Method for segmenting cardiac imagery, method for analyzing cardiac imagery

Abstract

This application discloses a cardiac image segmentation method and a cardiac image analysis method. The cardiac image segmentation method includes: acquiring the long-axis and short-axis data corresponding to the original cardiac magnetic resonance imaging (MRI) images; segmenting the long-axis and short-axis data respectively to obtain long-axis segmentation results and short-axis segmentation results; analyzing the long-axis and short-axis segmentation results based on a spatial registration method using covariance spectral decomposition to determine the short-axis apex and base positions, and removing redundant scan layers from the short-axis segmentation results to obtain short-axis data segmentation information containing only the heart. This method models the long-axis left ventricle segmentation results, achieves precise automatic localization of the apex and base of the heart through covariance spectral decomposition combined with spatial registration of projection extrema and short-axis slices, and then removes redundant scan layers from the short-axis segmentation results based on the short-axis apex and base positions to obtain accurate short-axis data segmentation information containing only the heart.

Description

Technical Field

[0001] This disclosure relates to the field of medical image processing technology, and in particular to a method for cardiac image segmentation and cardiac image analysis. Background Technology

[0002] In clinical cardiac function assessment, the dynamic changes in cardiac structure, such as the evolution of ventricular volume and myocardial thickness during the cardiac cycle, are crucial analytical indicators. Currently, such analyses typically rely on experienced physicians using specialized software to manually or semi-manually process short-axis cardiac image sequences frame by frame. This includes outlining the contours of the ventricles and myocardium, as well as identifying key anatomical locations such as the apex and base of the heart. This process is not only time-consuming but also demands a high level of expertise from the operator, resulting in significant challenges to fully automated analysis in clinical practice.

[0003] Despite advancements in medical image segmentation technology, significant limitations remain in the analysis of short-axis cardiac images. Because a single short-axis slice only provides local anatomical information and lacks a complete representation of the heart's overall three-dimensional structure, segmentation results based on this data often fall short of clinical accuracy requirements. In particular, short-axis images themselves cannot provide sufficient positional context for automatically and reliably identifying key landmarks such as the apex and base of the heart; manual localization confirmation is still necessary. These factors collectively limit the reliability and clinical application potential of fully automated cardiac function analysis methods. Summary of the Invention

[0004] In view of this, the present disclosure provides a cardiac image segmentation method and a cardiac image analysis method, which can solve the problems that the current technology mainly relies on manual or semi-automatic processing of a single short-axis image sequence, and its inherent two-dimensional locality limits the accuracy and automation of segmentation, and cannot automatically identify key anatomical locations.

[0005] In a first aspect, embodiments of this disclosure provide a method for cardiac image segmentation, including: Obtain the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging; The major axis data and the minor axis data are segmented separately to obtain the major axis segmentation result and the minor axis segmentation result; Based on the major axis segmentation results and the minor axis segmentation results, the apex position and basal position of the minor axis are determined using a spatial registration method based on covariance spectral decomposition. Based on the short-axis apical position and the short-axis basal position, redundant scan layers in the short-axis segmentation result are removed to obtain short-axis data segmentation information containing only the heart.

[0006] Secondly, embodiments of this disclosure provide a cardiac image analysis method, including: Obtain the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging; The major axis data and the minor axis data are segmented separately to obtain the major axis segmentation result and the minor axis segmentation result; Based on the major axis segmentation results and the minor axis segmentation results, the apex position and basal position of the minor axis are determined using a spatial registration method based on covariance spectral decomposition. Based on the short-axis apical position and the short-axis basal position, the redundant scan layers in the short-axis segmentation result are removed to obtain short-axis data segmentation information containing only the heart. Analysis results are obtained based on the segmentation information of the short axis data containing only the heart.

[0007] The cardiac image segmentation method disclosed in this embodiment first acquires the long-axis and short-axis data corresponding to the original cardiac magnetic resonance imaging (MRI) image. The long-axis and short-axis data are then segmented to obtain long-axis and short-axis segmentation results, respectively. Based on a spatial registration method using covariance spectral decomposition, the long-axis and short-axis segmentation results are analyzed to determine the short-axis apex and base positions. Based on this, redundant scan layers in the short-axis segmentation results are removed to obtain short-axis data segmentation information containing only the heart. This method models the long-axis left ventricular segmentation results, uses covariance spectral decomposition combined with spatial registration of projection extrema and short-axis slices to achieve precise automatic positioning of the apex and base. Then, based on the short-axis apex and base positions, redundant scan layers in the short-axis segmentation results are removed, resulting in accurate short-axis data segmentation information containing only the heart. The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic flowchart of a cardiac image segmentation method provided in an embodiment of this disclosure.

[0010] Figure 2 This is a flowchart illustrating the method for obtaining long-axis and short-axis data corresponding to raw cardiac magnetic resonance images provided in this embodiment of the disclosure.

[0011] Figure 3A framework diagram of a robust segmentation method based on hard example task enhancement provided in embodiments of this disclosure.

[0012] Figure 4 This is a flowchart illustrating the method for determining the position of the short shaft tip and the position of the short shaft bottom provided in the embodiments of this disclosure.

[0013] Figure 5 This is a schematic flowchart of the cardiac image analysis method provided in the embodiments of this disclosure. Detailed Implementation

[0014] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0015] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0016] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0017] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0018] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0019] Reference Figure 1 This application discloses a method for segmenting cardiac images, including: S100 acquires the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging.

[0020] Among them, the short-axis film sequence and the long-axis sequence are data from the same phase.

[0021] S200, the major axis data and minor axis data are segmented separately to obtain the major axis segmentation result and the minor axis segmentation result.

[0022] S300, based on the major axis segmentation results and the minor axis segmentation results, and using the spatial registration method based on covariance spectral decomposition, determine the position of the minor axis apex and the position of the minor axis base.

[0023] S400, based on the short-axis apex and base positions of the heart, removes redundant scan layers from the short-axis segmentation results to obtain short-axis data segmentation information containing only the heart.

[0024] The cardiac image segmentation method disclosed in this application studies the consistency characteristics of long and short axis data, and proposes a short axis segmentation method based on consistency constraints of cardiac magnetic resonance imaging data. The consistency constraint is used as an important component of the loss function, guiding the network to fully utilize multi-view information during training to achieve accurate segmentation of short axis data. For the problem of apex and base identification, a spatial registration method based on covariance spectral decomposition is proposed, combining the segmentation results of the long and short axes. This method models the long axis left ventricle segmentation results, and through covariance spectral decomposition combined with spatial registration of projection extrema and short axis slices, achieves accurate automatic localization of the apex and base, thereby obtaining precise short axis data segmentation information containing only the heart.

[0025] Reference Figure 2 The method for S100, "acquiring the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging," specifically includes: S110, acquire original cinema magnetic resonance imaging (MRI) long-axis view image data of the same heart as first image data, and original cinema magnetic resonance imaging (MRI) short-axis view image data as second image data.

[0026] Specifically, DICOM format image data (cine-MRI) of the same subject can be exported from the hospital's PACS system / MRI equipment terminal. The long-axis view (first image data) preferably includes a 2 / 3 / 4-chamber long-axis sequence, covering the entire cardiac cycle from systole to diastole (typically 20-30 frames / cycle); the short-axis view (second image data) preferably includes multi-slice short-axis sequences (typically 8-12 slices) from the base to the apex of the heart, also covering the entire cardiac cycle; key metadata such as pixel spacing, slice thickness, TR / TE (repetition / echo time), and cardiac cycle phase markers are preserved.

[0027] S120, resample the first image data and the second image data to unify their spatial resolution to a preset value.

[0028] If no clinical standard is available, the preset value can be the optimal compromise between the original resolution of the long and short axes, such as 1.2 × 1.2 mm for the original long axis and 0.8 × 0.8 mm for the original short axis, with a preset value of 1.0 × 1.0 mm.

[0029] For resampling, the scaling factor between the original resolution and the preset resolution can be calculated based on PixelSpacing (planar pixel spacing) and SliceThickness (layer thickness) in the DICOM header file: Scaling factor = Preset resolution / Original resolution; Resampling is performed using cubic linear interpolation or bilinear interpolation: For 2D images, the width / height pixel count is adjusted according to the scaling factor, and missing pixel values ​​are supplemented by interpolation; For 3D sequences, the resolution in the width / height / layer thickness directions is adjusted simultaneously to maintain the anatomical continuity between layers; After resampling and verification, the PixelSpacing / SliceThickness labels of the output image are updated to the preset values, and the pixel matrix size is adjusted synchronously according to the scaling factor, such as 256×256 originally → 256×256 when the preset is 1.0mm, and 512×512 originally → 256×256 after scaling by 0.5.

[0030] Furthermore, during resampling, zero-padding or mirror padding can be used on the image boundaries to avoid distortion of edge pixel values; the anatomical center position of the image (such as the centroid of the heart) is preserved to ensure spatial alignment of the major and minor axes after resampling.

[0031] S130, the pixel values ​​of the resampled first image data and second image data are truncated to remove abnormal pixel values ​​within a predetermined percentage range.

[0032] Among these, it is preferable to remove the values ​​of the first 1% and the last 1% of the data as outliers.

[0033] Specifically, merge the pixel values of all the long-axis / short-axis images after resampling to construct a global pixel value histogram; calculate the quantiles of the histogram: lower truncation threshold = 1% quantile (P1) of the pixel values, upper truncation threshold = 99% quantile (P99) of the pixel values; perform truncation on each pixel value: if the pixel value < P1, replace it with P1; if the pixel value > P99, replace it with P99; if it is between P1 - P99, keep the original value. Through this step, abnormal pixels commonly found in magnetic resonance images, such as extreme values caused by radio frequency noise, metal artifacts, and motion artifacts, can be removed, avoiding interference from outliers to subsequent normalization and feature extraction.

[0034] S140. Based on all the pixel values after truncation processing, calculate the global mean and variance, and perform normalization processing on the first image data and the second image data based on the global mean and variance to obtain normalized image data.

[0035] Through this step, the signal intensity differences between different subjects / different sequences, such as the signal being too strong / too weak caused by magnetic resonance equipment models, scanning parameters, and subject body types, can be eliminated, making the long-axis / short-axis data on the same numerical scale; the normalized pixel values conform to the input habits of machine learning models (such as CNN), avoiding gradient explosion / vanishing caused by too large numerical ranges; the degree to which the normalized pixel values deviate from 0 can directly reflect their differences from the global mean (such as the myocardial region is usually positive and the background is negative), facilitating subsequent anatomical structure segmentation.

[0036] In the first embodiment, for S200, it specifically includes: 1) Train the first initial segmentation model according to the historical cine magnetic resonance long-axis training data and the first target loss function, and segment the long-axis data according to the pre-trained long-axis segmentation model to obtain the long-axis segmentation result.

[0037] Among them, the long-axis segmentation result is the segmentation mask of the long-axis data, and the segmentation mask labels the pixel ownership of the left ventricle, right ventricle, myocardium, and background in the long-axis data. The cine magnetic resonance long-axis training data can be data from public datasets and hospital private data with segmentation labels.

[0038] Specifically, the first initial segmentation model is preferably nn-UNet; the first target loss function is preferably the sum of the Dice coefficient loss function, boundary loss function, and cross-entropy loss function (i.e., CE loss function) corresponding to the long-axis training data.

[0039] Specifically, the first target loss function is : [[ID=--]]

[0040] Among them,​ The Dice coefficient loss function is the one used for the training data along the major axis. The cross-entropy loss function is the one corresponding to the training data along the long axis. This is the boundary loss function corresponding to the training data along the major axis.

[0041] The Dice coefficient loss function measures the overlap between the segmentation mask output by the segmentation network and the segmentation labels inherent in the training data; the formula for calculating the Dice coefficient loss function is: ;,in, The segmentation mask (i.e., prediction value) for the corresponding category output by the long axis segmentation network. The segmentation labels (i.e., true values) that come with the training data for the long axis are used. For example, segmentation label 0 represents the background, etc.

[0042] Among them, CE loss is used to measure the difference between the class probability output by the long axis segmentation network and the true segmentation label; the cross-entropy loss function (i.e., CE loss function) = background CE loss function + left ventricular CE loss function + myocardial CE loss function + right ventricular loss function.

[0043] The formula for calculating the CE loss function is: , C To segment categories (including: background, left ventricle, right ventricle, myocardium). For the first Predicted probability of pixels corresponding to class segmentation structure For the first The actual pixel value corresponding to the class segmentation structure.

[0044] Among them, "the The method for obtaining the "predicted probability of pixel points corresponding to the class segmentation structure" includes: 1) obtaining the predicted values ​​of the four class labels corresponding to each pixel point according to the long axis segmentation network; 2) taking the largest predicted value as the first predicted value. Predicted probability of pixels corresponding to class segmentation structures.

[0045] The left ventricular CE loss function, myocardial CE loss function, and right ventricular CE loss function are preferably 1-2.5 times the background CE loss function. The boundary loss is preferably calculated using a distance transformation method to constrain the edge accuracy of each cardiac structure in the long axis segmentation result.

[0046] 2) Determine the area proportion of each structure based on the results of the major axis segmentation.

[0047] Specifically, the number of pixels in the left ventricle, right ventricle, and myocardium in the segmentation mask is counted, and the area ratio of each structure is calculated. The area ratio is the ratio of the number of pixels in each structure to the total number of pixels in the left ventricle, right ventricle, and myocardium.

[0048] 3) Determine the consistency constraint loss of major and minor axes based on the area ratio.

[0049] Specifically, based on the characteristics that the short-axis and long-axis data come from the same cardiac motion phase, differ only in imaging direction, and have consistent anatomical structures, since both the long-axis and short-axis sequences come from data of the same cardiac time phase and differ only in imaging direction, they should have high consistency in anatomical structure. In this embodiment, the area ratio corresponding to the long axis is taken as the area ratio corresponding to the short axis (i.e., label / true value). The corresponding long-short axis consistency constraint loss is LV² + RV² + MY², where LV, RV, and MYO are the area ratios of the left ventricle, right ventricle, and myocardium in the long axis, respectively.

[0050] 4) Obtain short-axis training data that matches the long-axis training data of historical film magnetic resonance imaging.

[0051] Specifically, the short-axis training data and the long-axis data from cinema magnetic resonance imaging belong to the same patient and the same cardiac motion phase. The short-axis training data has segmentation labels, which are used to label the left ventricle, right ventricle, myocardium, and background structures in the short-axis training data.

[0052] 5) Construct a second objective loss function based on the long and short axis consistency constraint loss and the first objective loss function.

[0053] The second objective loss function is the sum of the Dice coefficient loss function, boundary loss function, cross-entropy loss function, and major-minor axis consistency constraint loss corresponding to the minor axis training data.

[0054] Specifically, the second objective loss function is: : .

[0055] in, The Dice coefficient loss function is the one used for the training data with the minor axis. The cross-entropy loss function is the one used for the training data with the shorter axis. The boundary loss function is the one corresponding to the training data with the minor axis. The loss is due to the consistency constraint of the major and minor axes.

[0056] The Dice coefficient loss function measures the overlap between the segmentation mask output by the segmentation network and the segmentation labels inherent in the training data; the formula for calculating the Dice coefficient loss function is: ;in, The segmentation mask (i.e., prediction value) for the corresponding category output by the short-axis segmentation network. The segmentation labels (i.e., true values) that come with the training data for the short axis are used. For example, segmentation label 0 represents the background, etc.

[0057] Among them, CE loss is used to measure the difference between the class probability output by the short-axis segmentation network and the true segmentation label; the cross-entropy loss function (i.e., CE loss function) = background CE loss function + left ventricular CE loss function + myocardial CE loss function + right ventricular loss function.

[0058] The formula for calculating the CE loss function is: , C To segment categories (including: background, left ventricle, right ventricle, myocardium). For the first Predicted probability of pixels corresponding to class segmentation structure For the first The actual pixel value corresponding to the class segmentation structure.

[0059] Among them, "the The method for obtaining the "predicted probability of pixel points corresponding to the class segmentation structure" includes: 1) obtaining the predicted values ​​of the four class labels corresponding to each pixel point according to the short-axis segmentation network; 2) taking the largest predicted value as the first predicted value. Predicted probability of pixels corresponding to class segmentation structures.

[0060] Among them, the left ventricular CE loss function, myocardial CE loss function, and right ventricular loss function are preferably 1-2.5 times the background CE loss function.

[0061] The boundary loss is preferably calculated using a distance transformation method to constrain the edge accuracy of each cardiac structure in the short-axis segmentation results.

[0062] 6) Train the second initial segmentation model based on the minor axis training data and the second objective loss function, and segment the minor axis data according to the pre-trained minor axis segmentation model to obtain the minor axis segmentation result.

[0063] The second initial segmentation model is preferably a 3D UNETR minor axis segmentation network. Specifically, the minor axis training data is input into the constructed 3D UNETR minor axis segmentation network, the segmentation mask is output, the second objective loss function value corresponding to the segmentation mask is calculated, the parameters of the minor axis segmentation network are optimized by backpropagation through the gradient descent algorithm, and the above training process is repeated until the second objective loss function value tends to stabilize and the network converges.

[0064] Furthermore, it also includes: performing Spacing unification on the long-axis and short-axis training data, removing outliers, calculating the global mean, variance, and Z-scores normalization. Specifically, Spacing unification involves using a medical image processing library to interpolate all short-axis and long-axis training data, unifying the pixel physical space size of all data to a preset fixed value. Outlier removal specifically includes: statistically analyzing the pixel value distribution of all short-axis and long-axis training data, setting a reasonable threshold for pixel values, replacing outlier pixel values ​​exceeding the threshold with the threshold or the data median, and removing maxima and minima caused by scanning noise. Z-scores normalization involves calculating the global mean and global variance of all short-axis and long-axis training data, performing a normalization operation on the pixel value of each pixel, so that all data distributions conform to a standard normal distribution.

[0065] In another embodiment, S200 uses a robust segmentation method based on hard example task enhancement (add-KAN) for segmentation, specifically including: S210, constructing an initial teacher model and an initial student model.

[0066] The initial teacher model construction uses 3D U-Net++ / nnU-Net (the state-of-the-art model for medical image segmentation) as the basic architecture. The input is a medical image tensor (e.g., major / minor axis data from CT / MRI, dimensions: [B, C, D, H, W], where B = batch, C = channels, D = depth, H / W = height / width), and the output is a segmentation mask (with dimensions consistent with the input space, pixel-level classification). A lightweight 3D U-Net is chosen to reduce the number of channels / residual units and to reserve a KAN embedding interface; the feature layer is followed by a KAN layer.

[0067] S220: The initial teacher model is trained using the first target raw data. The target information in the trained initial teacher model is then transferred to the initial student model for initialization through the KAN network for interpretability. Feature distillation and output distillation are then performed to obtain the first target student model.

[0068] The teacher model uses a high-performance, complex segmentation network (to ensure segmentation accuracy), while the student model uses a lightweight network (to ensure deployment efficiency). Both are designed with medical image segmentation as the goal and have the same output dimension (to facilitate subsequent distillation). KAN (Kolmogorov-Arnold Network) is embedded in the model feature layer as the core component for interpretable transfer to achieve knowledge transfer.

[0069] Specifically, a difficult segmentation structure (myocardial ring) is selected to first train the teacher model to fit the original data (long axis data) of the first target. The loss function can be DiceLoss + CrossEntropyLoss (a classic combination for medical segmentation). The optimizer is preferably AdamW, with a learning rate of 1e-4 and a weight decay of 1e-5. The training strategy includes: batch size of 4, epoch of 100, and early stopping (stop if val_loss does not decrease for 10 consecutive epochs).

[0070] Next, key features / weights of the teacher model are extracted using KAN (for interpretable transfer) to initialize the student model. Finally, the student model is optimized through feature distillation and output distillation. The purpose of this step is to extract the feature mapping patterns of the teacher model (fitted by KAN) and use them as the initial weights of the student model. Specifically, the feature outputs of the last three layers of the teacher model (denoted as T) are taken. features ), calculate the weight distribution and activation pattern (interpretability index) of features; learn T using a KAN network. features The mapping to the initial features S_init_features of the student model makes the output of KAN correlated with T. features The error is minimized (MSELoss); the mapping parameters (weights, grid nodes) learned by KAN are assigned to the kan_layer of the student model, and the output layer of the teacher model is initialized by scaling the output layer weights of the teacher model proportionally.

[0071] The goal of distillation is to make the features / output of the student model approximate those of the teacher model, ensuring the model's segmentation accuracy for the left and right ventricles while enhancing its segmentation accuracy for myocardial rings. Feature distillation involves minimizing the MSE (Minimum Separation Equation) between student and teacher features (focusing on key layer features). Output distillation employs a SoftTarget loss (KL divergence between student and teacher soft outputs) with a temperature coefficient T=2, plus a hard label loss (DiceLoss between student output and ground truth labels); the optimal total loss is Loss = 0.3. Characteristic distillation loss +0.5 Output distillation loss +0.2 Hard label loss.

[0072] S240, based on the first target student model, the long axis data is segmented to obtain the long axis segmentation result.

[0073] Specifically, the pre-trained first target student model is loaded, forward inference is performed on the long axis data, the segmentation mask is output, and post-processing optimization is performed on the results.

[0074] S250: The initial teacher model is trained using the original data of the second target. The target information in the trained initial teacher model is then transferred to the initial student model for initialization through the KAN network for interpretability. Feature distillation and output distillation are then performed to obtain the second target student model.

[0075] The logic is the same as S220, but the target data is replaced with the original data of the second target (minor axis data). The spatial dimension of the minor axis data needs to be adapted (e.g., if the depth D of the minor axis data is smaller, the input dimension of the model needs to be adjusted).

[0076] S260, based on the second objective student model, the minor axis data is segmented to obtain the minor axis segmentation result.

[0077] Consistent with the logic of S240, load the second target student model, adapt it to the spatial dimension of the short axis data, and perform segmentation and post-processing.

[0078] In this embodiment, KAN can transform the implicit knowledge (features / weights) of the teacher model into an interpretable piecewise function, achieving white-box knowledge transfer rather than traditional black-box distillation. It should be noted that the short-axis data space dimension differs from the long-axis dimension, requiring adjustment of the model input dimension and post-processing parameters (such as connected component area) to ensure segmentation accuracy.

[0079] Among them, the KAN layer structure has: formally, d in Input dimension d out A Kolmogorov-Arnold layer with n output dimensions can be represented as: In practical applications, It is composed of a linear combination of the SiLU activation function and the B-spline function: Where G represents the interval The number of grid nodes used in the uniform division; k represents the order of the spline; Represents a linear weight vector; This represents the output weight vector of the B-spline basis function. It does not contribute to the model's output by activation or inhibition, so it can be pruned to leave only the positions that contribute more.

[0080] specifically refer to Figure 3This approach aims to train a complex teacher model by reducing the workload. The key knowledge learned by the teacher model is then transferred (added) to the student model across all tasks, and knowledge distillation is performed on the intermediate feature layers and the output. The steps are as follows: First, evaluate the most important knowledge representation nodes in the teacher model; second, project these important nodes onto row vectors; third, select the corresponding columns from the input features; fourth, transfer the weight parameters of these important nodes to the student model; fifth, perform local feature distillation based on the features selected in step three and the corresponding input features in the student model; sixth, perform output distillation. In this embodiment, a simple yet practical subtask-enhanced KAN knowledge self-distillation framework is proposed, leveraging the better domain-invariant representation capabilities of complex models and the interpretability of KAN. Since training a complex model is difficult due to the limited number of samples, the method proposed in this embodiment achieves the goal of training a complex teacher model by reducing the model's workload. Then, the important knowledge selected from the sub-task teacher model is transferred (added) to the full-task student model through the interpretable features of the KAN network for initialization and feature distillation and output distillation. This allows the full-task student segmentation model to learn both the common knowledge of the full task and the unique knowledge of the domain-invariant representation ability in the sub-task teacher model.

[0081] In this embodiment, a robust segmentation method based on hard-example enhancement (add-KAN) is proposed to enhance the segmentation of difficult-example structures. This method first trains an accurate teacher model using hard-example structures, and then adds the important knowledge learned by the teacher model to the student model through knowledge distillation, thereby improving the student model's ability to segment hard-example structures and achieving accuracy and robustness in whole-heart segmentation.

[0082] Reference Figure 4 The methods for determining the position of the short shaft tip and the position of the short shaft base of the S300 specifically include: S310: Based on the long axis segmentation results, obtain the physical coordinates of the left ventricle along the long axis.

[0083] Specifically, the left ventricle is determined in the long axis segmentation result; the metadata affine matrix corresponding to the left ventricle is obtained; and the pixel coordinates corresponding to the left ventricle are converted into physical coordinates in real space based on the metadata affine matrix to obtain the long axis left ventricle physical coordinates. This step involves converting the long axis segmentation pixel coordinates into the long axis left ventricle physical coordinates in real space based on the left ventricle in the long axis segmentation result and the metadata affine matrix corresponding to the long axis segmentation result.

[0084] The core source of the left ventricle's affine matrix is ​​the global affine matrix of the medical image in which it is located. It can be directly extracted using the img.affine file from the nibabel library. The image file is in the standard NIfTI format.

[0085] S320 performs spectral decomposition of the covariance matrix of the left ventricular physical coordinates along the major axis to obtain feature data, which includes eigenvectors and eigenvalues.

[0086] This step specifically includes: S321, obtaining the average value of all long-axis left ventricular physical coordinates.

[0087] ,in, This represents the total number of physical coordinates of the left ventricle along the major axis. Let i be the physical coordinates of the left ventricle on the i-th major axis, which is a three-dimensional vector. , , ), where x, y, and z represent the x, y, and z coordinates of the point in the world coordinate system, respectively. The final mean vector represents the average center position of all left ventricular coordinate points, in the form of ( , , ).

[0088] S322, perform mean-removal processing based on the average value to obtain a new data matrix.

[0089] The new data matrix is ; ;in, This is the original left ventricular coordinate data matrix, with dimensions N×3, where each row represents a coordinate point. , , ). Let N be a column vector of length N, consisting entirely of 1s, i.e., 1 = [1, 1, ..., 1]. ⊤ . Mean vector The transpose of is used to transform it from a column vector into a row vector with a dimension of 1×D. Given an N×D matrix, each row of which is identical, representing the mean vector. This formula calculates the coordinates of each point in the original data matrix world_lax_seg. , , Subtract the average center coordinates from all of them. , , ), to obtain new coordinates ( , , ), i.e., new coordinates The center of the coordinate system has moved to the origin (0,0,0).

[0090] S323: Obtain the covariance matrix based on the mean and the new data matrix.

[0091] The covariance matrix is: .

[0092] S324, perform spectral decomposition on the covariance matrix to obtain the characteristic data.

[0093] ,in, It is a diagonal matrix, and the values ​​on the diagonal are eigenvalues; The column vectors of U are the eigenvectors. Specifically, Λ is a 3×3 diagonal matrix with three eigenvalues ​​λ1≥λ2≥λ3≥0 on the diagonal. The columns of U are the corresponding eigenvectors u1, u2, u3, which are all 3×1 unit vectors and are pairwise orthogonal.

[0094] S330 projects the feature data according to the principal components to obtain the physical coordinates of the long axis apex and the long axis base.

[0095] Wherein, the physical coordinates of the apex of the major axis represent the maximum value in the projection direction, and the physical coordinates of the base of the major axis represent the minimum value in the projection direction. In this embodiment, K-th principal component projection is preferred.

[0096] This step specifically includes: 1) Constructing the projection vector: The eigenvector corresponding to the largest eigenvalue... As the direction of the heart's long axis (i.e., the direction of the first principal component); 2) Point-by-point projection calculation: for each feature point Calculate its projection along the major axis using the formula: This yields the set of projected values ​​for all feature points. , ,..., 3) Locating the extreme value index: Find the maximum value in the set of projected values. Record its index : Find the minimum value in the set of projected values. Record its index : 4) Mapping back to physical coordinates: Major axis apex physical coordinates: Take the index from the original feature data. Point: The coordinates of the apex are Long axis center physical coordinates: taken from the original feature data with the index as... Point: Coordinates of the heart are .

[0097] S340: Based on the short axis segmentation results, obtain the short axis physical coordinates of the left ventricle.

[0098] Specifically, the metadata affine matrix corresponding to the left ventricular region in the short axis segmentation result is obtained; based on the metadata affine matrix, the pixel coordinates corresponding to the left ventricular region in the short axis segmentation result are converted into physical coordinates in real space to obtain the physical coordinates of the left ventricular region in the short axis.

[0099] S350 spatially registers the short-axis left ventricular physical coordinates with the long-axis apical and base physical coordinates to obtain the short-axis apical and base positions.

[0100] Specifically, the coordinate closest to the long axis apical physical coordinate is determined from all short-axis left ventricular physical coordinates, and this coordinate is taken as the short-axis apical position; the coordinate closest to the long axis base physical coordinate is determined from all short-axis left ventricular physical coordinates, and this coordinate is taken as the short-axis base position.

[0101] In this embodiment, the affine matrix is ​​preferably a 4x4 transformation matrix, serving as a bridge connecting pixel coordinates and physical coordinates. It is typically stored in the metadata of medical image files (such as DICOM). This matrix encodes the following key spatial information: 1) Scaling: how many millimeters each pixel actually represents in the X, Y, and Z directions (i.e., pixel pitch and slice thickness); 2) Rotation: the orientation of the image slice relative to the patient's body; 3) Translation: the starting position of the first pixel (0,0,0) in physical space. The physical coordinates are the Affine matrix multiplied by the homogenized pixel coordinates.

[0102] Assuming the affine moments of a CT image show each pixel representing 0.5 mm horizontally and vertically, and the image center point corresponds to the point (100, 150, -200) mm in the physical coordinate system, the current segmentation result indicates that the apex of the left ventricle is located at the image pixel coordinates (256, 300). By applying the affine matrix for calculation, the physical coordinates of this point can be obtained as approximately (228.0, 0.0, -200.0) mm. This (228, 0, -200) is the actual three-dimensional position of the patient's left ventricular apex in the scanner coordinate system. This embodiment utilizes the spatial positioning information (affine matrix) inherent in the image to convert the left ventricular pixel position found by the segmentation software into a three-dimensional position with actual length and orientation in the real world.

[0103] The S400 method, which "removes redundant scan layers from the short-axis segmentation results based on the short-axis apex and base positions to obtain short-axis data segmentation information containing only the heart," specifically includes: based on the apex and base positions of the short-axis sequence, performing layer filtering on the original short-axis image and corresponding segmentation results, removing redundant scan layers above the apex and below the base that do not contain effective heart structures, and retaining only the short-axis layers containing complete heart structures to obtain pure heart region segmentation data for subsequent quantitative analysis of cardiac function.

[0104] Based on the obtained precise short-axis apex and base positions of the heart, this step retains only the layers from the apex to the base of the heart in the short-axis cardiac MRI sequence, while deleting the redundant scan layers above and below, those without the heart, or those with only a small portion of the heart, leaving only the segmentation results of the pure heart region. This process is simple, efficient, and accurate.

[0105] Reference Figure 5 Secondly, this application discloses a cardiac image analysis method, comprising: S10: Acquire the long axis data and short axis data corresponding to the original cardiac magnetic resonance imaging; S210, segment the major axis data and minor axis data respectively to obtain the major axis segmentation result and the minor axis segmentation result; S30. Based on the major axis segmentation results and the minor axis segmentation results, the spatial registration method based on covariance spectrum decomposition is used to determine the position of the minor axis apex and the position of the minor axis base. S40, based on the short-axis apex and base positions of the heart, removes redundant scan layers from the short-axis segmentation results to obtain short-axis data segmentation information containing only the heart. S50, obtain analysis results based on the segmentation information containing only the short axis data of the heart. Specifically, the analysis results include: calculating cardiac function parameters, which preferably include ventricular volume, ejection fraction, stroke volume, interventricular septum thickness, and myocardial mass.

[0106] Specifically, Simpson's method is used to calculate the volumes of the left ventricle, right ventricle, and myocardium at end-diastole and end-systole in the short-axis segmentation results. The end-diastole and end-systole periods are determined according to a 2:1 ratio based on the time dimension of the short-axis data. For example, in a 30-frame timeframe, the end-diastole can be taken as frame 0 or 30, and the end-systole as frame 10. Stroke volume is calculated from (end-diastole ventricular volume - end-systole ventricular volume); ejection fraction is calculated from (stroke volume / end-diastole ventricular volume); and the ejection fraction is calculated from (end-diastole myocardial volume). 1.05) The myocardial mass is calculated; the interventricular septum thickness is measured from the width of the myocardium adjacent to the left and right ventricles.

[0107] The method disclosed in this application can construct a fully automatic method for calculating cardiac function parameters based on joint segmentation and localization of cinema magnetic resonance imaging (MRI) long and short axis data. It is convenient to use, has high computational efficiency, and high accuracy.

[0108] It should be noted that the specific steps of S10-S30 are consistent with the specific steps of the cardiac image segmentation method disclosed in the first aspect of this application, so they will not be described in detail here.

[0109] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0110] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0111] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0112] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0113] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0114] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0115] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for segmenting cardiac images, characterized in that, include: Obtain the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging; The major axis data and the minor axis data are segmented separately to obtain the major axis segmentation result and the minor axis segmentation result; Based on the major axis segmentation results and the minor axis segmentation results, the apex position and basal position of the minor axis are determined using a spatial registration method based on covariance spectral decomposition. Based on the short-axis apical position and the short-axis basal position, redundant scan layers in the short-axis segmentation result are removed to obtain short-axis data segmentation information containing only the heart.

2. The cardiac image segmentation method according to claim 1, characterized in that, The step of determining the apex and basal positions of the minor axis based on the major axis segmentation results and the minor axis segmentation results, using a spatial registration method based on covariance spectral decomposition, includes: Based on the major axis segmentation results, the physical coordinates of the left ventricle along the major axis are obtained; The covariance matrix of the physical coordinates of the left ventricle along the major axis is decomposed spectrally to obtain feature data, which includes eigenvectors and eigenvalues. The feature data is projected according to the principal components to obtain the physical coordinates of the apex and base of the major axis; the physical coordinates of the apex of the major axis are the maximum values ​​in the projection direction, and the physical coordinates of the base of the major axis are the minimum values ​​in the projection direction. Based on the short axis segmentation results, the physical coordinates of the left ventricle along the short axis are obtained; Spatially register the short-axis left ventricular physical coordinates with the long-axis apical physical coordinates and the long-axis base physical coordinates to obtain the short-axis apical position and the short-axis base position.

3. The cardiac image segmentation method according to claim 1, characterized in that, The step of obtaining the left ventricular physical coordinates based on the major axis segmentation results includes: Determine the left ventricle location in the long axis segmentation results; Obtain the metadata affine matrix corresponding to the left ventricular region; Based on the metadata affine matrix, the pixel coordinates corresponding to the left ventricular region are converted into physical coordinates in real space to obtain the long axis left ventricular physical coordinates.

4. The cardiac image segmentation method according to claim 2, characterized in that, The spectral decomposition of the covariance matrix of the long-axis left ventricular physical coordinates yields feature data, including: Obtain the average value of all the aforementioned long-axis left ventricular physical coordinates; A new data matrix is ​​obtained by removing the mean from the average value. Based on the average value and the new data matrix, the covariance matrix is ​​obtained; The covariance matrix is ​​subjected to spectral decomposition to obtain the feature data.

5. The cardiac image segmentation method according to claim 2, characterized in that, The step of projecting the feature data according to principal components to obtain the physical coordinates of the apex and base of the major axis includes: The direction of the heart's long axis is determined based on the aforementioned feature data; Calculate the projection value of each feature point along the long axis of the heart to obtain the set of projection values ​​for all feature points; Identify the first index information and the second index information corresponding to the maximum and minimum values, respectively, from the set of projected values; Based on the mapping of the first index information and the second index information back to physical coordinates, the points in the feature data corresponding to the first index information are used as the physical coordinates of the apex of the major axis, and the points in the feature data corresponding to the second index information are used as the physical coordinates of the base of the major axis.

6. The cardiac image segmentation method according to claim 2, characterized in that, The step of spatially registering the short-axis left ventricular physical coordinates with the long-axis apical and base physical coordinates to obtain the short-axis apical and base positions includes: Determine the coordinate closest to the long axis apical physical coordinate from all the short axis left ventricular physical coordinates, and use it as the short axis apical position; The coordinate closest to the long axis basal body physical coordinates from all the short axis left ventricular physical coordinates is determined as the short axis basal body position.

7. The cardiac image segmentation method according to claim 1, characterized in that, The acquisition of the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging includes: The original cinema magnetic resonance imaging (MRI) long-axis view image data of the same heart was acquired as the first image data, and the original cinema magnetic resonance imaging (MRI) short-axis view image data was acquired as the second image data. The first image data and the second image data are resampled to unify their spatial resolution to a preset value; The pixel values ​​of the resampled first and second image data are truncated to remove abnormal pixel values ​​within a predetermined percentage range. Based on all pixel values ​​after truncation, the global mean and variance are calculated, and the first and second image data are standardized based on the global mean and variance to obtain normalized image data.

8. The cardiac image segmentation method according to claim 1, characterized in that, The step of segmenting the major axis data and the minor axis data respectively to obtain the major axis segmentation result and the minor axis segmentation result includes: Construct initial teacher and student models; The initial teacher model is trained using the first target raw data. The target information in the trained initial teacher model is then interpreted and transferred to the initial student model through the KAN network for initialization. Feature distillation and output distillation are then performed to obtain the first target student model. The long axis data is segmented based on the first target student model to obtain the long axis segmentation result; The initial teacher model is trained using the original data of the second target. The target information in the trained initial teacher model is then interpreted and transferred to the initial student model through the KAN network for initialization. Feature distillation and output distillation are then performed to obtain the second target student model. The minor axis data is segmented based on the second target student model to obtain the minor axis segmentation result.

9. The cardiac image segmentation method according to claim 1, characterized in that, The step of segmenting the major axis data and the minor axis data respectively to obtain the major axis segmentation result and the minor axis segmentation result includes: The first initial segmentation model is trained based on the long axis training data of historical film magnetic resonance imaging and the first objective loss function. The long axis data is then segmented according to the pre-trained long axis segmentation model to obtain the long axis segmentation result. Based on the major axis segmentation results, determine the area proportion of each structure; The major and minor axis consistency constraint loss is determined based on the area ratio. Obtain short-axis training data that matches the long-axis training data of the historical film magnetic resonance imaging; A second objective loss function is constructed based on the major and minor axis consistency constraint loss and the first objective loss function. The second initial segmentation model is trained based on the minor axis training data and the second objective loss function. The minor axis data is then segmented according to the pre-trained minor axis segmentation model to obtain the minor axis segmentation result.

10. A method for analyzing cardiac images, characterized in that, include: Obtain the long axis and short axis data corresponding to the original cardiac magnetic resonance imaging; The major axis data and the minor axis data are segmented separately to obtain the major axis segmentation result and the minor axis segmentation result; Based on the major axis segmentation results and the minor axis segmentation results, the apex position and basal position of the minor axis are determined using a spatial registration method based on covariance spectral decomposition. Based on the short-axis apical position and the short-axis basal position, the redundant scan layers in the short-axis segmentation result are removed to obtain short-axis data segmentation information containing only the heart. Analysis results are obtained based on the segmentation information of the short axis data containing only the heart.

Images