left and right ventricular ejection fraction; calculates the thickness of each segment of the left ventricular myocardium; and then outputs the cardiac function assessment result. The present invention improves the accuracy and generalization performance of the model, and solves the problems of low accuracy and long analysis time of traditional methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a cardiac magnetic resonance cardiac function assessment method based on multi-task learning. Background Art

[0002] Cardiac magnetic resonance imaging (CMR) is widely used in the diagnosis and treatment of heart diseases due to its noninvasive nature, high soft tissue contrast, and multi-parameter and multi-directional imaging capabilities. CMR provides detailed information on cardiac morphology, structure, function, blood perfusion, and tissue properties, making it crucial for the early diagnosis of primary myocardial diseases, risk assessment, and monitoring of the therapeutic efficacy of heart failure patients. Traditional CMR imaging techniques, such as cine cardiac imaging, can obtain dynamic images throughout the cardiac cycle, assessing the morphology and function of the left and right atria, left and right ventricles. However, these techniques have limited clinical application due to complex procedures, high technician requirements, the need for patients to cooperate with multiple breath-holds, and low success rates in patients with poor tolerance. With the continuous advancement of imaging technology, CMR examination sequences and methods have been continuously updated, resulting in the development of CMR-specific cardiac function assessment indices such as ejection fraction, ventricular mass index, strain, interventricular septal curvature ratio, right atrioventricular annular systolic excursion, and vascular compliance.

[0003] Despite its unique advantages in assessing cardiac function, CMR still faces challenges in practical application, such as long imaging times, high technician requirements, and the need for patients to cooperate with breath-holding. Furthermore, the quality of CMR imaging is affected by factors such as the position, size, and motion of the heart, resulting in image quality that may not meet expectations in some cases. Therefore, the development of high-quality, time-efficient CMR imaging technology, combined with advanced image segmentation algorithms to accurately assess cardiac structure and function, is of great significance in the diagnosis of heart disease. Summary of the Invention

[0004] In response to the above existing technologies, the present invention aims to provide a multi-task learning-based cardiac magnetic resonance cardiac function assessment method. This method aims to optimize the performance of cardiac anatomical structure segmentation and cardiac key point identification through deep learning techniques, thereby improving the accuracy and efficiency of cardiac function assessment. This method is expected to overcome the low assessment accuracy and long analysis time inherent in existing technologies, achieving efficient cardiac anatomical structure segmentation and key point detection, while reducing the model's reliance on training data and providing high-quality cardiac function assessment results.

[0005] The technical solutions of the present invention are as follows:

[0006] The present invention proposes a cardiac magnetic resonance cardiac function assessment method based on multi-task learning, comprising the following steps:

[0007] S1. Obtain cardiac magnetic resonance images and annotate cardiac segmentation labels and cardiac key point labels to obtain an annotated dataset, and perform data enhancement on the dataset;

[0008] S2. Preprocessing the data set constructed in S1 and training a multi-task learning model, wherein the multi-task learning model includes a cardiac anatomical structure segmentation sub-branch and a cardiac key point position recognition sub-branch;

[0009] S3. Input the cardiac magnetic resonance image to be evaluated into the trained multi-task learning model to obtain the left and right ventricular endocardial segmentation results, the left ventricular myocardial segmentation results, and the right ventricular insertion point and left ventricular center point in the cardiac anatomical structure;

[0010] S4. Calculating the ventricular volume to obtain the left and right ventricular ejection fractions based on the segmentation results of the left and right ventricular endocardium in the cardiac anatomical structure;

[0011] S5. Divide the left ventricular myocardium segmentation result based on the right ventricular insertion point and the left ventricular center point in the cardiac key point positions, and calculate the thickness of each segment of the left ventricular myocardium;

[0012] S6. Outputting a cardiac function assessment result based on the ejection fraction and the thickness of each segment of the left ventricular myocardium.

[0013] Preferably, the multi-task learning model in step S2 includes a multi-task shared encoder and a decoder customized for tasks corresponding to different sub-branches, and a Transformer module is introduced between the encoder and the decoder, so that the model's ability to capture long-range dependencies and understand complex patterns and representations is enhanced; the model input is a two-dimensional magnetic resonance short-axis movie sequence image of the heart, and the output is the left and right ventricular endocardium segmentation results, the left ventricular myocardium segmentation results, and the right ventricular insertion point and the left ventricular center point in the heart key point position.

[0014] Preferably, during the model training process of step S2, the total loss function L is constructed as follows:

[0015] L=ω seg ·L seg +ω lm ·L lm +ω rgc ·L rgc

[0016] Among them, ω seg is the segmentation task loss L seg Weight, ω lm Loss L for key point detection task lm Weight, ω rgc is the regional geometric consistency loss L rgc weight;

[0017] L seg =λ1·L CE +λ2·L Dice

[0018] Among them, λ1 is the corresponding cross entropy loss L CE The weight of λ2 corresponds to the Dice loss L Dice weight;

[0019]

[0020] Among them, ω is the proportional constraint value, which is a positive number; ε is the curvature constraint value; is a constant;

[0021]

[0022] Among them, p represents the key points of the heart, including the anterior insertion point of the right ventricle Right ventricular posterior insertion point and the left ventricular center point P LC ; For any given point q on the myocardial contour, C contains the set of points that define the contour; ‖·‖2 represents the L2 norm, which is used to quantify the Euclidean distance between points; |·| represents the Euclidean distance between two key points; is the sigmoid function, which is used to introduce nonlinearity; k is the proportional coefficient, which is used to adjust the impact of distance difference on the sensitivity of the loss function.

[0023] The innovations and advantages of the present invention are:

[0024] (1) This paper innovatively introduces the concept of multi-task learning into cardiac magnetic resonance cardiac function assessment. By improving the network structure, it efficiently segments the cardiac anatomical structure and identifies the locations of key points, and introduces regional geometric consistency loss to supervise the multi-task training process, thereby improving the accuracy and generalization performance of the model and solving the problems of low accuracy and long analysis time of traditional methods.

[0025] (2) The present invention adopts a rigid deformation method for data enhancement, which effectively expands the data set and improves the generalization ability of the model. At the same time, the Transformer module is introduced between downsampling and upsampling, which enhances the model's ability to capture long-range dependencies and understand complex patterns and representations.

[0026] (3) The present invention provides a comprehensive cardiac function assessment result by comprehensively evaluating the ejection fraction and myocardial wall thickness. This method not only improves the accuracy of diagnosis, but also has high operability and repeatability, and is of great value to clinical medical research. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Flowchart of the cardiac function assessment method of the present invention.

[0028] Figure 2 This is a flow chart of data preprocessing in the present invention.

[0029] Figure 3 This is a framework diagram of the heart segmentation and key point detection model based on multi-task learning in the present invention.

[0030] Figure 4 3. This is a comparison chart of the left ventricular myocardial segmentation results of an embodiment of the present invention and the two-stage TransUNet method.

[0031] Figure 5 Schematic diagram of the myocardial segment thickness calculation principle in the present invention.

[0032] Figure 6 This is the flow chart of the central function diagnosis report generation and auxiliary diagnosis and treatment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described and illustrated below in conjunction with specific embodiments. The embodiments are merely illustrative of the present disclosure and do not limit its scope. The technical features of the various embodiments of the present invention may be combined accordingly, provided that there is no conflict between them.

[0034] In this embodiment, if Figure 1 As shown, a cardiac magnetic resonance cardiac function assessment method based on multi-task learning is provided, which includes the following steps:

[0035] S1. Obtain cardiac magnetic resonance images and annotate cardiac segmentation labels and cardiac key point labels to obtain an annotated dataset, and perform data enhancement on the dataset;

[0036] In this embodiment, S1 is specifically as follows: obtain two-dimensional magnetic resonance movie sequence images collected from different centers and different equipment, from the heart base to the heart apex, and carefully annotate the images, including but not limited to cardiac segmentation labels such as the left and right ventricular blood pools and the left ventricular myocardium, and cardiac key point labels such as the right ventricular insertion point and the left and right ventricular center points. Then, in order to perform model training and performance evaluation, the present invention divides the data set into a training set and a test set in a 6:1 ratio, and introduces a rigid deformation method for data enhancement, including random distance translation, random angle rotation, mirror flipping, etc., to expand the data volume of the training set. This method enhances the model's adaptability to changes in cardiac morphology by simulating the geometric changes of the heart between different individuals.

[0037] S2. Preprocessing the data set constructed in S1 and training a multi-task learning model, wherein the multi-task learning model includes a cardiac anatomical structure segmentation sub-branch and a cardiac key point position recognition sub-branch;

[0038] Reference Figure 2 , a series of measures were taken to preprocess the dataset. To enhance the contrast of magnetic resonance images, the histogram equalization method was used to redistribute the image pixel values. Then, to remove irrelevant information, the region of interest (ROI) was cropped from the image, which contained the target research object. In addition, the datasets from different sources were resampled to unify the spatial resolution. The specific operation was to adjust the voxel size and slice spacing of the image, which helped to eliminate the difference in scanner resolution. Finally, the min-max normalization method was used to map the image pixel values ​​to the range [0,1]. This can reduce numerical errors in the calculation process and improve numerical stability. The calculation formula is:

[0039]

[0040] Where x is the original data, x min and x max are the minimum and maximum values ​​in the data set, respectively, and x′ is the standardized data.

[0041] S3. Input the cardiac magnetic resonance image to be evaluated into the trained multi-task learning model to obtain the left and right ventricular endocardial segmentation results, the left ventricular myocardial segmentation results, and the right ventricular insertion point and left ventricular center point in the cardiac anatomical structure;

[0042] S4. Calculating the ventricular volume to obtain the left and right ventricular ejection fractions based on the segmentation results of the left and right ventricular endocardium in the cardiac anatomical structure;

[0043] S5. Divide the left ventricular myocardium segmentation result based on the right ventricular insertion point and the left ventricular center point in the cardiac key point positions, and calculate the thickness of each segment of the left ventricular myocardium;

[0044] S6. Outputting a cardiac function assessment result based on the ejection fraction and the thickness of each segment of the left ventricular myocardium.

[0045] In cardiac magnetic resonance imaging analysis, the traditional two-stage model has obvious limitations, such as the independence of cardiac segmentation and cardiac key point detection, which leads to the loss of inter-task dependency information and low computational efficiency. Therefore, in S3 of the cardiac magnetic resonance cardiac function assessment method proposed in the present invention, a multi-task learning network MLTUNet (Multi-task Learning TransUNet) is introduced, such as Figure 3As shown in the figure, for the input cardiac 2D MRI short-axis movie sequence images, ResNet50 is first applied to downsample the images three times to obtain feature maps; then the images are input to the Transformer module; the image size is adjusted through the convolution layer, BatchNorm normalization layer and ReLU activation function in sequence; then two sub-task branches are input, each of which contains three upsampling layers. Each layer uses a skip connection mechanism to fuse the low-level image features extracted by the encoder at each stage with the decoder. During the training process, the network output includes the segmentation prediction value of the cardiac anatomical structure (S pre ) and the predicted value of the heart key point position (L pre ), these predictions will be compared with the actual cardiac anatomical structure labels (S gt ) and heart key point location labels (L gt ) are combined to calculate the loss function. Through the back-propagation algorithm, these losses are used to guide and optimize the network training process to improve prediction accuracy. The multi-task learning network can focus on multiple related tasks simultaneously, improving its ability to capture changes in cardiac motion and structure, allowing the model to achieve better performance on each individual task.

[0046] The proposed network MLTNUet includes a multi-task shared encoder and a decoder customized for different subtasks. The Transformer module is introduced between the encoder and decoder, which enhances the model's ability to capture long-range dependencies and understand complex patterns and representations, thereby coping with the complexity of left and right ventricular endocardium segmentation in cardiac anatomy. This design allows the model to not only focus on local features when processing cardiac images, but also capture global contextual information, which is crucial for accurate cardiac structure segmentation. The Transformer module contains two submodules: the multi-head self-attention mechanism (MSA) and the feedforward neural network (FFN). The multi-head self-attention mechanism is a technology that allows the model to simultaneously focus on different position information in different representation subspaces. It splits the input vector into multiple heads, each head independently calculates the attention output, and then merges these outputs to enhance the model's ability to capture information. Let the input sequence be Q, K, V, where Q is the query, K is the key, and V is the value. The calculation of multi-head self-attention can be expressed as:

[0047] MHead(Q,K,V)=Concat(head1,…,head h )W o

[0048] head i =Att(QWi i Q ,KW i K ,VWi V )

[0049]

[0050] Among them, W i Q ,W i K ,W i V is the learnable weight matrix, d k is the dimension of the key (K) vector, Used to scale dot products to prevent the vanishing gradient problem.

[0051] A feedforward neural network (FFN) consists of two linear transformation layers and a ReLU activation function. The first linear transformation layer is responsible for expanding the input data to a higher-dimensional representation space. This process allows the model to capture richer feature information. Then, by introducing the ReLU activation function, the network can introduce the necessary nonlinear characteristics, which is crucial for simulating complex function mappings. Finally, the second linear transformation layer converts the nonlinearly processed data back to the original dimension or target dimension to ensure that the output is compatible with the rest of the model. Let the input be x, then the calculation process can be expressed as:

[0052] FFN(x)=max(0,xW1+b1)W2+b2

[0053] Among them, W i represents a series of trainable weight matrices, and b i represents the corresponding bias vector.

[0054] Furthermore, during the model training process, Dice loss and cross entropy loss are used to supervise the performance of the segmentation branch, while Wing loss is used to supervise the performance of key point detection. In addition, the regional geometric consistency loss L is introduced. rgc To measure the common performance of the two tasks, the back-propagation algorithm is used to update the model parameters. The loss functions are introduced as follows.

[0055] Segmentation branch loss L seg :

[0056] L seg =λ1·L CE +λ2·L Dice

[0057] Among them, L seg is the segmentation branch loss, λ1 is the corresponding cross entropy loss L CE The weight of λ2 corresponds to the Dice loss L Dice In this embodiment, both λ1 and λ2 are selected as 0.5.

[0058] Keypoint detection branch loss L lm :

[0059]

[0060] Among them L lm is the key point detection branch loss, ω is the proportional constraint value, which is a positive number that limits the range of the nonlinear part to [-ω, ω]; ∈ is the curvature constraint value, which is used to constrain the curvature of the nonlinear part; is a constant used to smoothly connect the linear and nonlinear parts of the above segmentation.

[0061] Regional geometric consistency loss L rgc :

[0062]

[0063] Among them, p represents a key cardiac landmark, including the right ventricular anterior insertion point Right ventricular posterior insertion point and the left ventricular center point P LC For any given point q on the myocardial contour, C contains the set of points that define the contour; ‖·‖2 represents the L2 norm, which is used to quantify the Euclidean distance between points; |·| represents the Euclidean distance between two key points; is the sigmoid function, which is used to introduce nonlinearity; k is the proportional coefficient, which is used to adjust the impact of distance difference on the sensitivity of the loss function.

[0064] By integrating the above multi-faceted loss functions, the final loss function is constructed as follows:

[0065] L=ω seg ·L seg +ω lm ·L lm +ω rgc ·L rgc

[0066] Among them, ω seg is the weight of the segmentation task loss function, ω lm is the weight of the loss function for the key point detection task, ω rgc is the weight of the regional geometric consistency loss function.

[0067] Furthermore, S4 specifically includes: based on the segmentation results of the left and right ventricular endocardium in the cardiac anatomical structure, identifying the ventricular contour by an edge tracking algorithm, counting the number of pixels within the contour, and estimating the single-layer ventricular cavity area in combination with the imaging resolution; then accumulating the areas of the left and right ventricles in all slices and multiplying them by the inter-slice resolution of the image to obtain the volumes of the left and right ventricles, and calculating the left and right ventricular ejection fractions based on the following formula:

[0068]

[0069] Where EF is the ejection fraction, EDV and ESV are the end-diastolic volume and end-systolic volume, respectively.

[0070] Table 1 shows the ejection fraction calculation results for the proposed MLTUNet and two-stage TransUNet models, using mean absolute error (MAE) and root mean square error (RMSE) as evaluation metrics. The results demonstrate that the proposed method achieves higher left ventricular ejection fraction accuracy than the existing two-stage model.

[0071] Table 1 Quantitative comparison of ejection fraction calculation accuracy between the present invention and the two-stage TransUNet algorithm

[0072]

[0073] Furthermore, S5 is specifically as follows: based on the position information of key cardiac points such as the left ventricular center point and the right ventricular insertion point, the left ventricular myocardium is divided into 16 segments. Figure 4 This is a comparison chart of the left ventricular myocardial segmentation results of the embodiment of the present invention and the two-stage TransUNet method. The two-stage TransUNet method has the problem of missed detection at the base. The segmentation results of each segment are thinner than those of the method of the present invention and are more different from the annotation labels. The results show that MLTUNet has excellent performance in the myocardial segmentation task; then refer to Figure 5 The maximum, minimum, and average wall thickness of each myocardial segment were calculated as part of the cardiac function assessment. The method involved extracting the center of gravity of the left ventricular myocardial contour and intersecting a ray from that center with the inner and outer contours of the left ventricular myocardium. The straight-line distance between the two intersection points, calculated based on the imaging resolution, was used as the thickness value at that location.

[0074] The ejection fraction and myocardial wall thickness calculation results in S6 can be directly output as cardiac function assessment results, or can be referred to Figure 6 As shown in the figure, the ejection fraction (EF) and myocardial wall thickness measurements are combined with clinical guidelines and expert knowledge graphs to generate a detailed cardiac function assessment report. This report comprehensively measures the heart's diastolic and systolic function, providing a scientific basis for clinical decision-making and achieving an accurate assessment of heart health.

[0075] This embodiment also provides a cardiac magnetic resonance cardiac function assessment system based on multi-task learning, including at least one memory and at least one processor, the memory storing a computer program, when the program is executed by the processor, enabling the processor to execute the above-mentioned cardiac magnetic resonance cardiac function assessment method based on multi-task learning.

[0076] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. Persons skilled in the art will readily appreciate that variations and modifications may be made without departing from the scope of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A cardiac magnetic resonance cardiac function assessment method based on multi-task learning, characterized in that: The method comprises the following steps: S1. Acquire cardiac magnetic resonance images and label cardiac segmentation labels and cardiac key point labels to obtain a labeled data set, and perform data enhancement on the data set; the cardiac magnetic resonance images include two-dimensional magnetic resonance short-axis movie sequence images of the heart; S2. Preprocessing the data set constructed in S1 and training a multi-task learning model, wherein the multi-task learning model includes a cardiac anatomical structure segmentation sub-branch and a cardiac key point position recognition sub-branch; In step S2, for the input cardiac 2D MRI short-axis cine image sequence, the multi-task learning model first downsamples the image three times using ResNet50 to obtain a feature map. This is then fed into the Transformer module. The image is resized sequentially through a convolutional layer, a BatchNorm normalization layer, and a ReLU activation function. Subsequently, two subtask branches are fed into each branch, each containing three upsampling layers. Each layer uses a skip connection mechanism to fuse the low-level image features extracted by each encoder stage with the decoder. During training, the multi-task learning model outputs include segmentation predictions for cardiac anatomical structures and predicted locations of cardiac key points. These predictions are combined with the actual cardiac anatomical structure labels and cardiac key point location labels to calculate the total loss function. The back-propagation algorithm is then used to optimize the network parameters of the multi-task learning model. During the model training process of step S2, the total loss function L is constructed as follows: L=ω seg ·L seg +oh lm ·L lm +oh rgc ·L rgc Among them, ω seg is the segmentation task loss L seg Weight, ω lm Loss L for key point detection task lm Weight, ω rgc is the regional geometric consistency loss L rgc weight; L seg =λ1·L CE +λ2·L Dice Among them, λ1 is the corresponding cross entropy loss L CE The weight of λ2 corresponds to the Dice loss L Dice weight; Among them, ω is the proportional constraint value, which is a positive number; ∈ is the curvature constraint value; is a constant; Among them, p represents the key points of the heart, including the anterior insertion point of the right ventricle Right ventricular posterior insertion point and the left ventricular center point P LC ; For any given point q on the myocardial contour, C contains the set of points that define the contour; ‖· ‖2 represents the L2 norm, which is used to quantify the Euclidean distance between points; |·| represents the Euclidean distance between two key points; is the sigmoid function, which is used to introduce nonlinearity; k is the proportional coefficient, which is used to adjust the effect of distance difference on the sensitivity of the loss function; S3. Input the cardiac magnetic resonance image to be evaluated into the trained multi-task learning model to obtain the left and right ventricular endocardial segmentation results, the left ventricular myocardial segmentation results, and the right ventricular insertion point and left ventricular center point in the cardiac anatomical structure; S4. Calculating the ventricular volume to obtain the left and right ventricular ejection fractions based on the segmentation results of the left and right ventricular endocardium in the cardiac anatomical structure; S5. Divide the left ventricular myocardium segmentation result based on the right ventricular insertion point and the left ventricular center point in the cardiac key point positions, and calculate the thickness of each segment of the left ventricular myocardium; S6. Outputting a cardiac function assessment result based on the ejection fraction and the thickness of each segment of the left ventricular myocardium.

2. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: In step S1, the two-dimensional magnetic resonance short-axis movie sequence images of the heart are annotated with heart segmentation labels and heart key point labels, and the images and their corresponding annotation information are regarded as a data set; the heart segmentation labels include the left and right ventricular blood pools and the left ventricular myocardium; the heart key point labels include the right ventricular insertion point and the left ventricular center point.

3. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: In step S1, a rigid deformation method is used to perform data enhancement on the dataset, and the data enhancement method includes translation, rotation and scaling.

4. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: The preprocessing operations in step S2 include: a. Use histogram equalization method to redistribute image pixel values; b. Crop the region of interest from the image; c. Resample datasets from different sources to unify spatial resolution by adjusting the voxel size and slice spacing of the images; d. Finally, the min-max normalization method is used to map the image pixel values ​​to the interval [0, 1] to reduce the numerical error in the calculation process.

5. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: The multi-task learning model in step S2 includes a multi-task shared encoder and decoders customized for tasks corresponding to different sub-branches. At the same time, a Transformer module is introduced between the encoder and decoder, which enhances the model's ability to capture long-range dependencies and understand complex patterns and representations. The model input is a two-dimensional magnetic resonance short-axis movie sequence image of the heart, and the output is the left and right ventricular endocardium segmentation results, the left ventricular myocardium segmentation results, and the right ventricular insertion point and left ventricular center point in the heart key point positions.

6. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: In step S4, the ventricular volumes are calculated to obtain the left and right ventricular ejection fractions, specifically: Based on the segmentation results of the left and right ventricular endocardium in the cardiac anatomy, the ventricular contour is identified using an edge tracking algorithm. The number of pixels within the contour is counted and the single-layer ventricular cavity area is estimated based on the imaging resolution. The areas of the left and right ventricles in all slices are then accumulated and multiplied by the inter-slice resolution of the image to obtain the volumes of the left and right ventricles. The left and right ventricular ejection fractions are then calculated based on the following formula: Where EF is the ejection fraction, EDV and ESV are the end-diastolic volume and end-systolic volume, respectively.

7. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: In step S5, the thickness of each segment of the left ventricular myocardium is calculated, specifically: Based on the position information of the left ventricular center point and the right ventricular insertion point, the left ventricular myocardium was divided into 16 segments, and the maximum thickness, minimum thickness and average thickness of the ventricular wall of each myocardial segment were calculated.

8. The cardiac magnetic resonance cardiac function assessment method based on multi-task learning according to claim 1, characterized in that: In step S6, the ejection fraction and myocardial wall thickness calculation results are directly output as cardiac function assessment results; Alternatively, based on the ejection fraction, myocardial wall thickness, and in combination with clinical guidelines and knowledge graphs, a cardiac function assessment report including cardiac diastolic function and systolic function is output as the cardiac function assessment result.

Images