Left atrial wall thickness evaluation and intraoperative navigation method and system based on artificial intelligence

Abstract

The invention provides a left atrium wall thickness evaluation and intraoperative navigation method and system based on artificial intelligence. Comprising the steps of collection and preprocessing of multi-modal medical image data, artificial intelligence image segmentation under topological constraints, three-dimensional structure reconstruction and inner and outer surface separation, wall thickness calculation and risk assessment, visual mapping interaction of wall thickness and risk levels, and intraoperative navigation and dynamic adjustment. The system automatically identifies a left atrium structure through a deep learning model, introduces topological consistency constraint to improve segmentation precision, and realizes quantitative measurement of wall thickness and high-risk region identification through three-dimensional reconstruction and geometric analysis. The method supports intra-operative real-time image synchronization and path dynamic updating, and has closed-loop navigation and risk early warning capabilities. According to the invention, accurate evaluation of the thickness of the left atrial wall and intelligent navigation in the operation can be realized, and the safety and individualized level of treatment of arrhythmia such as atrial fibrillation and the like are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing and artificial intelligence, and specifically relates to a method and system for left atrial wall thickness evaluation and intraoperative navigation based on artificial intelligence. The method and system comprehensively use multi-modal medical image acquisition, deep learning segmentation, three-dimensional reconstruction, real-time visualization interaction, and closed-loop risk management, and proposes a complete technical solution combining preoperative accurate evaluation and intraoperative intelligent navigation. TECHNICAL BACKGROUND

[0002] Atrial fibrillation is the most common persistent arrhythmia, usually manifested as rapid and irregular atrial electrical activity, and is closely related to various serious cardiovascular diseases such as stroke and heart failure. The occurrence of atrial fibrillation is closely related to structural changes, mechanical dysfunction, and changes in atrial wall thickness of the left atrium. In particular, during the preoperative evaluation of atrial fibrillation, accurate measurement of left atrial wall thickness and its distribution is crucial for developing personalized treatment plans, optimizing preoperative planning, predicting treatment effectiveness, and reducing the risk of postoperative complications.

[0003] Changes in left atrial wall thickness are considered one of the key factors in the pathogenesis of atrial fibrillation. Atrial fibrillation is often accompanied by structural remodeling of the left atrial wall, particularly pathological changes such as atrial wall hypertrophy and fibrosis. These abnormalities not only affect the electrophysiological properties of the atrium, but also can lead to atrial dilation and dysfunction, further increasing the risk of atrial fibrillation. Therefore, there is an urgent need in clinical practice for a technology that can accurately evaluate left atrial wall thickness preoperatively, identify weak or abnormal areas, and provide dynamic navigation and real-time risk warning during surgery, in order to optimize individualized treatment strategies for atrial fibrillation.

[0004] The current mainstream methods for left atrial wall thickness evaluation and intraoperative navigation in clinical practice have the following limitations:

[0005] 1. Traditional two-dimensional or single-modality methods are difficult to comprehensively and accurately measure the complex three-dimensional structure of the left atrial wall;

[0006] 2. Point-by-point or simple visualization-based solutions lack comprehensive analysis of anatomical integrity and dynamic risk;

[0007] 3. Some studies focus on offline statistics or two-dimensional heat maps, providing only limited guidance value and being difficult to meet the real-time and individualized intraoperative application needs.

[0008] With the rapid development of deep learning technology in the field of medical image segmentation and analysis, the introduction of artificial intelligence for automatic fine segmentation of left atrium and its wall structure, combined with anatomical structure preservation and dynamic risk identification, can become an important means to improve the evaluation accuracy. In view of the high requirements for anatomical continuity and topological consistency in atrial wall reconstruction and thickness calculation, the topological loss function (such as mathematical methods based on persistent homology) can be combined to strengthen the model's accurate grasp of the atrial structure. In addition, the fusion of multi-modal images (CT, MRI, 3DE, etc.) further improves the segmentation robustness and evaluation accuracy, thereby laying a technical foundation for the closed-loop intelligent system of preoperative evaluation-intraoperative navigation. SUMMARY

[0009] The present application combines artificial intelligence technology and image processing algorithms to provide a left atrial wall thickness evaluation and intraoperative navigation method based on artificial intelligence, including the following steps:

[0010] Step 1: Multi-modal data acquisition and preprocessing, obtaining multi-modal medical image data containing complete left heart anatomical information, and preprocessing to improve data compatibility and calculation accuracy;

[0011] Step 2: Artificial intelligence segmentation under topological constraints, when analyzing the preprocessed medical image data, using artificial intelligence segmentation method to automatically identify the left atrium and related anatomical structures, and optimizing the segmentation results through topological consistency constraints;

[0012] Step 3: Three-dimensional reconstruction and surface separation, based on the segmentation results, constructing a three-dimensional structure model of the left atrium, and analyzing and optimizing its surface; using geometric feature extraction and spatial separation methods to distinguish key anatomical regions to support subsequent wall thickness calculation and visualization analysis;

[0013] Step 4: Wall thickness calculation and regional risk assessment, measuring the thickness and partitioning the regions of the three-dimensional structure of the left atrial wall, combining various anatomical and physiological parameters to assess potential high-risk areas and provide quantitative analysis to support individualized treatment strategies;

[0014] Step 5: Visualization mapping and three-dimensional interaction, based on the calculation results, mapping the left atrial wall thickness and risk information to a unified visualization model, using interactive three-dimensional rendering technology to provide intuitive visual feedback, supporting real-time query and labeling of target areas by users;

[0015] Step 6: Closed-loop intraoperative navigation and dynamic adjustment, in the operation, combining real-time image and position information to dynamically update the left atrial model and provide personalized navigation scheme and risk warning.

[0016] Further, in the step 1, the multi-modal data includes CT (CTA), MRI and three-dimensional echocardiogram (3DE) for obtaining complete left heart anatomical information; in view of the possible coordinate system differences of multi-source data, a rigid or non-rigid registration algorithm is used to align the multi-modal images; in view of the images with different resolutions or sampling intervals, a linear interpolation or high-order interpolation method is used for spatial resampling to ensure the accuracy and consistency of subsequent processing.

[0017] Further, in the step 2, after the image preprocessing is completed, a segmentation algorithm based on deep learning is used to automatically extract the left heart anatomical structure, including the left atrium, left atrial wall, left atrial appendage, pulmonary vein and aorta, to realize accurate segmentation of key regions, and the deep learning algorithm includes nnUNet, SwinUNETR and MedCLIP-SAMv2; in order to improve the segmentation accuracy of complex atrial structure and maintain the topological consistency, a topological loss function based on persistent homology is introduced to constrain the deep learning segmentation network, so as to ensure the continuity of the atrial wall and the anatomical rationality, and to minimize the misclassification or fracture phenomenon; the segmented left heart main anatomical region includes the left atrium, left atrial wall, left atrial appendage, pulmonary vein and aorta; the mathematical expression of the topological loss function is as follows:

[0018]

[0019] Wherein, And The persistence bar graph of the segmented image S and the real label G is represented by p and q respectively.

[0020] Further, in the step 3, based on the segmentation result obtained in the step 2, an improved MarchingCubes three-dimensional reconstruction algorithm is used to construct a three-dimensional model of the left atrial wall, and the inner surface and the outer surface are separated; through geometric analysis and region growing method, four kinds of quantifiable geometric parameters of the surface normal vector, average curvature, Gaussian curvature and principal curvature of the three-dimensional model are calculated and processed comprehensively; the inner surface and the outer surface are saved in the form of independent grid or point cloud respectively to ensure that the corresponding points of the inner and outer surfaces can be accurately matched when calculating the wall thickness later, and the mathematical expression is as follows:

[0021]

[0022] Wherein, f(p) is a scalar function of the model surface point, The gradient of the point is the normal vector The unitized vector of the gradient is.

[0023] Further, in the step 4, the wall thickness calculation and the regional risk assessment specifically include the following steps:

[0024] Step 4.1: To achieve precise local analysis, the left atrial wall is automatically divided into the following six anatomical regions: superior wall, posterior wall, septal wall, anterior wall, left lateral wall, and inter-pulmonary vein region. In the region automatic identification process, first, the division is performed according to the pre-defined anatomical marker points or based on the deep learning partition model, and the user is allowed to make necessary manual fine-tuning according to actual needs, to ensure that the partition boundary meets the clinical judgment standard.

[0025] Step 4.2: For each anatomical region, the shortest distance algorithm between the inner and outer surfaces is used to calculate the average wall thickness; at the same time, the left atrial volume index (LAVI) and the left atrial wall signal intensity normalization index (IIR) are introduced for multi-dimensional risk assessment; LAVI reflects the ratio of left atrial volume to body surface area, combined with average thickness analysis, wall thickness change rate and local expansion coefficient, to assess the local or overall risk level; based on the above parameters, each region is automatically scored, and a warning is issued for high-risk areas, prompting clinicians to pay special attention.

[0026] Further, in the step 5, the visualization mapping and interaction method includes: based on the wall thickness and risk grading data obtained in step 4, the numerical value is mapped to a unified color space to realize intuitive visualization of left atrial wall thickness and risk grading; using a segmented color mapping algorithm and OpenGL three-dimensional rendering technology, interactive model operation is realized to support model rotation, scaling and translation, and allow users to select, label and query information in real time for local areas; users can directly obtain the thickness parameters and risk prompts of the target region in the three-dimensional visualization interface, providing visual auxiliary support for clinical decision-making;

[0027] The color mapping mathematical model is as follows:

[0028]

[0029] wherein each Color i corresponds to the color value of RB=(r i ,g i ,b i ).

[0030] Further, in the step 6, the intraoperative navigation and dynamic adjustment method includes: by establishing a data interface, real-time communication is realized with the electrophysiology mapping platform through HL7 / FHIR protocol or hospital information system (HIS) standard, to realize real-time fusion of catheter position and local electrical activity characteristics; when the catheter tip or ablation instrument enters the high-risk area, the system will give a high-light warning in the three-dimensional visualization interface, and prompt the operator to adjust the path when necessary;

[0031] Further, in step 6, if intraoperative X-ray fluoroscopy, intracardiac echocardiography (ICE) or transesophageal echocardiography (TEE) is equipped, the pose offset of intraoperative images can be automatically corrected based on the non-rigid registration method to ensure that the navigation view is consistent with the patient's anatomy.

[0032] Further, in step 6, the improved A* or Dijks tra path planning algorithm can also be used to provide alternative ablation path suggestions to the doctor according to real-time data, and allow the operator to manually confirm and synchronize with one key, realizing closed-loop navigation and dynamic risk control.

[0033] The application also provides an artificial intelligence-based left atrial wall thickness evaluation and intraoperative navigation system, comprising:

[0034] A data acquisition and preprocessing module is used to acquire multi-modal medical image data and preprocess it to improve data compatibility and calculation accuracy.

[0035] An image segmentation module is used to automatically identify the left atrium and related anatomical structures when analyzing the preprocessed medical image data, and to optimize the segmentation results through topological consistency constraints.

[0036] A three-dimensional modeling and structure analysis module is used to construct a three-dimensional structure model of the left atrium based on the segmentation results, analyze and optimize its surface, and use geometric feature extraction and spatial separation methods to distinguish key anatomical regions to support subsequent wall thickness calculation and visual analysis.

[0037] A wall thickness calculation and risk assessment module is used to measure the thickness of the three-dimensional structure of the left atrium and to partition the region, combine various anatomical and physiological parameters, assess potential high-risk areas, and provide quantitative analysis to support individualized treatment strategies.

[0038] A visualization and interaction module is used to map the left atrial wall thickness and risk information to a unified visualization model based on the calculation results, use interactive three-dimensional rendering technology to provide intuitive visual feedback, and support user real-time query and annotation of target areas.

[0039] An intraoperative navigation and dynamic adjustment module is used to dynamically update the left atrial model in combination with real-time images and position information during surgery, and to provide personalized navigation solutions and risk warnings.

[0040] Technical effects

[0041] The present application realizes high-precision reconstruction and comprehensive risk analysis of the left atrial wall through key technical links such as multi-modal medical data acquisition and preprocessing, artificial intelligence segmentation (introducing topological loss function to reduce missegmentation), three-dimensional reconstruction and surface separation, wall thickness calculation and regional risk assessment, visualization mapping and interactive navigation, and closed-loop intraoperative dynamic adjustment. Specifically:

[0042] Segmentation optimization: Deep learning network combined with persistent homology topological loss function significantly improves the consistency of segmentation results with actual anatomical structure, reducing "breakage" and "false channel" phenomena.

[0043] Multi-modal fusion: CT / MRI data registration and resampling enhancement algorithm robustness, adapt to different patient and equipment needs.

[0044] Partition and risk assessment: From the perspective of anatomy, functional partitions (such as the upper wall, posterior wall, etc.) are divided, and multi-dimensional risk assessment is conducted in combination with average thickness, LAVI and IIR indicators to accurately predict fibrosis and potential complications.

[0045] Visualization interaction: Segmental color mapping and OpenGL rendering provide an intuitive and convenient operation interface, supporting model rotation, scaling, translation and local parameter point selection viewing.

[0046] Intraoperative navigation: Breakthrough offline planning limitations, linkage with intraoperative mapping system, real-time update of model and risk warning information, realization of individualized dynamic navigation, greatly improving the safety and treatment effect of surgery. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flowchart of an artificial intelligence-based left atrial wall thickness real-time evaluation and interactive navigation method.

[0048] Figure 2 is a structural diagram of multi-modal image registration and fusion.

[0049] Figure 3 is a comparison chart of artificial intelligence segmentation results before and after introducing topological constraints.

[0050] Figure 4 is a three-dimensional reconstruction and surface separation processing diagram of the left atrium.

[0051] Figure 5 is a left atrial anatomical region division and thickness distribution diagram.

[0052] Figure 6 is a segmented color mapping and three-dimensional interactive operation interface diagram.

[0053] Figure 7 is an application diagram of the intraoperative navigation system in the actual surgical scene. DETAILED DESCRIPTION

[0054] The application will be further described in conjunction with specific embodiments. It should be understood that the embodiments are only used to illustrate but not to limit the protection scope of the application. Moreover, it should be understood that those skilled in the art can make various modifications or amendments to the application after reading the disclosure of the application, and these equivalent forms also fall within the protection scope defined by the application.

[0055] The application proposes an artificial intelligence-based left atrial wall thickness evaluation and intraoperative navigation method, the process is as shown in Figure 1 , comprising the following steps:

[0056] Step 1: Multimodal data acquisition and preprocessing, as shown in Figure 2 . Specifically, the following steps are included:

[0057] Step 1.1: Multimodal data acquisition and preprocessing.

[0058] This step mainly obtains preoperative cardiac imaging data of atrial fibrillation patients, including computed tomography (CT), magnetic resonance imaging (MRI), and three-dimensional echocardiography (3DE). During image acquisition, it is necessary to ensure that the acquired image data completely covers the left atrium and its surrounding key anatomical structure area, so as to ensure the integrity of subsequent processing and the accuracy of analysis.

[0059] Further, the acquired medical image data needs to be standardized and preprocessed, including image noise reduction, contrast enhancement, and spatial standardization, etc. to upgrade the image quality, enhance the consistency between multiple modalities, and improve the accuracy of subsequent wall thickness evaluation.

[0060] Specifically, for CT images, Hounsfield units (HU) are used for standardization processing; for MRI images, signal intensity normalization processing is used; and for three-dimensional echocardiography (3DE), time-depth correction and spatial resolution optimization are used to reduce artifact interference and enhance boundary clarity.

[0061] Step 1.2: Alignment and resampling.

[0062] In order to meet the requirements of accurate measurement of left atrial wall thickness, the spatial resolution of preprocessed image data should not be less than 0.5mm×0.5mm×0.5mm. Considering the possible inconsistency of coordinate systems and spatial resolution of multimodal data, a non-rigid registration algorithm is used to align the images to the standard world physical coordinate system. After completing the coordinate alignment, in view of the resolution difference of each modality image, linear interpolation or cubic interpolation algorithm is further used for resampling operation, so that the pixel voxel size of each image in the three-dimensional direction is as consistent as possible, providing compatible input data basis for subsequent segmentation, three-dimensional reconstruction and thickness calculation.

[0063] Step 1.3: Standardization and denoising.

[0064] After image registration and resampling, the multi-modal image data is further standardized and denoised to improve the overall image quality and eliminate random interference. Specifically, CT images are standardized in intensity using Hounsfield units (HU), and MRI images are normalized in signal intensity. Wavelet transform or bilateral filtering is used for denoising to remove artifacts and random noise. In the preprocessing of three-dimensional echocardiography (3DE) data, inter-frame registration technology is used to align the images within the cardiac cycle to address the image blurring caused by cardiac motion. In addition, due to the low spatial resolution of 3DE data, a super-resolution reconstruction algorithm is introduced to improve image detail performance and ensure compatibility with CT and MRI data. After the above processing, high-quality multi-modal fusion images are obtained, providing reliable input support for subsequent intelligent segmentation and three-dimensional reconstruction.

[0065] Step 2: Artificial intelligence segmentation under topological constraints

[0066] Step 2.1: Deep learning model construction.

[0067] The multi-modal fusion images obtained in step 1 are input into the deep learning segmentation network based on nnUNet, SwinUNETR, or MedCLIP-SAMv2 improved model to realize automatic identification and segmentation of key anatomical structures such as left atrial wall, left atrial appendage, and pulmonary vein orifice.

[0068] Step 2.2: Introduction of topological loss function

[0069] To improve the recognition accuracy and continuity of the complex structure of the left atrial wall, the invention introduces a topological loss function based on the theory of persistent homology (Persis tent Homology) during model training. As shown in Figure 3 The results show that this method can effectively improve the anatomical consistency of the segmentation results.

[0070] The topological loss function is used to capture the anatomical topological properties of the atrial wall, such as closure and connectivity. If there are structural abnormalities in the segmentation results that do not conform to the true anatomy (such as holes or breaks that should not exist), the topological loss will penalize them, guiding the segmentation model to generate segmentation results that conform to the anatomical structure of the left atrial wall and maintain the correct connection relationship with the surrounding tissues.

[0071] The mathematical expression of the topological loss function is as follows:

[0072]

[0073] where, and p and q are points in the persistence bar plots of S and G, respectively.

[0074] Step 3: Three-dimensional reconstruction and surface separation

[0075] Step 3.1: Three-dimensional model generation

[0076] As shown in Figure 4 , based on the segmentation mask obtained in step 2, a modified Marching Cubes algorithm is used to perform three-dimensional mesh reconstruction on the left atrial region, thereby preliminarily generating a three-dimensional geometric model of the left atrium. High-frequency noise mesh patches generated during the reconstruction process are smoothed, and preferably using Laplacian smoothing or Taubin smoothing methods, while preserving important anatomical features, effectively removing redundant noise points.

[0077] Step 3.2: Inner and outer surface separation

[0078] In order to accurately calculate the left atrial wall thickness in subsequent steps, the left atrial wall model needs to be divided into inner and outer surfaces. Let f(p) be the implicit function at point P on the model surface, then the gradient of this point is:

[0079]

[0080] This gradient vector can be expressed as the normal vector at this point, and after unitization processing, the normalized normal vector expression is obtained:

[0081]

[0082] Through the analysis of the above normal vector and curvature characteristics, the inner and outer surfaces can be extracted and represented as independent point clouds or mesh data structures, to support subsequent wall thickness calculation and spatial correspondence matching.

[0083] Step 4: Wall thickness calculation and regional risk assessment

[0084] Step 4.1: Anatomical region division

[0085] The present application automatically divides the left atrial wall into six regions: Superior Wall, Posterior Wall, Septal Wall, Anterior Wall, Left Lateral Wall, and Between Superior and Inferior Homolateral Pulmonary Veins, from the clinical needs. As shown in Figure 5After the main anatomical landmarks are labeled on the 3D model, the system automatically generates region boundaries based on deep learning-based automatic partitioning model or geometric rules.

[0086] To ensure the clinical usability of the partition boundaries, the system supports user manual adjustment of the automatic partitioning results, such as adjusting the region boundaries by dragging or selecting, to match the actual surgeon's judgment and anatomical experience, thereby enhancing the consistency of the model with clinical practice.

[0087] Step 4.2: Thickness calculation and multi-index risk warning

[0088] Step 4.2.1: Average thickness calculation

[0089] Let d(p) represent the shortest distance between a point p on the inner surface and the outer surface, and S outer For the outer surface point set, the wall thickness of the point can be defined as:

[0090]

[0091] The average wall thickness of the region can be obtained by calculating the average value of all inner surface point sets in a certain anatomical region:

[0092]

[0093] Step 4.2.2: LAVI index calculation

[0094] Combined with the patient's body surface area (Body Surface Area, BSA), according to the image segmentation results or clinical electronic medical records, the left atrial volume V LA is obtained, and then the left atrial volume index LAVI

[0095]

[0096] If LAVI exceeds the clinical threshold (e.g., >58.6 mL / m 2 ) and the local wall thickness is higher than the set standard (e.g., >1.69 mm), it indicates that the left atrial structure is remodeling, which may have potential risk of arrhythmia.

[0097] Step 4.2.3: IIR index calculation

[0098] The left atrial wall signal intensity normalization index IIR (Intensity Index Ratio) is introduced to judge whether the tissue has a fibrosis trend. Its calculation formula is as follows:

[0099]

[0100] Where, SI ROISignal intensity of the region of interest, SI Reference Signal intensity of the reference region. When IIR>1.20, it is determined that there is a tendency of fibrosis, and the operator is advised to pay attention.

[0101] Step 5: Based on the left atrial wall thickness and risk information obtained in step 4, the present application maps the relevant numerical parameters to a unified color space, realizing the intuitive visualization of the left atrial wall thickness distribution and risk classification. As shown in Figure 6 , a segmented color mapping method is used in combination with OpenGL three-dimensional rendering technology to construct an interactive visualization model.

[0102] The model supports user rotation, scaling, and translation operations on the three-dimensional atrial structure, and can perform real-time point selection or labeling on local areas. Users can directly view the thickness values and risk prompt information of the target area in the three-dimensional view, providing visual decision support for preoperative evaluation and intraoperative navigation.

[0103] The color mapping model is defined as follows:

[0104]

[0105] where each color value Color i represents an RGB=(r i ,g i ,b i ).

[0106] Step 6: Closed-loop intraoperative navigation and dynamic adjustment

[0107] As shown in Figure 7 , during the intraoperative stage, the system realizes closed-loop navigation function, dynamically fusing real-time images, electrophysiological signals, and navigation feedback to improve the individual precision and safety of atrial fibrillation ablation surgery.

[0108] Step 6.1: Data acquisition and image synchronization

[0109] The system is communicated with the platform used in clinical practice, and the three-dimensional coordinates of the catheter tip (positioning accuracy up to ±0.5mm) and electrical activity information such as activation time are received in real time through the HL7 / FHIR interface. If X-ray fluoroscopy or intracardiac echocardiography (ICE) is used during atrial fibrillation ablation, the fluoroscopy or ultrasound image is transmitted to the system in the form of DICOM real-time stream (frame rate ≥15fps). To reduce the left atrial position offset caused by heartbeat and respiration, a B-spline non-rigid registration algorithm (control point grid spacing about 5mm) is used to update the image registration state every 2-3 seconds during the operation, ensuring that the alignment error between the virtual model and the patient's physical heart anatomy is controlled within 1.5mm.

[0110] Step 6.2: Risk Linkage and Path Guidance

[0111] When the mapping data indicates that the catheter tip has entered or is close to a high-risk area (e.g. wall thickness <2.5mm, IIR>1.20), the system visually identifies it with red highlights on the left atrial three-dimensional model, and prompts the doctor with a pop-up window or sound to be cautious.

[0112] At the same time, the system dynamically calls the improved A* algorithm based on intraoperative data to generate alternative ablation routes (can be set to two or three) with the dual objective function of "avoiding high-risk areas" and "shortest path". Superimpose it on the three-dimensional model, different colors correspond to different priorities. If the real-time collected local potential shows that the lesion range has increased, the risk boundary radius can also be adaptively expanded (for example, from 2.0mm to 3.0mm), and the optimal path is recalculated.

[0113] Step 6.3: Intraoperative Interaction and Path Execution

[0114] After the operator clicks the "path preview" function on the navigation interface, the system will display the candidate paths (color-coded priority) in the three-dimensional model, and mark the minimum distance (accuracy 0.1mm) between each path and the surrounding high-risk area. After confirmation, the path coordinate sequence is sent to the ablation catheter control device through RS232 or other communication protocols (transmission delay <100ms), and the predetermined catheter trajectory is displayed in the interface as a dashed line.

[0115] When there is deviation or abnormal data during operation, the operator can perform "emergency pause" and re-plan to ensure real-time dynamic adjustment. After the operation is completed, the system will automatically generate a report document of the operation trajectory, risk trigger alarm frequency and local IIR change, etc., to provide support for postoperative review and recurrence prediction.

Claims

1. An artificial intelligence-based left atrial wall thickness evaluation and intraoperative navigation method, characterized by Comprising the following steps: Step 1: Multi-modal data acquisition and preprocessing, acquiring multi-modal medical image data containing complete left heart anatomical information, and preprocessing it to improve data compatibility and calculation accuracy; Step 2: Artificial intelligence segmentation under topological constraints, when analyzing the preprocessed medical image data, using artificial intelligence segmentation method to automatically identify left atrium and related anatomical structures, and optimizing the segmentation results through topological consistency constraints; Step 3: Three-dimensional reconstruction and surface separation, based on the segmentation results, constructing a three-dimensional structure model of the left atrium, and analyzing and optimizing its surface; using geometric feature extraction and spatial separation method, distinguishing key anatomical regions to support subsequent wall thickness calculation and visual analysis; Step 4: Wall thickness calculation and regional risk assessment, measuring the thickness of the left atrial wall and partitioning the region, combining various anatomical and physiological parameters to assess potential high-risk areas and provide quantitative analysis to support individualized treatment strategies; Step 5: Visualization mapping and three-dimensional interaction, based on the calculation results, mapping the left atrial wall thickness and risk information to a unified visualization model, using interactive three-dimensional rendering technology to provide intuitive visual feedback to support user real-time query and annotation of target areas; Step 6: Closed-loop intraoperative navigation and dynamic adjustment, combining real-time images and position information to dynamically update the left atrial model and provide personalized navigation solutions and risk warnings during surgery.

2. The left atrial wall thickness evaluation and intraoperative navigation method according to claim 1, in the step 1, the multi-modal data includes CT, MRI and three-dimensional echocardiogram for obtaining complete left heart anatomical information; rigid or non-rigid registration algorithm is used to align multi-modal images for possible coordinate system differences; linear interpolation or high-order interpolation method is used for spatial resampling of images with different resolutions or sampling intervals to ensure the accuracy and consistency of subsequent processing.

3. The left atrial wall thickness evaluation and intraoperative navigation method according to claim 2, in the step 2, after completing image preprocessing, a deep learning-based segmentation algorithm is used to automatically extract left heart anatomical structures, including left atrium, left atrial wall, left atrial appendage, pulmonary vein and aorta, to achieve accurate segmentation of key areas, the deep learning algorithm includes nnUNet, SwinUNETR and MedCLIP-SAMv2; to improve the segmentation accuracy of complex atrial structures and maintain topological consistency, the invention introduces a topological loss function based on persistent homology to constrain the deep learning segmentation network to ensure the continuity of the atrial wall and anatomical rationality, and to minimize misclassification or fracture phenomenon; the segmented left heart main anatomical regions include left atrium, left atrial wall, left atrial appendage, pulmonary vein and aorta; the mathematical expression of the topological loss function is as follows: wherein and denote the persistence bar charts of the segmented image S and the ground truth G, respectively, and p and q are points in the persistence bar charts of S and G, respectively.

4. The left atrial wall thickness evaluation and intraoperative navigation method according to claim 3, in step 3, based on the segmentation result obtained in step 2, a modified Marching Cubes three-dimensional reconstruction algorithm is used to construct a three-dimensional model of the left atrial wall, and the inner and outer surfaces are separated; through geometric analysis and region growing method, four types of quantifiable geometric parameters of the surface normal vector, average curvature, Gaussian curvature and principal curvature of the three-dimensional model are calculated and processed comprehensively; the inner and outer surfaces are saved in the form of independent grid or point cloud respectively to ensure accurate matching of the corresponding points of the inner and outer surfaces during subsequent wall thickness calculation, and the mathematical expression is as follows: wherein f(p) is a scalar function of the model surface point, is the gradient of the point, the normal vector is the unitized vector of the gradient.

5. The left atrial wall thickness evaluation and intraoperative navigation method according to any one of claims 1-4, in step 4, the wall thickness calculation and regional risk assessment specifically includes the following steps: Step 4.1: To achieve accurate local analysis, the left atrial wall is automatically divided into the following six anatomical regions: upper wall, posterior wall, septal wall, anterior wall, left lateral wall and pulmonary vein interstitial region; in the automatic region identification process, first, the division is made according to the pre-defined anatomical marker points or the partition model based on deep learning, and necessary manual fine-tuning is allowed according to actual needs to ensure that the partition boundary meets the clinical judgment standard. Step 4.2: For each anatomical region, the shortest distance algorithm between the inner and outer surfaces is used to calculate the average wall thickness; At the same time, left atrial volume index (LAVI) and left atrial wall signal intensity normalization index (IIR) are introduced for multi-dimensional risk assessment; LAVI reflects the ratio of left atrial volume to body surface area, which is combined with average thickness analysis, wall thickness change rate and local expansion coefficient to evaluate the local or overall risk level; based on the above parameters, each region is automatically scored, and a warning is issued for high-risk areas to prompt clinicians to pay special attention.

6. The left atrial wall thickness assessment and intraoperative navigation method of claim 5, in the step 5, the visualization mapping and interaction method comprises: Based on the wall thickness and risk classification data obtained in step 4, the numerical values are mapped to a unified color space to realize the intuitive visualization of left atrial wall thickness and risk classification; using segmented color mapping algorithm and OpenGL three-dimensional rendering technology, interactive model operation is realized to support model rotation, scaling and translation, and users are allowed to select, label and query information of local areas in real time; users can directly obtain the thickness parameters and risk prompts of the target area in the three-dimensional visualization interface to provide visual auxiliary support for clinical decision-making; The mathematical model of color mapping is as follows: where each Color i corresponds to an RGB = (r i ,g i ,b i ) color value.

7. The left atrial wall thickness evaluation and intraoperative navigation method of claim 6, in the step 6, the intraoperative navigation and dynamic adjustment method comprises: Through the establishment of a data interface, real-time communication is realized between the HL7 / FHIR protocol or hospital information system (HIS) standard and the electrophysiological mapping platform to realize the real-time fusion of catheter position and local electrical activity characteristics; when the catheter tip or ablation instrument enters the high-risk area, the system will give a high-light warning in the three-dimensional visualization interface and prompt the operator to adjust the path if necessary.

8. The left atrial wall thickness evaluation and intraoperative navigation method according to claim 7, wherein in step 6, if equipped with X-ray fluoroscopy, intracardiac echocardiography (ICE) or transesophageal echocardiography (TEE) during the operation, the pose offset of the intraoperative image can be automatically corrected based on the non-rigid registration method to ensure that the navigation view is consistent with the patient's anatomy.

9. The left atrial wall thickness evaluation and intraoperative navigation method according to claim 8, wherein in step 6, the improved A* or Dijkstra path planning algorithm can be used to provide the surgeon with alternative ablation path recommendations based on real-time data, and allow the operator to perform manual confirmation and one-key synchronization to achieve closed-loop navigation and dynamic risk control.

10. An artificial intelligence-based left atrial wall thickness evaluation and intraoperative navigation system for implementing the method of claims 1-9, comprising: a data acquisition and preprocessing module for acquiring multi-modal medical image data and preprocessing it to improve data compatibility and calculation accuracy; an image segmentation module for automatically identifying the left atrium and related anatomical structures when analyzing the preprocessed medical image data, and optimizing the segmentation results through topological consistency constraints; a three-dimensional modeling and structure analysis module for constructing a three-dimensional structure model of the left atrium based on the segmentation results, and analyzing and optimizing its surface, using geometric feature extraction and spatial separation methods to distinguish key anatomical regions to support subsequent wall thickness calculation and visual analysis; a wall thickness calculation and risk assessment module for measuring the thickness of the three-dimensional structure of the left atrium and partitioning the region, combining various anatomical and physiological parameters to assess potential high-risk areas and provide quantitative analysis to support individualized treatment strategies; a visualization and interaction module for mapping the left atrial wall thickness and risk information to a unified visualization model based on the calculation results, using interactive three-dimensional rendering technology to provide intuitive visual feedback and support user real-time query and annotation of target areas; and an intraoperative navigation and dynamic adjustment module for dynamically updating the left atrial model in combination with real-time image and position information during the operation, and providing personalized navigation solutions and risk warnings. ​

Images