Heart ablation planning method and system based on multi-modal data fusion
Through multimodal data fusion technology, the automation, quantitative prediction and automatic optimization of cardiac ablation targets are achieved, which solves the problems of incomplete target identification, large positioning errors and low planning efficiency in existing technologies, and improves the accuracy and safety of ablation planning.
Patent Information
- Application Number
- CN202510833545.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-09-23
AI Technical Summary
In existing cardiac ablation technology, insufficient multimodal data fusion leads to incomplete target identification, large positioning errors, low planning efficiency, insufficient safety and repeatability, lack of dynamic coupling analysis of cardiac motion and electrical activity, no unified energy model for path planning, reliance on operator experience, and difficulty in forming a standardized path.
Through multimodal data fusion, intracardiac electrophysiological point clouds, dynamic images, and preoperative heart sound signals are collected and uniformly mapped to a standard anatomical coordinate system. Dynamic features are extracted and cross-modal attention fusion is performed. An ablation target probability map and a path energy optimization model are constructed. Anatomical boundary constraints are set, and the ablation point sequence is automatically calculated.
It realizes the automated prediction of target probability and the automatic calculation of the optimal ablation point sequence, improves planning efficiency and repeatability, reduces positioning errors, enhances the feasibility and safety of the path, and is suitable for the precise identification of lesions in complex cardiac chambers.
Smart Images
Figure CN120690380A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ablation planning, and in particular to a cardiac ablation planning method and system based on multimodal data fusion. Background Art
[0002] With the increasing prevalence of arrhythmias and atrial fibrillation, cardiac catheter ablation has become an important clinical treatment. A typical ablation procedure requires the surgeon to manually or semi-automatically determine a series of ablation targets and plan the ablation path based on three-dimensional anatomical images (such as CT and MRI), combined with intraoperative electrophysiological (EP) point clouds and cardiac motion information. Currently, related technologies focus on the following aspects:
[0003] (1) Single modality imaging and electrophysiological mapping. CT / MRI anatomical images can obtain high-resolution structural information of the cardiac chambers and walls, but lack functional dynamic data and are only static references before surgery. Electrophysiological point cloud mapping systems (such as 3D electrophysiological navigation systems) can acquire intracardiac voltage points in real time, but they are usually operated in two-dimensional projections or simplified three-dimensional models. Registration with structural images requires manual adjustment, which may result in registration errors. Quantitative analysis of cardiac chamber motion: Some studies have performed myocardial motion calibration based on Cine-MRI, but this is only used to evaluate the overall motion function of the heart and has not yet been deeply integrated with ablation target planning.
[0004] (2) Multimodal fusion attempts. Some scientific research or commercial systems attempt to superimpose CT anatomy and EP point clouds through rigid or elastic registration, but most of them are static registrations, ignoring the dynamic deformation of the heart during the cardiac cycle, resulting in inaccurate electrophysiological signal mapping positions. A small number of studies have introduced Cine-MRI dynamic information, but they are often only qualitatively visualized and lack quantitative fusion algorithms and automated processes, making it difficult to support efficient and accurate target generation.
[0005] (3) Path planning and energy optimization. Traditional path planning relies heavily on operator experience or simple geometric shortest path algorithms, ignoring the physical kinematic characteristics of the catheter and the differences in cardiac tissue stability. It also lacks a comprehensive quantitative model for multiple factors such as path energy consumption, position stability, and surgical efficiency, resulting in planned paths that are often tortuous and lengthy or with uneven target distribution.
[0006] The main deficiencies of the existing technology include the following aspects:
[0007] (1) Functional deficiencies. Most systems only provide static anatomy and electrophysiological superposition, lacking dynamic coupling analysis of cardiac motion and electrical activity; they are unable to utilize auxiliary information such as heart sounds and myocardial stress, resulting in incomplete identification of the lesion area.
[0008] (2) Insufficient accuracy. Rigid or rough registration methods cannot accurately align the electrophysiological point cloud with the dynamic changes of the cardiac chamber, resulting in positioning errors of several millimeters. Target probability prediction is based only on a single modality and is difficult to reflect the comprehensive risk distribution of multi-source information.
[0009] (3) Low efficiency and long time consumption. Data integration and manual target selection consume a lot of preoperative preparation and intraoperative operation time; path planning is not supported by a unified energy model, and the catheter trajectory needs to be repeatedly adjusted, making the overall process cumbersome.
[0010] (4) Insufficient safety and repeatability. There is a lack of a systematic anatomical boundary protection mechanism, and the surgeon needs to manually set multiple safety parameters. The planning results rely on the surgeon's experience, making it difficult to form a standardized and repeatable optimal path. Summary of the Invention
[0011] In view of this, the purpose of the embodiments of the present invention is to provide a cardiac ablation planning method and system based on multimodal data fusion, which can realize automated and quantified target probability prediction, automatically calculate the optimal ablation point sequence, and greatly improve planning efficiency and repeatability.
[0012] In a first aspect, an embodiment of the present invention provides a cardiac ablation planning method based on multimodal data fusion, which includes:
[0013] Multi-source heterogeneous cardiac function data are collected and uniformly mapped to the standard anatomical coordinate system to obtain the registered cardiac function data.
[0014] Dynamic features are extracted from the registered cardiac function data.
[0015] The dynamic features are fused with cross-modal attention to obtain a fused feature map.
[0016] The ablation possibility of each voxel is predicted based on the fused feature map to obtain an ablation target probability map generation model.
[0017] Anatomical boundary constraints are set for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range.
[0018] An ablation path energy optimization model is constructed based on the safe ablation target probability map, and an ordered ablation point sequence is calculated through a path planning algorithm.
[0019] In combination with the first aspect, an embodiment of the present invention provides a first possible implementation of the first aspect, wherein the collecting multi-source heterogeneous cardiac function data and uniformly mapping them to a standard anatomical coordinate system to obtain registered cardiac function data includes:
[0020] Collect intracardiac electrophysiological point cloud data ,in, For the The three-dimensional coordinates of the electrodes in the standard anatomical coordinate system, For the corresponding time The voltage amplitude is the number of electrophysiological points, is the time frame index.
[0021] Acquire dynamic cardiac image sequences , indicating that each frame size is height ,width , depth layers 3D voxel data.
[0022] Acquisition of static cardiac CT volume data ,in, is the image height, is the image width, is the number of image depth layers.
[0023] Collect preoperative heart sound signals ,in, Indicates that the heart sound signal is of length One-dimensional time series.
[0024] Calculate the registration objective function ,in, is the rigid body transformation matrix of the electrophysiological point cloud, Indicates converting the point cloud into the corresponding spatial surface or shape expression. Indicates the extraction of endocardial deformation contours in dynamic images. Indicates applying the matrix to point cloud data.
[0025] The static cardiac CT volume data The coordinate system is the standard anatomical coordinate system.
[0026] The electrophysiological point cloud data Mapping to the standard anatomical coordinate system to obtain time-series registered electrophysiological data .
[0027] The cardiac dynamic image sequence The images were acquired under the standard anatomical coordinate system.
[0028] The preoperative heart sound signal The timestamps of the dirty dynamic image sequence The timestamps are aligned to obtain aligned heart sound signals .
[0029] Its technical effect is: by uniformly mapping electrophysiological point clouds, dynamic MRI, preoperative heart sounds and CT anatomical images into a standard three-dimensional coordinate system, it not only eliminates the spatial offset and temporal drift between different data sources, but also for the first time accurately aligns the timing information of heart sounds with electrical, biological and imaging data within the same anatomical framework, allowing subsequent feature extraction to synchronously consider electrical activity, mechanical movement and acoustic signals under a completely consistent anatomical reference.
[0030] The embodiment of the invention provides a second possible implementation of the first aspect, wherein extracting dynamic features from the registered cardiac function data includes:
[0031] Based on the temporal registration of electrophysiological data , calculate each pair of electrode points Local conduction velocity characteristics and the potential change sequence characteristics representing the potential waveform of each electrode , modeled by Transformer encoder to obtain electrophysiological features ,in, for point The three-dimensional coordinates of and is the sampling time of the corresponding point, A pair of electrode points The transmission speed, is the characteristic mapping function used to combine conduction velocity and potential information, Modeling temporal sequences with multi-head attention networks.
[0032] The cardiac dynamic image sequence Perform inter-frame difference to obtain temporal motion features , extracting cardiac motion features through 3D convolutional modeling ,in, It is a three-dimensional convolutional network.
[0033] The static cardiac CT volume data Perform 3D segmentation and extract anatomical structure features ,in, is a convolution-based medical image segmentation model. is the weight matrix of the endocardial region, It is a pixel-by-pixel multiplication operation.
[0034] The preoperative heart sound signal Perform bispectral analysis to generate a two-dimensional bispectral diagram of heart sounds The bispectral graph of heart sounds is input into a neural network based on partitioned attention and lightweight convolution for high-dimensional encoding to extract the deep features of the bispectral graph of heart sounds. ,in, Indicates a time point The bispectral diagram of the heart sound at is the second-order spectrum calculation method, It is a module that combines partitioned attention mechanism with lightweight feature enhancement.
[0035] Its technical effect is: by organically combining the timing dependence of conduction speed and potential waveform, the three-dimensional dynamic changes of motor motion, the fine anatomical structure segmentation and the nonlinear coupling characteristics of heart sounds, it not only realizes the deep and multi-scale encoding of four-dimensional information of electricity, biology, machine and sound, but also leverages the respective advantages of Transformer, 3D-CNN, U-Net and PANet-FusionGhost to effectively capture the multi-level interaction of cardiac function and structure, and significantly improves the accuracy and robustness of ablation target positioning.
[0036] In combination with the first aspect, an embodiment of the present invention provides a third possible implementation of the first aspect, wherein performing cross-modal attention fusion on the dynamic features to obtain a fused feature map includes:
[0037] For each time frame , the attention weights of each modality feature are adaptively calculated by the neural network to obtain the dynamic weights of the electrophysiological features , dynamic weight of cardiac motion characteristics , dynamic weight of deep features of heart sound bispectrum ,in, is the Sigmoid activation function, MLP is the multi-layer perceptron, is global average pooling.
[0038] Weighted fusion of multimodal features to obtain a unified fusion feature map ,in, is the total number of time frames, 、 、 represents the voxel dimension in three-dimensional space, Indicates the number of feature channels.
[0039] Its technical effects are: introducing a dynamic attention mechanism at the time frame level, using neural networks to adaptively calculate the weight distribution of modalities such as electrophysiology, motion and heart sounds at different time points, breaking through the limitation of traditional static fusion that cannot respond to changes in the cardiac cycle, and achieving accurate modeling of the dynamic evolution of the importance of multimodal features over time, so that the fused feature map not only has the ability to collaboratively express multi-source information, but also has high sensitivity to the rhythmicity of lesions and the dynamics of functional disorders, thereby providing a more discriminative and clinically relevant high-dimensional semantic representation for target probability prediction.
[0040] In combination with the first aspect, an embodiment of the present invention provides a fourth possible implementation of the first aspect, wherein predicting the ablation possibility of each voxel based on the fused feature map to obtain an ablation target probability map generation model includes:
[0041] In the fused feature map Based on this, a feature expansion layer is constructed to enhance spatial continuity. ,in, represents the dilated convolutional layer, is the number of channels after expansion.
[0042] The feature expansion layer Perform three-dimensional convolution mapping to obtain a three-dimensional feature expansion layer ,in, is a three-dimensional convolution operator.
[0043] The three-dimensional feature expansion layer Perform Softmax normalization calculation to obtain the ablation target probability map generation model representing the original ablation probability ,in, is an exponential function, Voxel The original ablation score.
[0044] Its technical effect is: by introducing a feature expansion layer to expand the receptive field, the spatial continuity expression ability of the fusion feature is enhanced without increasing the calculation parameters, and the voxel-level features are structurally sensitive mapped by combining three-dimensional convolution, and the probability distribution is obtained through Softmax normalization, thereby achieving high-resolution, context-aware modeling of the spatial distribution of ablation targets, significantly improving the spatial consistency and positioning accuracy of the prediction results, and is particularly suitable for determining areas with blurred lesion boundaries or sparse local signals in complex cardiac chambers.
[0045] In combination with the first aspect, an embodiment of the present invention provides a fifth possible implementation of the first aspect, wherein setting anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range includes:
[0046] Calculate the minimum physical distance from each voxel to the anatomical boundary ,in, is the set of endocardial boundary voxels, is the voxel side length.
[0047] Create an indicator function .
[0048] A safe ablation target probability map representing the probability of target distribution within a safe range is obtained by calculating the indicator function. , wherein the safe ablation target probability map , which means that only the ablation probability that meets the distance threshold is retained, and the high-risk area within 2 mm of the endocardium is shielded.
[0049] Its technical effect is: by introducing the endocardial boundary as a spatial constraint condition into the probability map generation process, and using the minimum physical distance calculation and indicator function to screen the prediction results voxel by voxel, it is possible to actively eliminate anatomical high-risk areas during the ablation probability modeling stage, avoiding the subjectivity and inconsistency caused by relying on the operator's experience or manually setting the safety area in the traditional method, and effectively ensuring the tissue safety, algorithm controllability and result applicability of the ablation plan.
[0050] In combination with the first aspect, an embodiment of the present invention provides a sixth possible implementation of the first aspect, wherein constructing an ablation path energy optimization model based on the safe ablation target probability map and obtaining an ordered ablation point sequence through calculation using a path planning algorithm includes:
[0051] Based on the safe ablation target probability map, threshold screening is performed to obtain a set of candidate ablation points. ,in, is the preset probability threshold.
[0052] For ablation pathways , define the total energy function ,in, The first The coordinates of the ablation points, is the total number of points in the ablation path, For Move to The mobile energy consumption, Based on probability The target stability penalty term, A coefficient that balances mobile energy consumption and stability;
[0053] Based on the total energy function Solve the shortest path problem and output an ordered sequence of ablation points .
[0054] Its technical effect is: by constructing a path energy optimization model that combines catheter movement energy consumption and target stability penalty, and performing graph search or shortest path solution based on the candidate point set screened out in the safety probability graph, it realizes automated, multi-factor collaborative optimization planning of the ablation path. It not only avoids the subjectivity and local optimality problems of manual path setting, but also enables the output ablation point sequence to prioritize coverage of high-probability, low-risk areas while ensuring surgical efficiency, significantly improving the feasibility, energy economy and operational stability of the path.
[0055] In a second aspect, an embodiment of the present invention further provides a cardiac ablation planning system based on multimodal data fusion, which includes:
[0056] The data registration module is used to collect multi-source heterogeneous cardiac function data, uniformly map them to the standard anatomical coordinate system, and obtain the registered cardiac function data.
[0057] The dynamic feature extraction module is used to extract dynamic features from the registered cardiac function data.
[0058] The feature fusion module is used to perform cross-modal attention fusion on the dynamic features to obtain a fused feature map.
[0059] The ablation target probability map generation module is used to predict the ablation possibility of each voxel based on the fused feature map to obtain an ablation target probability map generation model.
[0060] The constraint module is used to set anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range.
[0061] A path planning module is used to construct an ablation path energy optimization model based on the safe ablation target probability map, and obtain an ordered ablation point sequence through calculation using a path planning algorithm.
[0062] In a third aspect, an embodiment of the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cardiac ablation planning method based on multimodal data fusion as described above is implemented.
[0063] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cardiac ablation planning method based on multimodal data fusion as described above.
[0064] The beneficial effects of the embodiments of the invention are:
[0065] This method performs spatiotemporal registration of heterogeneous multi-source data, including electrophysiological point clouds, dynamic MRI, CT anatomy, and preoperative heart sounds. By utilizing a joint optimization objective function for point cloud surface mapping and image feature extraction, it accurately aligns each modality into a single standard anatomical coordinate system. This eliminates the several-millimeter error associated with traditional rigid overlay, providing a highly accurate data foundation for subsequent feature extraction and target location.
[0066] The present invention not only extracts the local conduction velocity and potential waveform of electrophysiological signals, but also combines Cine-MRI's inter-frame difference to capture myocardial motion abnormalities, and uses bispectral analysis and PANet+FusionGhost to extract the nonlinear coupling characteristics of heart sounds. This allows the fused feature map to simultaneously cover multi-dimensional information of electrical activity, mechanical motion, heart sound dynamics, and anatomical structure, significantly surpassing the single-modality method's ability to comprehensively depict the lesion area.
[0067] The present invention adopts a dynamic attention mechanism to perform temporal weighting on the features of each modality, thereby adaptively balancing the contribution of electrophysiological, motion and heart sound features in different cardiac phases; and introduces an anatomical boundary indicator function in the predicted probability map to shield high-risk areas within 2mm of the endocardium, which not only ensures the accuracy of the ablation target distribution, but also takes into account tissue safety.
[0068] Based on a safe ablation probability map, this method constructs an energy minimization model that integrates motion energy consumption and target stability. This model maps the candidate point set into a graph structure, automatically solving for an ordered ablation sequence using a shortest path or autoregressive generation algorithm. Compared to traditional empirical or shortest distance approaches, this method's path better aligns with the physical characteristics of the catheter, reduces procedure time, and improves stability and repeatability. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.
[0070] Figure 1 This is a flow chart of the cardiac ablation planning method based on multimodal data fusion of the present invention. DETAILED DESCRIPTION
[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.
[0072] Please refer to Figure 1The first embodiment of the invention provides a cardiac ablation planning method based on multimodal data fusion, which includes: collecting multi-source heterogeneous cardiac function data, uniformly mapping them into a standard anatomical coordinate system, and obtaining registered cardiac function data; extracting dynamic features from the registered cardiac function data; performing cross-modal attention fusion on the dynamic features to obtain a fusion feature map; predicting the ablation possibility of each voxel on the basis of the fusion feature map, and obtaining an ablation target probability map generation model; setting anatomical boundary constraints for the ablation target probability map generation model, and obtaining a safe ablation target probability map representing the probability of target distribution within a safe range; constructing an ablation path energy optimization model based on the safe ablation target probability map, and obtaining an ordered ablation point sequence through a path planning algorithm.
[0073] The acquisition of multi-source heterogeneous cardiac function data and uniform mapping to a standard anatomical coordinate system to obtain registered cardiac function data includes: acquiring intracardiac electrophysiological point cloud data ,in, For the The three-dimensional coordinates of the electrodes in the standard anatomical coordinate system, For the corresponding time The voltage amplitude is the number of electrophysiological points, Index of time series frames; collect cardiac dynamic image sequences , indicating that each frame size is height ,width , depth layers 3D voxel data; Acquire static cardiac CT volume data ,in, is the image height, is the image width, is the image depth layer, collecting preoperative heart sound signals ,in, Indicates that the heart sound signal is of length One-dimensional time series; calculate the registration objective function ,in, is the rigid body transformation matrix of the electrophysiological point cloud, Indicates converting the point cloud into the corresponding spatial surface or shape expression. Indicates the extraction of endocardial deformation contours in dynamic images. Indicates that the matrix is applied to the point cloud data; the static cardiac CT volume data The coordinate system of the electrophysiological point cloud data is the standard anatomical coordinate system; Mapping to the standard anatomical coordinate system to obtain time-series registered electrophysiological data The cardiac dynamic image sequence The preoperative heart sound signal is collected under the standard anatomical coordinate system; The timestamps of the dirty dynamic image sequence The timestamps are aligned to obtain aligned heart sound signals .
[0074] Its technical effect is: by uniformly mapping electrophysiological point clouds, dynamic MRI, preoperative heart sounds and CT anatomical images into a standard three-dimensional coordinate system, it not only eliminates the spatial offset and temporal drift between different data sources, but also for the first time accurately aligns the timing information of heart sounds with electrical, biological and imaging data within the same anatomical framework, allowing subsequent feature extraction to synchronously consider electrical activity, mechanical movement and acoustic signals under a completely consistent anatomical reference.
[0075] The dynamic features extracted from the registered cardiac function data include: based on the time series registration electrophysiological data , calculate each pair of electrode points Local conduction velocity characteristics and the potential change sequence characteristics representing the potential waveform of each electrode , modeled by Transformer encoder to obtain electrophysiological features ,in, for point The three-dimensional coordinates of and is the sampling time of the corresponding point, A pair of electrode points The transmission speed, is the characteristic mapping function used to combine conduction velocity and potential information, A temporal modeling network with a multi-head attention mechanism; for the cardiac dynamic image sequence Perform inter-frame difference to obtain temporal motion features , represents the voxel intensity changes of adjacent frames, captures the contraction and relaxation dynamics of the myocardial area, and extracts cardiac motion features through three-dimensional convolution modeling ,in, For a three-dimensional convolutional network, the output Representing a high-dimensional feature map of myocardial motion in anatomical space; Perform 3D segmentation and extract anatomical features ,in, It is a convolution-based medical image segmentation model that outputs the probability of each anatomical region. is the weight matrix of the endocardial region, used to control the penetration depth, It is a pixel-by-pixel multiplication operation, integrating the structural distribution and ablation boundary security; the preoperative heart sound signal Perform bispectral analysis to generate a two-dimensional bispectral diagram of heart sounds The bispectral graph of heart sounds is input into a neural network based on partitioned attention and lightweight convolution for high-dimensional encoding to extract the deep features of the bispectral graph of heart sounds. ,in, Indicates a time point The bispectral diagram of the heart sound at It is a second-order spectrum calculation method used to reveal the nonlinear coupling characteristics of the signal. A module that combines the partitioned attention mechanism PANet with lightweight features to enhance FusionGhost.
[0076] Its technical effect is: by organically combining the timing dependence of conduction speed and potential waveform, the three-dimensional dynamic changes of motor motion, the fine anatomical structure segmentation and the nonlinear coupling characteristics of heart sounds, it not only realizes the deep and multi-scale encoding of four-dimensional information of electricity, biology, machine and sound, but also leverages the respective advantages of Transformer, 3D-CNN, U-Net and PANet-FusionGhost to effectively capture the multi-level interaction of cardiac function and structure, and significantly improves the accuracy and robustness of ablation target positioning.
[0077] The cross-modal attention fusion of the dynamic features to obtain the fusion feature map includes: for each time frame , the attention weights of each modality feature are adaptively calculated by the neural network to obtain the dynamic weights of the electrophysiological features , dynamic weight of cardiac motion characteristics , dynamic weight of deep features of heart sound bispectrum ,in, is the Sigmoid activation function, MLP is the multi-layer perceptron, It is global average pooling; multimodal features are weighted and fused to obtain a unified fusion feature map ,in, is the total number of time frames, 、 、 represents the voxel dimension in three-dimensional space, Indicates the number of feature channels.
[0078] Its technical effects are: introducing a dynamic attention mechanism at the time frame level, using neural networks to adaptively calculate the weight distribution of modalities such as electrophysiology, motion and heart sounds at different time points, breaking through the limitation of traditional static fusion that cannot respond to changes in the cardiac cycle, and achieving accurate modeling of the dynamic evolution of the importance of multimodal features over time, so that the fused feature map not only has the ability to collaboratively express multi-source information, but also has high sensitivity to the rhythmicity of lesions and the dynamics of functional disorders, thereby providing a more discriminative and clinically relevant high-dimensional semantic representation for target probability prediction.
[0079] The ablation probability of each voxel is predicted based on the fused feature map to obtain the ablation target probability map generation model, which includes: Based on this, a feature expansion layer is constructed to enhance spatial continuity. ,in, Represents the dilated convolutional layer, which is used to expand the receptive field without increasing parameters and enhance the spatial connectivity of features. is the number of channels after expansion; the feature expansion layer Perform three-dimensional convolution mapping to obtain a three-dimensional feature expansion layer ,in, It is a three-dimensional convolution operator with a kernel size of k×k×k, which is used to map high-dimensional features to the original score of each voxel; the three-dimensional feature expansion layer Perform Softmax normalization calculation to obtain the ablation target probability map generation model representing the original ablation probability ,in, is an exponential function, Voxel The original ablation score.
[0080] Its technical effect is: by introducing a feature expansion layer to expand the receptive field, the spatial continuity expression ability of the fusion feature is enhanced without increasing the calculation parameters, and the voxel-level features are structurally sensitive mapped by combining three-dimensional convolution, and the probability distribution is obtained through Softmax normalization, thereby achieving high-resolution, context-aware modeling of the spatial distribution of ablation targets, significantly improving the spatial consistency and positioning accuracy of the prediction results, and is particularly suitable for determining areas with blurred lesion boundaries or sparse local signals in complex cardiac chambers.
[0081] The method of setting anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range includes: calculating the minimum physical distance from each voxel to the anatomical boundary; ,in, is the set of endocardial boundary voxels, is the voxel side length; establish indicator function ; The safe ablation target probability map representing the target distribution probability within the safe range is calculated by the indicator function , wherein the safe ablation target probability map , which means that only the ablation probability that meets the distance threshold is retained, and the high-risk area within 2 mm of the endocardium is shielded.
[0082] Its technical effect is: by introducing the endocardial boundary as a spatial constraint condition into the probability map generation process, and using the minimum physical distance calculation and indicator function to screen the prediction results voxel by voxel, it is possible to actively eliminate anatomical high-risk areas during the ablation probability modeling stage, avoiding the subjectivity and inconsistency caused by relying on the operator's experience or manually setting the safety area in the traditional method, and effectively ensuring the tissue safety, algorithm controllability and result applicability of the ablation plan.
[0083] The ablation path energy optimization model is constructed based on the safe ablation target probability map, and the ordered ablation point sequence is obtained by calculating the path planning algorithm, including: performing threshold screening based on the safe ablation target probability map to obtain a candidate ablation point set. ,in, is the preset probability threshold; for the ablation path , define the total energy function ,in, The first The coordinates of the ablation points, is the total number of points in the ablation path, For Move to The mobile energy consumption, Based on probability The target stability penalty term, The larger it is, the smaller the penalty is, which encourages the selection of high probability points. A coefficient for balancing mobile energy consumption and stability; Based on the total energy function Solve the shortest path problem and output an ordered sequence of ablation points .
[0084] Its technical effect is: by constructing a path energy optimization model that combines catheter movement energy consumption and target stability penalty, and performing graph search or shortest path solution based on the candidate point set screened out in the safety probability graph, it realizes automated, multi-factor collaborative optimization planning of the ablation path. It not only avoids the subjectivity and local optimality problems of manual path setting, but also enables the output ablation point sequence to prioritize coverage of high-probability, low-risk areas while ensuring surgical efficiency, significantly improving the feasibility, energy economy and operational stability of the path.
[0085] The second embodiment of the present invention provides a cardiac ablation planning system based on multimodal data fusion, which includes: a data registration module for collecting multi-source heterogeneous cardiac function data, uniformly mapping them into a standard anatomical coordinate system, and obtaining registered cardiac function data; a dynamic feature extraction module for extracting dynamic features from the registered cardiac function data; a feature fusion module for performing cross-modal attention fusion on the dynamic features to obtain a fusion feature map; an ablation target probability map generation module for predicting the ablation possibility of each voxel based on the fusion feature map to obtain an ablation target probability map generation model; a constraint module for setting anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the target distribution probability within a safe range; a path planning module for constructing an ablation path energy optimization model based on the safe ablation target probability map, and obtaining an ordered ablation point sequence through calculation by a path planning algorithm.
[0086] A third embodiment of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cardiac ablation planning method based on multimodal data fusion as described above is implemented.
[0087] A fourth embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the aforementioned cardiac ablation planning method based on multimodal data fusion.
[0088] The computer program product of the cardiac ablation planning method and apparatus based on multimodal data fusion provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method in the previous method embodiment. For specific implementation, please refer to the method embodiment and will not be repeated here.
[0089] Specifically, the storage medium can be a general storage medium, such as a mobile disk, a hard disk, etc. When the computer program on the storage medium is run, it can execute the above-mentioned cardiac ablation planning method based on multimodal data fusion, thereby realizing automated and quantitative target probability prediction, automatically calculating the optimal ablation point sequence, and greatly improving planning efficiency and repeatability.
[0090] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.
[0091] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.
Claims
1. A cardiac ablation planning method based on multimodal data fusion, characterized in that: include: Collect multi-source heterogeneous cardiac function data, uniformly map them to the standard anatomical coordinate system, and obtain registered cardiac function data; extracting dynamic features from the registered cardiac function data; Performing cross-modal attention fusion on the dynamic features to obtain a fused feature map; Predicting the ablation possibility of each voxel based on the fused feature map to obtain an ablation target probability map generation model; Setting anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range; An ablation path energy optimization model is constructed based on the safe ablation target probability map, and an ordered ablation point sequence is calculated through a path planning algorithm.
2. The cardiac ablation planning method based on multimodal data fusion according to claim 1, characterized in that: The multi-source heterogeneous cardiac function data is collected and uniformly mapped to a standard anatomical coordinate system to obtain the registered cardiac function data, including: Collect intracardiac electrophysiological point cloud data ,in, For the The three-dimensional coordinates of the electrodes in the standard anatomical coordinate system, For the corresponding time The voltage amplitude is the number of electrophysiological points, is the time frame index; Acquire dynamic cardiac image sequences , indicating that each frame size is height ,width , depth layers 3D voxel data; Acquisition of static cardiac CT volume data ,in, is the image height, is the image width, is the number of image depth layers; Collect preoperative heart sound signals ,in, Indicates that the heart sound signal is of length One-dimensional time series of; Calculate the registration objective function ,in, is the rigid body transformation matrix of the electrophysiological point cloud, Indicates converting the point cloud into the corresponding spatial surface or shape expression. Indicates the extraction of endocardial deformation contours in dynamic images. Indicates applying the matrix to point cloud data; The static cardiac CT volume data The coordinate system is the standard anatomical coordinate system; The electrophysiological point cloud data Mapping to the standard anatomical coordinate system to obtain time-series registered electrophysiological data ; The cardiac dynamic image sequence The data were collected under the standard anatomical coordinate system; The preoperative heart sound signal The timestamps of the dirty dynamic image sequence The timestamps are aligned to obtain aligned heart sound signals .
3. The cardiac ablation planning method based on multimodal data fusion according to claim 2, characterized in that: The dynamic features extracted from the registered cardiac function data include: Based on the temporal registration of electrophysiological data , calculate each pair of electrode points Local conduction velocity characteristics and the potential change sequence characteristics representing the potential waveform of each electrode , modeled by Transformer encoder, and obtained electrophysiological features ,in, for point The three-dimensional coordinates of and is the sampling time of the corresponding point, A pair of electrode points The transmission speed, is the characteristic mapping function used to combine conduction velocity and potential information, Modeling networks for temporal sequences with multi-head attention mechanisms; The cardiac dynamic image sequence Perform inter-frame difference to obtain temporal motion features , extracting cardiac motion features through 3D convolutional modeling ,in, It is a three-dimensional convolutional network; The static cardiac CT volume data Perform 3D segmentation and extract anatomical structure features ,in, is a convolution-based medical image segmentation model. is the weight matrix of the endocardial region, It is a pixel-by-pixel multiplication operation; The preoperative heart sound signal Perform bispectral analysis to generate a two-dimensional bispectral diagram of heart sounds The bispectral graph of heart sounds is input into a neural network based on partitioned attention and lightweight convolution for high-dimensional encoding to extract the deep features of the bispectral graph of heart sounds. ,in, Indicates a time point The bispectral diagram of the heart sound at is the second-order spectrum calculation method, It is a module that combines partitioned attention mechanism with lightweight feature enhancement.
4. The cardiac ablation planning method based on multimodal data fusion according to claim 3, characterized in that: The cross-modal attention fusion of the dynamic features to obtain a fusion feature map includes: For each time frame , the attention weights of each modal feature are adaptively calculated by the neural network to obtain the dynamic weights of the electrophysiological features , dynamic weight of cardiac motion characteristics , dynamic weight of deep features of heart sound bispectrum ,in, is the Sigmoid activation function, MLP is the multi-layer perceptron, is global average pooling; Weighted fusion of multimodal features to obtain a unified fusion feature map ,in, is the total number of time frames, 、 、 represents the voxel dimension in three-dimensional space, Indicates the number of feature channels.
5. The cardiac ablation planning method based on multimodal data fusion according to claim 4, characterized in that: The ablation possibility of each voxel is predicted based on the fused feature map to obtain an ablation target probability map generation model, which includes: In the fused feature map Based on this, a feature expansion layer is constructed to enhance spatial continuity. ,in, represents the dilated convolutional layer, is the number of channels after expansion; The feature expansion layer Perform three-dimensional convolution mapping to obtain a three-dimensional feature expansion layer ,in, is a three-dimensional convolution operator; The three-dimensional feature expansion layer Perform Softmax normalization calculation to obtain the ablation target probability map generation model representing the original ablation probability ,in, is an exponential function, Voxel The original ablation score.
6. The cardiac ablation planning method based on multimodal data fusion according to claim 5, characterized in that: Setting anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range includes: Calculate the minimum physical distance from each voxel to the anatomical boundary ,in, is the set of endocardial boundary voxels, is the voxel side length; Create an indicator function ; A safe ablation target probability map representing the probability of target distribution within a safe range is obtained by calculating the indicator function. , wherein the safe ablation target probability map , which means that only the ablation probability that meets the distance threshold is retained, and the high-risk area within 2 mm of the endocardium is shielded.
7. The cardiac ablation planning method based on multimodal data fusion according to claim 6, characterized in that: The ablation path energy optimization model is constructed based on the safe ablation target probability map, and the ordered ablation point sequence is obtained by calculating the path planning algorithm, including: Based on the safe ablation target probability map, threshold screening is performed to obtain a set of candidate ablation points. ,in, is the preset probability threshold; For ablation pathways , define the total energy function ,in, The first The coordinates of the ablation points, is the total number of points in the ablation path, For Move to The mobile energy consumption, Based on probability The target stability penalty term, A coefficient that balances mobile energy consumption and stability; Based on the total energy function Solve the shortest path problem and output an ordered sequence of ablation points .
8. A cardiac ablation planning system based on multimodal data fusion, characterized in that: include: The data registration module is used to collect multi-source heterogeneous cardiac function data, uniformly map them to the standard anatomical coordinate system, and obtain the registered cardiac function data; A dynamic feature extraction module, configured to extract dynamic features from the registered cardiac function data; A feature fusion module, configured to perform cross-modal attention fusion on the dynamic features to obtain a fused feature map; an ablation target probability map generation module, configured to predict the ablation possibility of each voxel based on the fused feature map, and obtain an ablation target probability map generation model; a constraint module, configured to set anatomical boundary constraints for the ablation target probability map generation model to obtain a safe ablation target probability map representing the probability of target distribution within a safe range; A path planning module is used to construct an ablation path energy optimization model based on the safe ablation target probability map, and obtain an ordered ablation point sequence through calculation using a path planning algorithm.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the cardiac ablation planning method based on multimodal data fusion according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the cardiac ablation planning method based on multimodal data fusion according to any one of claims 1 to 7 is implemented.