System for constructing a model of a heart chamber measurement and storage medium

Abstract

The application provides a system for constructing a heart chamber measurement model and a storage medium, and the heart chamber precision measurement model comprises an image segmentation and three-dimensional reconstruction module, a spatial distance calculation module, a dynamic function parameter evaluation module and a calcification artifact compensation module. The application breaks through the limitation of low accuracy of traditional static straight-line measurement, and through the image segmentation and three-dimensional reconstruction technology driven by deep learning, the model can realize sub-millimeter level measurement accuracy, and construct an accurate three-dimensional model of the heart chamber. The application can assist in the size selection and path planning of interventional instruments, thereby effectively improving the efficiency and safety of structural heart disease surgery.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted medical image processing technology, and in particular to a system for constructing and storing a cardiac chamber measurement model. Background Technology

[0002] Precise measurement of cardiac chambers plays an indispensable role in the accurate planning of interventional procedures for structural heart disease, directly affecting the safety and effectiveness of the surgery. With the development of multimodal image fusion technology, cardiac image analysis has made some progress. For example, Li et al. pointed out that multimodal cardiac imaging, by fusing anatomical, morphological, and functional information, can significantly improve the diagnostic accuracy and interventional treatment outcomes of cardiovascular diseases (Med. Image Anal, 2023).

[0003] However, existing technologies still have significant shortcomings: most current chamber measurements rely on simple geometric models (such as the straight-line method), which are difficult to adapt to complex situations such as curved channels, thin-walled structures, and calcification artifacts; secondly, the measurement accuracy of fine structures (such as the left atrial appendage opening, valve annulus margins, and chamber grooves) is generally low, and the influence of dynamic motion (such as heartbeat and respiration) on images is not adequately compensated, often resulting in measurement errors as high as 2-3 mm in clinical applications. Although some models have achieved static image segmentation and fusion, the overall process of applying real-time measurement and multimodal image co-registration to the accurate measurement of cardiac chambers is still imperfect. Summary of the Invention

[0004] The purpose of this invention is to provide a system for constructing and storing a cardiac chamber measurement model.

[0005] To address the above problems, this invention provides a system for constructing a cardiac chamber measurement model, comprising:

[0006] The acquisition unit is used to establish a precise measurement dataset of cardiac chambers, which includes: original multimodal cardiac images, fine structural annotations matched with the original multimodal cardiac images, and clinical regurgitation grading;

[0007] The modeling unit is used to establish an initial cardiac chamber measurement model, which includes: an image segmentation and three-dimensional reconstruction module, a spatial distance calculation module, a dynamic functional parameter evaluation module, and a calcification artifact compensation module.

[0008] The training unit is used to train the initial heart chamber measurement model based on the heart chamber precision measurement dataset to obtain a trained heart chamber precision model.

[0009] The multimodal images of the patient to be tested are input into the trained accurate measurement model of the heart chambers to obtain the measurement data of the heart chambers of the patient to be tested.

[0010] Furthermore, in the above system, the original multimodal cardiac images include: preoperative images, intraoperative images, and postoperative images; the formats of the original multimodal cardiac images include: CTA, CT, MRA, TEE, DSA, Xray, and CBCT.

[0011] Furthermore, in the above system, the acquisition unit is used to generate initial annotations from the original multimodal cardiac images according to the preset graded anatomical annotation specifications and using a semi-automatic pre-annotation tool; then, the initial annotations are refined using a doctor's drawing correction tool to obtain an image annotation draft with preliminary anatomical structure labels.

[0012] Based on the image annotation draft with preliminary anatomical labels, the Hounsfield threshold slider is used to distinguish soft plaques and calcified tissues in the original multimodal cardiac images and supplement the corresponding pathological tissue annotations. The pathological tissue annotations are integrated into the image annotation draft with preliminary anatomical structure labels to obtain an image annotation draft containing anatomical structure and pathological tissue annotations.

[0013] Multiple cardiac imaging experts provided image annotation drafts, each based on anatomical structures and pathological tissues. These drafts were used to fully annotate the original multimodal cardiac images, resulting in three independent annotation results. The STAPLE algorithm was used to weightedly fuse overlapping and dissimilar regions from each independent annotation result, generating an initial fused annotation result. The Dice similarity coefficient between the annotations from different experts was calculated to assess the degree of difference, resulting in a Dice coefficient evaluation report for each region. If the Dice coefficient of the dissimilar region in the evaluation report is greater than or equal to a preset threshold (e.g., 0.85), the initial fused annotation result is directly used as the unified annotation draft with consensus. If the Dice coefficient of the dissimilar region in the evaluation report is less than the preset threshold (e.g., 0.85), expert arbitration and negotiation are initiated. Based on the anatomical features and pathological manifestations of the original multimodal cardiac images, the final annotation boundaries and labels in the initial fused annotation result are corrected to obtain a unified annotation draft with consensus.

[0014] Using preset verification indicators, the quality of the unified annotation draft that has reached a consensus is verified; based on the annotation quality, the unified annotation draft that has passed the verification and reached a consensus is used as fine-structure annotation.

[0015] Furthermore, in the aforementioned system, the trained image segmentation and 3D reconstruction module is used to call the trained 3D U-Net variant model to perform pixel-level automatic segmentation on the preoperative original multimodal cardiac images of the patient to be examined, identify and delineate the contour boundaries of complex anatomical structures, and obtain images with segmentation masks for each anatomical structure. Based on the spatial coordinate information of the images with anatomical structure segmentation masks, the 2D image layers of the images with anatomical structure segmentation masks are fused into a 3D model through a 3D reconstruction algorithm, and the accuracy is calibrated to the sub-millimeter level to obtain a sub-millimeter accuracy 3D initial cardiac model. An anatomical perception loss function is introduced to iteratively optimize the thin-walled structural contours of the 3D initial cardiac model, correct the segmentation error, and obtain a 3D cardiac chamber model with fine anatomical labels.

[0016] Furthermore, in the above system, the trained spatial distance calculation module is used to locate the spatial coordinates and contour data of various key anatomical structures based on a three-dimensional cardiac chamber model with fine anatomical labels; based on the spatial coordinates and contour data of various key anatomical structures, quantitative measurement results of each key anatomical structure are obtained; when calcification is identified in a key anatomical structure, the anatomical structure containing calcification is measured first to obtain the initial measurement data of the calcified region.

[0017] Furthermore, in the above system, the trained spatial distance calculation module is used for:

[0018] Based on the coordinate and contour data of the nonlinear structure in the three-dimensional cardiac chamber model, the central axis of the nonlinear structure is automatically identified; the nonlinear structure is segmented along the central axis; and the effective anchoring length of each segment is calculated using a geodesic algorithm to obtain the segmented effective anchoring length data of the nonlinear structure.

[0019] For measuring the valve annulus circumference of a saddle-shaped curved surface structure in a 3D heart model: Based on the coordinate and contour data of a non-saddle-shaped curved surface structure in the 3D heart model, the edge contour of the valve annulus surface is extracted; based on the edge contour of the valve annulus surface, a geodesic algorithm is used to fit the closed path of the surface edge; the length of the closed path is calculated to obtain the valve annulus circumference, and the valve annulus circumference measurement value is obtained.

[0020] For measuring the height of coronary artery ostia in vascular structures in a 3D heart model: Based on the relative spatial position data of the vascular structure ostia and the corresponding chamber in the 3D heart model, the center point of the vascular structure ostia is located; using the normal distance algorithm, the vertical height from the center point of the vascular structure ostia to the wall of the corresponding chamber is measured as the height of the vascular structure ostia.

[0021] Furthermore, in the aforementioned system, the trained dynamic functional parameter evaluation module is used to automatically identify key phases of the cardiac cycle in the intraoperative or postoperative cardiac cycle time-series images of the patient under test, based on the three-dimensional heart model generated by the image segmentation and three-dimensional reconstruction module and the quantitative measurement results of each key anatomical structure output by the spatial distance calculation module. It spatially matches the valve morphology under the key phase of the cardiac cycle with the quantitative measurement results of each key anatomical structure to locate the relative position of the valve within the anatomical structure, serving as the valve morphology data for the corresponding key phase of the cardiac cycle. Based on the valve morphology data under the key phase of the cardiac cycle and the quantitative measurement results of each key anatomical structure, the maximum leaflet separation point is extracted. Using the leaflet anatomical length in the quantitative measurement results of each key anatomical structure as a reference scale, the ratio of the spatial distance of the maximum leaflet separation point to the leaflet anatomical length is calculated to determine the degree of dynamic offset of the maximum leaflet separation point relative to the leaflet itself, serving as valve dynamic feature data. Based on the valve dynamic feature data and the quantitative measurement results of each key anatomical structure, the Hausdorff method is used to... The width of the valve leaflet tear was calculated by distance, and the relative proportion of the tear was converted from the circumference of the valve annulus in the quantitative measurement results of each key anatomical structure. The degree of valve regurgitation was analyzed by time-series variation curves, and the spatial range of regurgitation was determined by combining the anatomical length of the leaflet. The relative proportion of the tear and the spatial range of regurgitation were used as quantitative indicators of valve function.

[0022] Furthermore, in the aforementioned system, the trained calcification artifact compensation module is used to automatically detect calcification areas in the preoperative CT images of the patient under test with CT values ​​greater than a preset CT threshold based on the initial measurement data of the calcification area output by the spatial distance calculation module, thereby obtaining the localization result of the calcification artifact area. Based on the localization result of the calcification artifact area, thickness deviation compensation is performed on the initial measurement data of the calcification area to obtain the calibrated measurement data of the calcification area.

[0023] Furthermore, in the above system, the training unit is used for:

[0024] The image segmentation and 3D reconstruction module was trained in a supervised manner using preoperative raw multimodal cardiac images from the precise cardiac chamber measurement dataset. The fine structure annotations in the precise cardiac chamber measurement dataset were used as ground truth, and Dice and cross-entropy loss were used to optimize the segmentation effect, resulting in the trained image segmentation and 3D reconstruction module.

[0025] Freeze the parameters of the image segmentation and 3D reconstruction module after training, and train the spatial distance calculation module based on the 3D heart chamber model output by the image segmentation and 3D reconstruction module to minimize the L1 and / or L2 errors between the quantitative measurement results of each key anatomical structure predicted by the spatial distance calculation module and the accurate measurement dataset of the heart chamber.

[0026] Based on the quantitative measurement results of key anatomical structures output by the spatial distance calculation module, combined with the intraoperative or postoperative cardiac cycle time series images in the precise measurement dataset of cardiac chambers and the clinical regurgitation grading in the precise measurement dataset of cardiac chambers, the dynamic functional parameter assessment module is trained so that the valvular function quantitative indicators output by the dynamic functional parameter assessment module are consistent with the clinical evaluation.

[0027] Based on the initial measurement data of calcified regions output by the spatial distance calculation module and the precise measurement data of calcified regions in the fine structural annotation of the precise cardiac chamber measurement dataset, the bias of the spatial distance calculation module is calculated to obtain the structural size deviation value caused by calcification artifacts. Based on the valvular function quantification index of the dynamic functional parameter assessment module and the valve clinical truth index in the fine structural annotation of the precise cardiac chamber measurement dataset, the bias of the dynamic functional parameter assessment module is calculated to obtain the functional assessment deviation value transmitted by the structural size deviation. The structural size deviation value and the functional assessment deviation value are integrated to locate the correspondence between the source of the deviation and the calcified region, resulting in a calcification-related measurement deviation dataset. The preoperative CT in the original multimodal cardiac images in the precise cardiac chamber measurement dataset is identified. The calcified regions in the image are located. The calcified region localization results are obtained. The calcification-related measurement deviation dataset and the calcified region localization results are correlated to establish a mapping model between calcified region features and measurement deviations. With the goal of reducing system errors, the parameters of the calcification artifact compensation module are adjusted. The adjusted parameters of the calcification artifact compensation module are used to compensate the initial measurement data of the calcified region. The deviation between the compensated data and the accurate measurement data of the calcified region is verified. Finally, the optimized calcification artifact compensation module is output.

[0028] According to another aspect of the invention, a computer-readable storage medium is also provided, having stored thereon computer-executable instructions, wherein when executed by a processor, the computer-executable instructions cause the processor to: execute the construction system for the cardiac chamber measurement model as described in any of the preceding claims.

[0029] Compared with existing technologies, this invention provides a system for constructing a cardiac chamber measurement model. The precise cardiac chamber measurement model includes: an image segmentation and 3D reconstruction module, a spatial distance calculation module, a dynamic functional parameter evaluation module, and a calcification artifact compensation module. This invention overcomes the limitations of low accuracy in traditional static straight-line measurements. Through deep learning-driven image segmentation and 3D reconstruction technology, the model can achieve sub-millimeter-level measurement accuracy, constructing a precise 3D model of the cardiac chamber. This invention can assist in the size selection and path planning of interventional devices, thereby effectively improving the efficiency and safety of structural heart disease surgery. Attached Figure Description

[0030] Figure 1 A schematic diagram of a system for constructing a cardiac chamber measurement model according to an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the framework of a system for constructing a cardiac chamber measurement model in one example of the present invention;

[0032] Figure 3 A schematic diagram illustrating the principle of precise measurement of cardiac chambers according to an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the structure of a cardiac chamber measurement model training device provided in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the structure of a precise cardiac chamber measurement device provided in an embodiment of the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various implementation methods of the present invention will be described in detail below with reference to the embodiments. Experimental methods in the embodiments that do not specify specific conditions are generally operated under conventional conditions, such as those described in textbooks and experimental guides, or according to the conditions recommended by the manufacturer and the accompanying software.

[0037] like Figure 1 As shown, the cardiac chamber precision measurement model construction system provided in this embodiment of the invention includes:

[0038] Step 101, the acquisition unit is used to establish a precise measurement dataset of cardiac chambers, which includes: original multimodal cardiac images, fine structural annotations matched with the original multimodal cardiac images, and clinical regurgitation grading;

[0039] Here, the original multimodal cardiac images may include:

[0040] 1) Preoperative imaging:

[0041] Computed tomography angiography (CTA) or enhanced CT: used for fine anatomical localization of cardiac structures;

[0042] Magnetic resonance angiography (MRA) or cardiac MRI: provides multidimensional information on cardiac morphology and function;

[0043] Echocardiography: includes conventional ultrasound and transesophageal echocardiography (TEE) to help assess valvular function;

[0044] 2) Intraoperative imaging:

[0045] Dynamic X-ray: Real-time tracking of the position and operation process of interventional instruments;

[0046] Digital subtraction angiography (DSA) or cone-beam computed tomography (CBCT): used for precise intraoperative localization and surgical path planning;

[0047] 3) Postoperative imaging:

[0048] Follow-up imaging similar to that before surgery (CTA, ultrasound, etc.): used to assess surgical outcomes and changes in cardiac structure.

[0049] Specifically, in step 101, the acquired raw multimodal cardiac images may include: preoperative CTA / CBCT and MRA data, intraoperative dynamic X-ray and TEE data, and postoperative multimodal data. Acquisition methods include using public databases and self-built databases. Major public databases may include:

[0050] ①CAMUS dataset (500 echocardiograms): Its annotations cover the basic structures of the left ventricular endocardium, left ventricular epicardium, and left atrium. However, it cannot support the analysis of valve leaflet motion trajectory and valve function;

[0051] ②ASOCA2020 dataset (60 CTA cases): This dataset annotates seven major cardiac structures, including the left / right ventricular chambers, left / right atrial chambers, left ventricular myocardium, ascending aorta, and pulmonary artery, but does not include fine-grained annotations of any valves or left atrial appendage. This dataset focuses on coronary artery analysis.

[0052] ③MM-WHS 2017 dataset (60 CT scans): The annotation range is similar to ASOCA, covering 7 major cardiac substructures, with a focus on cardiac chamber segmentation.

[0053] Here, public databases provide basic cardiac structural annotations, such as core structures like the left ventricle, atrium, and myocardium, serving as the foundational data source for datasets and supporting basic training of models for segmentation and measurement of the heart's main chambers. However, a shortcoming of public databases is the lack of fine-grained valve structural annotations, which prevents them from supporting valve-related functional analysis.

[0054] Specifically, the public databases used to obtain annotation information in step 101 generally suffer from a lack of detailed annotations of valvular structures, necessitating the creation of a self-built database for further supplementation. This invention can rely on a clinical team to build a multimodal registration system database based on imaging data of patients who underwent interventional treatment for structural heart disease at a certain hospital from 2018 to 2024, including approximately 800 cases. Among these, approximately 320 cases underwent transcatheter aortic valve replacement (TAVI), approximately 210 cases underwent transcatheter mitral and tricuspid valve edge-to-edge repair (TEER), approximately 190 cases underwent transcatheter left atrial appendage occlusion (LAAO), and approximately 80 cases underwent other structural heart disease interventions (such as PFO / ASD occlusion, tricuspid valve intervention, etc.).

[0055] For each case, preoperative CTA or enhanced CT (slice thickness approximately 0.5–0.625 mm), cardiac MRI and transthoracic / transesophageal ultrasound when necessary, intraoperative dynamic X-ray fluoroscopy sequences, rotating DSA / CBCT, 3D TEE volume data, and postoperative follow-up CTA / ultrasound and electrocardiograms were collected as multimodal images. Under ethical review and with informed consent from patients, all DICOM data underwent anonymization in a trusted environment, removing private information such as names and hospital numbers, retaining only random numbers and examination times. Standardized data queues were constructed through uniform format conversion, resampling, registration, and intensity normalization to provide a data foundation for subsequent fine-scale structural annotation and model training.

[0056] Self-built databases can compensate for the deficiencies of public databases, supplementing labeled data such as "detailed valve structure and interventional cases of different structural heart diseases", allowing the dataset to cover more refined needs such as valve motion and valve function assessment, thereby supporting the model to achieve more accurate measurements of heart chambers (including valves).

[0057] The detailed structural annotations consist of basic annotations from public databases and supplementary annotations from a self-built database. This unified and precise annotation set is the result of processing and optimization using a standardized and efficient annotation system. The public databases primarily provide preoperative basic images such as CT, MRI, and ultrasound, while the self-built databases supplement intraoperative dynamic images, postoperative follow-up images, and full-process images for more structural heart disease interventional cases. Specifically, the detailed structural annotations can include two core categories of content:

[0058] 1) Basic cardiac structure annotations (derived from public libraries and standardized and optimized), covering the anatomical contours of the core cardiac chambers and major blood vessels, ensuring the model's ability to recognize basic structures:

[0059] Chambers: left ventricular endocardium or epicardium, left atrium, right ventricular blood chamber, right atrial blood chamber, left ventricular myocardium;

[0060] Vascular vessels: ascending aorta, pulmonary artery;

[0061] 2) Detailed anatomical and pathological structure annotations (derived from a self-built database and standardized and optimized) compensate for the shortcomings of public databases and support the needs of valve-related measurements and interventional surgery evaluation:

[0062] Valve and its associated structures: valve annulus, valve leaflets, mitral valve occlusion region, dynamic valve motion trajectory;

[0063] Special structures: left atrial appendage and opening details, coronary sinus.

[0064] In this specific embodiment, fine structural annotations matching the original multimodal cardiac images are obtained, including:

[0065] Step 1011: Generate initial annotations for the original multimodal cardiac images (CTA / MRA, etc.) according to the preset graded anatomical annotation specifications (organ level / subregion level) using a semi-automatic pre-annotation tool; then refine the initial annotations using a doctor's drawing correction tool to obtain an image annotation draft with preliminary anatomical structure labels.

[0066] Step 1012: Based on the image annotation draft with preliminary anatomical labels, use the Hounsfield threshold slider (400-1000HU) to distinguish soft plaques and calcified tissues in the original multimodal cardiac images, and supplement the corresponding pathological tissue annotations. Integrate the pathological tissue annotations into the image annotation draft with preliminary anatomical structure labels to obtain an image annotation draft containing anatomical structure and pathological tissue annotations.

[0067] Here, we can first identify the regions corresponding to soft plaques and calcified tissues in the original CT images using the Hounsfield threshold slider (400-1000 HU range); then, we can add pathological tissue annotations for these identified regions, and finally integrate these pathological tissue annotations into the previous anatomical structure annotations. Simply put, it involves locating pathological tissues on the original images and labeling them, so that the annotations simultaneously include anatomical structure and pathological tissue information.

[0068] Step 1013: Obtain image annotation drafts from multiple experts (e.g., 3 cardiac imaging experts), each annotating the original multimodal cardiac images based on anatomical structures and pathological tissues. These experts will provide complete annotations of anatomical structures (chambers, valves, etc.) and pathological tissues (soft plaques, calcifications), resulting in 3 independent annotation results. Using the STAPLE algorithm (Statistical Parametric Map Label Fusion Algorithm), the overlapping and differing regions of these 3 independent annotation results are weighted and fused to generate an initial fused annotation result. Simultaneously, the Dice similarity coefficient between the annotations from different experts is calculated to determine the degree of difference, resulting in a Dice coefficient evaluation report for each region's annotation differences. If the Dice coefficient of the differing region in the evaluation report is greater than or equal to a preset threshold (e.g., 0.85), the initial fused annotation result is directly used as the unified annotation draft for consensus. If the Dice coefficient of the differing region in the evaluation report is less than the preset threshold (e.g., 0.85), expert arbitration and consultation are initiated: multiple experts (e.g., 3 experts) are organized to participate in the process. Experts jointly assessed the discrepancies in the regions and, based on the anatomical features and pathological manifestations of the original multimodal cardiac images, revised the final annotation boundaries and labels in the initial fusion annotation results, resulting in a unified annotation draft that reached a consensus, thereby eliminating subjective bias and meeting accuracy requirements.

[0069] Here, three experts independently completed the annotation work based on the original image and the initial annotation draft. Each expert would independently annotate the image based on their own professional experience, including anatomical structures and pathological tissues. Then, the STAPLE algorithm was used to fuse the results of the three independent annotations. This can avoid the subjective bias of a single expert and ensure the objectivity of the annotation.

[0070] Step 1014: Use preset verification indicators such as left atrial appendage opening ICC ≥ 0.92 and mitral valve occlusion distance error ≤ 0.6mm to verify the annotation quality of the consensus-reaching unified annotation draft; based on the annotation quality, the consensus-reaching unified annotation draft that has passed verification is used as fine structure annotation.

[0071] Here, from the unified annotation draft that reached a consensus after expert arbitration, the quantitative parameters of key anatomical structures in the unified annotation draft are extracted and verified using preset quality control indicators. The core verification content may include:

[0072] The intraclass correlation coefficient (ICC) of the left atrial appendage opening was ≥0.92, verifying the consistency of the annotation of this structure in different annotation stages;

[0073] The measurement error of the mitral valve occlusion distance is ≤0.6mm, verifying the accuracy of this parameter labeling;

[0074] Other supporting quality control indicators (such as the deviation in valve annulus circumference measurement, the fit of the calcified area marking, etc.).

[0075] It can output: a quality inspection report with a "compliant or non-compliant" result. If non-compliant, it will mark the specific problem area.

[0076] In addition, the entire process correction trajectory from step 1011 to step 1014 can be recorded.

[0077] By combining the annotation quality verification report with the intermediate products of the entire annotation process (preliminary annotation draft, 3 independent expert annotation files, initial fusion annotation results, and expert arbitration records), the correction actions at each step can be fully recorded according to the annotation process sequence. This includes the adjustment records of semi-automatic pre-annotation, the differences in independent expert annotation, the STAPLE algorithm fusion parameters, the decision basis for expert arbitration, and the secondary correction scheme for areas that do not meet quality control standards. This will result in the final fine-grained structure annotation file with correction trajectories.

[0078] Here, the detailed structural annotation can be achieved using a systematic approach that integrates expert experience and human-computer collaboration to form a standardized and efficient annotation system: establishing hierarchical anatomical annotation standards (organ-level and subregion-level annotation of left atrial appendage, valve annulus, valve leaflets, etc.), combined with semi-automatic pre-annotation and physician stroke correction tools; setting a Hounsfield threshold slider (400-1000HU) to help distinguish between soft plaques and calcifications; each case is independently annotated by 3 cardiac imaging experts and fused using the STAPLE algorithm, with arbitration initiated for discrepancies (Dice < 0.85), and approval only after consultation and confirmation; setting quality control indicators such as left atrial appendage ostium ICC ≥ 0.92 and mitral valve occlusion distance error ≤ 0.6 mm, and recording all correction trajectories in the annotation traceability database.

[0079] Here, basic annotations from public databases plus fine annotations from self-built databases are the sources of the dataset content, while the standardized and efficient annotation system is the process of processing and optimizing these annotated contents. It is equivalent to first gathering basic and finely annotated materials, and then using this system to polish the materials into high-quality, standardized annotated data.

[0080] Whether it's the basic structural annotation of public databases or the valve and interventional case annotations supplemented by self-built databases, all will undergo the same processing steps: using hierarchical standards to unify annotation logic (e.g., left atrial annotations in public databases and left atrial appendage subregion annotations in self-built databases must conform to the same set of anatomical definitions); using semi-automatic tools + expert correction to optimize annotation efficiency and accuracy; and using expert fusion + quality control to eliminate errors and ensure quality. The final output is a standardized dataset with comprehensive content coverage (basic + detailed) and meeting quality standards (accurate + consistent), rather than a patchwork of fragmented public database annotations and self-built database annotations.

[0081] The dataset in step 101 is a multi-dimensional labeled dataset, which includes not only preoperative CTA / MRA images, intraoperative / postoperative cardiac cycle time series images, and anatomical parameters measured manually by experts, i.e. finely labeled data, but also corresponding clinical diagnostic labeling information—clinical regurgitation grading is the core diagnostic labeling content.

[0082] Clinical regurgitation grading is not generated by a model, but is an authoritative evaluation (e.g., mild, moderate, severe regurgitation) given by cardiologists based on the patient's imaging findings, intraoperative observations, and postoperative follow-up results, according to clinical guidelines (such as the "Guidelines for the Diagnosis and Treatment of Valvular Heart Disease"). This evaluation serves as the ground truth label during the training of the dynamic functional parameter assessment module, and is used to ensure that the indicators such as the tear width output by the model are consistent with the clinical evaluation.

[0083] The precise cardiac chamber measurement dataset in step 101 is a complete set of “images + anatomical measurement annotations + clinical diagnostic annotations”, with clinical reflux grading being a part of the clinical diagnostic annotations.

[0084] Step 102, Modeling Unit, used to establish an initial cardiac chamber measurement model, the initial cardiac chamber measurement model includes: image segmentation and three-dimensional reconstruction module, spatial distance calculation module, dynamic functional parameter evaluation module and calcification artifact compensation module;

[0085] It should be noted that the initial accurate measurement model of the heart chambers in step 102 can be considered as a model that has not yet been trained.

[0086] Preferably, the trained image segmentation and 3D reconstruction module is used to call the trained 3D U-Net variant model to perform pixel-level automatic segmentation on the original multimodal cardiac images such as preoperative CTA / MRA of the patient to be examined. This accurately identifies and delineates the contours of complex anatomical structures such as dynamic valves, left atrial appendage, and millimeter-level coronary sinuses, resulting in 2D or 3D images with segmentation masks for each anatomical structure. The segmentation mask is a region label marking different anatomical structures. Based on the spatial coordinate information of the image with the anatomical structure segmentation mask, the 2D image layers of the image with the anatomical structure segmentation mask are fused into a 3D model using a 3D reconstruction algorithm (such as volume rendering or surface rendering), and the accuracy is calibrated to the sub-millimeter level to ensure the spatial scale accuracy of the model, resulting in a sub-millimeter level image. A three-dimensional initial model of the heart with a precision of meters is obtained. The initial model includes the three-dimensional contours of various anatomical structures, but the segmentation accuracy of thin-walled structures such as valve annulus and valve leaflets needs to be optimized. To address the segmentation deviation problem of thin-walled structures such as valve annulus and valve leaflets, an anatomical perception loss function is introduced (this function combines prior anatomical knowledge to constrain the model) to iteratively optimize the contours of thin-walled structures in the initial three-dimensional heart model, correcting the segmentation error and obtaining a three-dimensional heart chamber model with fine anatomical labels. In the three-dimensional heart chamber model with fine anatomical labels, the segmentation accuracy of thin-walled structures is improved, meeting the requirements of subsequent measurements.

[0087] Here, the trained image segmentation and 3D reconstruction module is used to convert the acquired preoperative multimodal cardiac images (unlabeled images) of the patient into a 3D cardiac chamber model with detailed anatomical labels. For example... Figure 2 As shown, complex anatomical structures in the original multimodal cardiac images, including dynamic valves, meandering left atrial appendages, and millimeter-scale coronary sinuses, are automatically segmented using a 3D U-Net variant to generate a three-dimensional cardiac model with sub-millimeter precision. Furthermore, an anatomical perception loss function is introduced to differentially optimize thin-walled structures such as valve annulus and leaflets, providing a foundation for subsequent accurate spatial distance calculations.

[0088] Once the image segmentation and 3D reconstruction module is trained, inputting an unlabeled preoperative CTA / MRA image of the patient to be tested will automatically output a final 3D model of the heart chambers with detailed anatomical labels.

[0089] Preferably, the trained spatial distance calculation module is used to locate the spatial coordinates and contour data of various key anatomical structures (left atrial appendage, valve annulus, coronary artery ostia, etc.) based on a three-dimensional cardiac chamber model with fine anatomical labels. Based on the spatial coordinates and contour data of various key anatomical structures (left atrial appendage, valve annulus, coronary artery ostia, etc.), quantitative measurement results of each key anatomical structure are obtained to effectively anchor length, valve annulus circumference, coronary artery ostium height, etc. In addition, when calcification is identified in key anatomical structures (such as valve annulus, left atrial appendage), the anatomical structure containing calcification is measured first to obtain initial measurement data of the calcified region. The initial measurement data of the calcified region at this time includes: the diameter, length, circumference and other parameters of the calcified region are initial values ​​containing calcification artifact interference. For example, calcification makes the structure look "thicker", resulting in larger measurement values. The deviation needs to be corrected by the calcification artifact compensation module.

[0090] Here, in step 102, the spatial distance calculation module completes the precise quantitative measurement of various key anatomical structures. For example... Figure 2 As shown, based on the generated three-dimensional heart model, an innovative algorithm is used to calculate the spatial distance between heart chambers. This innovative algorithm includes: calculating the effective anchoring length of nonlinear structures (such as the left atrial appendage) segmentally along the central axis; measuring the annular circumference of the saddle-shaped surface; and measuring the height of the coronary artery ostia; and using geodesic and normal distance algorithms to improve measurement accuracy.

[0091] Even better, based on the spatial coordinates and contour data of various key anatomical structures (left atrial appendage, valve annulus, coronary artery ostia, etc.), quantitative measurement results of each key anatomical structure are obtained, including:

[0092] Step 1021: Calculate the effective anchoring length for nonlinear structures (such as the left atrial appendage): Based on the coordinates and contour data of the nonlinear structure in the three-dimensional cardiac chamber model, automatically identify the central axis of the nonlinear structure; divide the nonlinear structure into segments along the central axis; combine the geodesic algorithm (to calculate the shortest path between two points on the surface) to calculate the effective anchoring length of each segment, and obtain the segmented effective anchoring length data of the nonlinear structure.

[0093] For example, to calculate the effective anchoring length for nonlinear structures (such as the left atrial appendage): based on the three-dimensional contour data of the left atrial appendage in the three-dimensional heart model (including spatial coordinates and surface morphology), the central axis of the left atrial appendage is automatically identified; the left atrial appendage is segmented along the central axis; combined with the geodesic algorithm (to calculate the shortest path between two points on the surface), the effective anchoring length of each segment is calculated to obtain the segmented effective anchoring length data of the left atrial appendage.

[0094] Step 1022: Measure the circumference of the valve annulus for a saddle-shaped surface structure (such as a valve annulus) in a 3D heart model: Based on the coordinate and contour data of a non-saddle-shaped surface structure in the 3D heart model, extract the edge contour of the valve annulus surface; based on the edge contour of the valve annulus surface, fit the closed path of the surface edge using a geodesic algorithm; calculate the length of the closed path to obtain the circumference of the valve annulus, and obtain the measured value of the valve annulus circumference.

[0095] For example, based on the saddle-shaped surface contour data of the valve annulus in a three-dimensional heart model, the edge contour of the valve annulus surface is extracted; based on the edge contour of the valve annulus surface, the geodesic algorithm fits the closed path of the surface edge; the length of the closed path is calculated to obtain the circumference of the valve annulus, and the circumference measurement value of the valve annulus is obtained.

[0096] Step 1023: Measure the height of the coronary artery ostium for vascular structures (such as coronary arteries) in the 3D heart model: Based on the relative spatial position data of the vascular structure ostium and the corresponding chamber (such as the left ventricle) in the 3D heart model, locate the center point of the vascular structure ostium; use the normal distance algorithm (to calculate the vertical distance from the point to the surface) to measure the vertical height from the center point of the vascular structure ostium to the wall of the corresponding chamber, and use this as the measured value of the height of the vascular structure ostium.

[0097] For example, based on the relative spatial position data of the coronary artery ostium and the corresponding chamber (such as the left ventricle) in the three-dimensional heart model, the center point of the coronary artery ostium is located; using the normal distance algorithm (to calculate the vertical distance from the point to the surface), the vertical height from the center point of the coronary artery ostium to the wall of the corresponding chamber is measured as the height measurement value of the coronary artery ostium.

[0098] Preferably, the trained dynamic functional parameter assessment module is used to automatically identify intraoperative or postoperative cardiac cycle time-series images (such as dynamic ultrasound, X-ray, etc.) of the patient based on the three-dimensional cardiac model generated by the image segmentation and three-dimensional reconstruction module and the quantitative measurement results of key anatomical structures output by the spatial distance calculation module. In the cardiac cycle key phases (such as mid-systole in mitral regurgitation) of a linear sequence, the valve morphology at these key phases is spatially matched with the quantitative measurements of key anatomical structures such as the annular circumference to locate the relative position of the valve within the anatomical structures, serving as the valve morphology data for the corresponding key phase of the cardiac cycle. Based on the valve morphology data and the quantitative measurements of key anatomical structures, dynamic features such as the maximum leaflet separation point (i.e., the point where the leaflets are separated by the greatest distance during the cardiac cycle) are extracted. Using the leaflet anatomical length from the quantitative measurements of key anatomical structures as a reference scale, the ratio of the spatial distance of the maximum leaflet separation point to the leaflet anatomical length is calculated to determine the degree of dynamic offset of the maximum leaflet separation point relative to the leaflet itself (the larger the ratio, the more significant the offset), serving as the valve dynamic feature data. Based on the valve dynamic feature data and the quantitative measurements of key anatomical structures, Hausdorff... The distance is used to calculate the leaflet tear width of the valve, and the relative proportion of the tear is calculated by using the circumference of the valve annulus in the quantitative measurement results of each key anatomical structure as a benchmark. The degree of valve regurgitation is analyzed by time-series variation curves, and the spatial range of regurgitation is determined by combining the anatomical length of the leaflet. The relative proportion of the tear and the spatial range of regurgitation are used as quantitative indicators of valve function, such as the ratio of "tear width / valve annulus circumference" and the ratio of "regurgitation range / leaflet length".

[0099] Specifically, from an anatomical perspective, heart valves (such as the mitral valve, tricuspid valve, aortic valve, and pulmonary valve) are composed of 1 to 3 thin and elastic leaflets, the edges of which are connected to the myocardium via chordae tendineae.

[0100] Here, the final output of the completed dynamic functional parameter evaluation module is: cardiac cycle valvular dynamic functional indicators associated with static anatomical parameters (quantitative measurement results of each key anatomical structure), which includes both absolute values ​​and the proportion of relative anatomical structures, making it more suitable for model training and device selection needs.

[0101] The trained dynamic functional parameter evaluation module is used to quantify dynamic functional parameters such as valvular motion during the cardiac cycle. For example... Figure 2As shown, based on the reconstructed three-dimensional heart model and time-series images, key phases in the cardiac cycle (such as mid-systole during mitral regurgitation) are automatically captured, and dynamic features such as the maximum leaflet separation point are extracted. By using methods such as Hausdorff distance and time-series change curves, functional information such as leaflet tear width and degree of insufficiency is quantified into spatial parameters. Thus, valve function, which could only be evaluated qualitatively or semi-quantitatively, is transformed into precise numerical indicators that can be directly used for model training and device selection, providing a dynamic functional supplement to the accurate measurement of cardiac chambers.

[0102] Preferably, the trained calcification artifact compensation module is used to automatically detect calcification areas with CT values ​​>1000HU in the preoperative CT images of the patient based on the initial measurement data of the calcified area output by the spatial distance calculation module (including parameters such as the diameter, length, and perimeter of the calcified area). This results in the localization of the calcification artifact area (including the area range and CT value distribution). Based on the localization results of the calcification artifact area, thickness deviation compensation is performed on the initial measurement data of the calcified area to obtain the calibrated measurement data of the calcified area.

[0103] In this context, calcified tissue corresponding to areas with CT values ​​> 1000 HU is prone to measurement artifacts. Thickness deviation compensation can be performed on the initial measured parameters such as diameter, length, and perimeter for calcified areas to correct the interference of calcification artifacts on structural dimension measurements. This can provide accurate measurement data after calibration for the calcified areas and eliminate the deviation of calcification artifacts.

[0104] Here, the trained calcification artifact compensation module plays a role in identifying calcification artifacts and correcting measurement errors. For example... Figure 2 As shown, the distance calculation engine automatically identifies areas with CT values ​​>1000HU and compensates for thickness deviations, eliminating measurement deviations in diameter, length, and circumference caused by calcification, and ensuring measurement robustness under pathological conditions.

[0105] Step 103, training unit, used to train the initial heart chamber measurement model based on the heart chamber precision measurement dataset, to obtain the trained heart chamber precision model.

[0106] Preferably, step 103, the training unit, is used to train the initial accurate measurement model of cardiac chambers in stages, including:

[0107] Step 1031: Using the preoperative original multimodal cardiac images in the precise cardiac chamber measurement dataset from Step 101, the image segmentation and 3D reconstruction module is first trained in a supervised manner. The fine structure annotations in the precise cardiac chamber measurement dataset are used as the ground truth. Dice and cross-entropy loss are used to optimize the segmentation effect, and the trained image segmentation and 3D reconstruction module is obtained.

[0108] Step 1032: Then freeze the parameters of the trained image segmentation and 3D reconstruction module. Based on the 3D heart chamber model output by the image segmentation and 3D reconstruction module, train the spatial distance calculation module to minimize the L1 and / or L2 errors of the quantitative measurement results of key anatomical structures predicted by the spatial distance calculation module, such as anchorage length, valve annulus circumference, and coronary artery ostium height, with the heart chamber precision measurement dataset in step 101.

[0109] Here, the L1 / L2 error is a metric that measures the difference between the model's predicted values ​​and the true values ​​of the accurate cardiac chamber measurement dataset from step 101. Minimizing this difference during training means making the model's predictions as close as possible to the expert's accurate measurements. The L1 error is the absolute error, and the L2 error is the squared error.

[0110] Step 1033: Based on the quantitative measurement results of each key anatomical structure output by the trained spatial distance calculation module, combined with the intraoperative or postoperative cardiac cycle time series images in the automatic identification of the precise measurement dataset of cardiac chambers and the clinical regurgitation grading in the precise measurement dataset of cardiac chambers, train the dynamic functional parameter evaluation module so that the valvular function quantitative indicators output by the dynamic functional parameter evaluation module are highly consistent with the clinical evaluation.

[0111] Step 1034: Based on the initial measurement data of the calcified region output by the spatial distance calculation module (including anatomical parameters interfered by calcification artifacts, such as the circumference of the valve annulus with calcification and the anchoring length of the left atrial appendage) and the precise measurement data of the calcified region in the fine structural annotation of the precise measurement dataset of the heart chambers, calculate the deviation of the spatial distance calculation module to obtain the structural size deviation value caused by calcification artifacts; based on the valve function quantification indicators (such as tear width and regurgitation degree generated based on the initial measurement data) of the dynamic functional parameter evaluation module and the valve clinical truth indicators in the fine structural annotation of the precise measurement dataset of the heart chambers, calculate the deviation of the dynamic functional parameter evaluation module to obtain the functional evaluation deviation value transmitted by the structural size deviation; integrate the structural size deviation value and the functional evaluation deviation value to locate the correspondence between the deviation source and the calcified region (such as the measurement deviation of the circumference corresponding to valve annulus calcification and the regurgitation evaluation deviation) to obtain the calcification-related measurement deviation dataset (including deviation value, the location of the calcified region corresponding to the deviation, and the deviation transmission path); identify the preoperative CT in the original multimodal cardiac images in the precise measurement dataset of the heart chambers. The calcified areas in the image are located. The calcified area localization results are obtained. The calcification-related measurement deviation dataset and the calcified area localization results are correlated to establish a mapping model between calcified area features (size, location, CT value) and measurement deviations. With the goal of reducing systematic errors, the parameters of the calcification artifact compensation module (such as correction coefficients and thickness compensation formulas) are adjusted. The adjusted parameters of the calcification artifact compensation module are used to compensate for the initial measurement data of the calcified area. The deviation between the compensated data and the accurate measurement data of the calcified area is verified. Finally, the optimized calcification artifact compensation module is output (the compensation algorithm has been adapted to the deviation correction of different calcification scenarios).

[0112] Alternatively, a separate calcification case validation dataset (containing preoperative CT images, initial measurement data, and expert-precated ground truth) can be used. The optimized module can then be used to compensate for the initial measurement data in the validation dataset, calculate the error between the compensated data and the expert-precated ground truth, confirm that the error is below a preset threshold, and finally output the optimized calcification artifact compensation module (which can be directly embedded into the overall model workflow). The calcification case validation dataset is independent of the cardiac chamber precision measurement dataset.

[0113] Here, the initial accurate measurement model of the heart chambers is trained in stages. The Adam optimizer and preset learning rate can be used uniformly. Training stops when the validation set index reaches the set threshold, thus obtaining an accurate measurement model of the heart chambers based on multimodal images.

[0114] Step 104: Input the multimodal images of the patient to be tested into the trained accurate measurement model of the heart chambers to obtain the measurement data of the heart chambers of the patient to be tested.

[0115] like Figure 3As shown, this embodiment of the invention also provides a method for precise measurement of cardiac chambers based on multimodal imaging, including:

[0116] Step 301: Acquire multimodal images of the patient to be tested;

[0117] Step 302: Input the multimodal images of the patient to be tested into the trained accurate measurement model of the heart chambers to obtain the heart chamber measurement data of the patient to be tested. The accurate measurement model of the heart chambers is trained by the above-mentioned training method of the accurate measurement model of the heart chambers based on multimodal images.

[0118] Similarly, in this embodiment, based on the application scenario described above, the original multimodal cardiac images are input into a precise measurement model of the cardiac chambers. The image segmentation and 3D reconstruction module generates a 3D cardiac model with detailed anatomical labels. Subsequently, the spatial distance calculation module uses methods such as central axis, geodesic, and surface normal distance on this model to accurately calculate key anatomical structural parameters. The dynamic functional parameter evaluation module combines time-series images to extract features such as maximum leaflet separation at key phases of the cardiac cycle and quantifies valve function into spatial parameters using indicators such as Hausdorff distance. The calcification artifact compensation module automatically identifies high CT value calcification areas and compensates for thickness deviations to obtain precise measurement data of the cardiac chambers under pathological conditions.

[0119] like Figure 4 As shown, this embodiment of the invention also provides a training device for a precise measurement model of cardiac chambers based on multimodal imaging, comprising:

[0120] First establishment unit 401: used to establish a precise measurement dataset of cardiac chambers, the dataset including but not limited to original multimodal cardiac images and fine structural annotations matched with the original multimodal cardiac images;

[0121] The second establishment unit 402 is used to establish an initial accurate measurement model of the heart chambers. The initial accurate measurement model of the heart chambers includes an image segmentation and three-dimensional reconstruction module, a spatial distance calculation module, a dynamic functional parameter evaluation module, and a calcification artifact compensation module.

[0122] Training unit 403: Used to train the initial accurate measurement model of the heart chambers based on the multimodal data in the accurate measurement dataset of the heart chambers, so as to obtain the accurate measurement model of the heart chambers.

[0123] It should be noted that the cardiac chamber precision measurement model training device provided in the embodiments of the present invention is a device capable of executing the above-described cardiac chamber precision measurement model training method. Therefore, all embodiments of the above-described cardiac chamber precision measurement model training method are applicable to the cardiac chamber precision measurement model training device and can achieve the same or similar beneficial effects.

[0124] like Figure 5 As shown, this embodiment of the invention also provides a precise measurement device for cardiac chambers based on multimodal imaging, comprising:

[0125] Acquisition unit 501: used to acquire multimodal images;

[0126] Prediction unit 502: used to input the multimodal image into the accurate measurement model of the heart chamber to obtain the measurement data of the heart chamber. The accurate measurement model of the heart chamber is trained by the above-mentioned training method of the accurate measurement model of the heart chamber based on multimodal image.

[0127] It should be noted that the precise cardiac chamber measurement device provided in the embodiments of the present invention is a system capable of performing the above-described precise cardiac chamber measurement method. Therefore, all embodiments of the above-described precise cardiac chamber measurement method are applicable to the precise cardiac chamber measurement device and can achieve the same or similar beneficial effects.

[0128] This invention also provides an intelligent terminal that integrates the above-mentioned precise measurement model of cardiac chambers based on multimodal imaging into a domestically produced DSA device, enabling real-time intraoperative acquisition, processing and analysis of multimodal images, and assisting in the size selection and path planning of interventional devices.

[0129] In this embodiment, the intelligent terminal is integrated with the domestically produced DSA equipment. Through a preset data interaction interface and communication protocol, it receives multimodal images such as CTA / MRA and DSA / CBCT. During the procedure, it performs real-time image acquisition and transmission, and calls upon a precise cardiac chamber measurement model based on multimodal images for processing and analysis, ensuring that the automatic latency calculation meets the real-time clinical requirements. Measurement results are displayed on the DSA display screen or an external monitor in the form of overlaid annotations and numerical lists. The operator can select the target structure on the touchscreen interface and trigger measurement and result updates with a single click. The intelligent terminal can also generate recommended interventional device sizes and a summary of key anatomical parameters based on the measurement parameters for preoperative / intraoperative assessment and physician decision-making reference.

[0130] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described system for constructing a precise measurement model of cardiac chambers, or the above-described method for precise measurement of cardiac chambers.

[0131] The technical advantage of this invention lies in that it provides a method for precise measurement of cardiac chambers based on multimodal imaging and a model training method, overcoming the limitations of low accuracy in traditional static linear measurements. Through deep learning-driven image segmentation and 3D reconstruction technology, the model can achieve sub-millimeter-level measurement accuracy, constructing a precise 3D model of the cardiac chambers. The invention also includes a model training device, a measuring device, a smart terminal, and a storage medium, which can assist in the size selection and path planning of interventional devices, thereby effectively improving the efficiency and safety of structural heart disease surgery.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0133] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0134] Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A system for constructing a cardiac chamber measurement model, characterized in that, include: The acquisition unit is used to establish a precise measurement dataset of cardiac chambers. This dataset includes: original multimodal cardiac images, fine structural annotations matched with the original multimodal cardiac images, and clinical regurgitation grading. The fine structural annotations include: basic cardiac structural annotations and fine anatomical and pathological structural annotations. The basic cardiac structural annotations include: Left ventricular endocardium or epicardium, left atrium, right ventricular blood chamber, right atrial blood chamber, and left ventricular myocardium; Ascending aorta and pulmonary artery; Among them, detailed anatomical and pathological structural annotations include: Valve annulus, leaflets, mitral valve occlusion region, and dynamic valve movement trajectory; Left atrial appendage and opening details and coronary sinus; The modeling unit is used to establish an initial cardiac chamber measurement model, which includes: an image segmentation and three-dimensional reconstruction module, a spatial distance calculation module, a dynamic functional parameter evaluation module, and a calcification artifact compensation module. The training unit is used to train the initial heart chamber measurement model based on the heart chamber precision measurement dataset to obtain a trained heart chamber precision model. The measurement unit is used to input the multimodal images of the patient to be tested into the trained accurate measurement model of the heart chambers to obtain the measurement data of the heart chambers of the patient to be tested. The original multimodal cardiac images include: preoperative images, intraoperative images, and postoperative images; the formats of the original multimodal cardiac images include: CTA, CT, MRA, TEE, DSA, Xray, and CBCT. The trained spatial distance calculation module is used to locate the spatial coordinates and contour data of various key anatomical structures based on a three-dimensional cardiac chamber model with fine anatomical labels; based on the spatial coordinates and contour data of various key anatomical structures, quantitative measurement results of each key anatomical structure are obtained; when calcification is detected in a key anatomical structure, the anatomical structure containing calcification is measured first to obtain the initial measurement data of the calcified area. The trained spatial distance calculation module is used for: Based on the coordinate and contour data of the nonlinear structure in the three-dimensional cardiac chamber model, the central axis of the nonlinear structure is automatically identified; the nonlinear structure is segmented along the central axis; and the effective anchoring length of each segment is calculated using a geodesic algorithm to obtain the segmented effective anchoring length data of the nonlinear structure. For measuring the valve annulus circumference of a saddle-shaped curved surface structure in a 3D heart model: Based on the coordinate and contour data of a non-saddle-shaped curved surface structure in the 3D heart model, the edge contour of the valve annulus surface is extracted; based on the edge contour of the valve annulus surface, a geodesic algorithm is used to fit the closed path of the surface edge; the length of the closed path is calculated to obtain the valve annulus circumference, and the valve annulus circumference measurement value is obtained. To measure the height of coronary artery ostia in 3D heart models: Based on the relative spatial position data of the vascular ostia and their corresponding chambers in the 3D heart model, the center point of the vascular ostia is located; using a normal distance algorithm, the vertical height from the center point of the vascular ostia to the wall of the corresponding chamber is measured as the height of the vascular ostia. The trained dynamic functional parameter evaluation module is used to automatically identify key phases of the cardiac cycle in intraoperative or postoperative cardiac cycle time-series images of the patient, based on the 3D heart model generated by the image segmentation and 3D reconstruction module and the quantitative measurement results of key anatomical structures output by the spatial distance calculation module. It spatially matches the valve morphology under the key phases of the cardiac cycle with the quantitative measurement results of each key anatomical structure to locate the relative position of the valve within the anatomical structure, serving as the valve morphology data for the corresponding key phase of the cardiac cycle. Based on the valve morphology data under the key phases of the cardiac cycle and the quantitative measurement results of each key anatomical structure, the maximum leaflet separation point is extracted. Using the leaflet anatomical length in the quantitative measurement results of each key anatomical structure as a reference scale, the ratio of the spatial distance of the maximum leaflet separation point to the leaflet anatomical length is calculated to determine the dynamic offset of the maximum leaflet separation point relative to the leaflet itself, serving as the valve dynamic feature data. Based on the valve dynamic feature data and the quantitative measurement results of each key anatomical structure, Hausdorff... The width of the valve leaflet tear was calculated by distance, and the relative proportion of the tear was converted from the circumference of the valve annulus in the quantitative measurement results of each key anatomical structure. The degree of valve regurgitation was analyzed by time-series variation curves, and the spatial range of regurgitation was determined by combining the anatomical length of the leaflet. The relative proportion of the tear and the spatial range of regurgitation were used as quantitative indicators of valve function. The trained calcification artifact compensation module is used to automatically detect calcification areas with CT values ​​greater than a preset CT threshold in the preoperative CT images of the patient based on the initial measurement data of the calcification area output by the spatial distance calculation module, and obtain the localization result of the calcification artifact area. Based on the localization result of the calcification artifact area, the initial measurement data of the calcification area is compensated for thickness deviation to obtain the calibrated measurement data of the calcification area.

2. The system for constructing a cardiac chamber measurement model as described in claim 1, characterized in that, The trained image segmentation and 3D reconstruction module is used to call the trained 3D U-Net variant model to perform pixel-level automatic segmentation on the preoperative original multimodal cardiac images of the patient to be examined, identify and delineate the contour boundaries of complex anatomical structures, and obtain images with segmentation masks for each anatomical structure. Based on the spatial coordinate information of the images with anatomical structure segmentation masks, the 2D image layers of the images with anatomical structure segmentation masks are fused into a 3D model through a 3D reconstruction algorithm, and the accuracy is calibrated to the sub-millimeter level to obtain a sub-millimeter accuracy 3D initial cardiac model. An anatomical perception loss function is introduced to iteratively optimize the thin-walled structural contours of the 3D initial cardiac model, correct the segmentation error, and obtain a 3D cardiac chamber model with fine anatomical labels.

3. The system for constructing a cardiac chamber measurement model as described in claim 1, characterized in that, The training unit is used for: The image segmentation and 3D reconstruction module was trained in a supervised manner using preoperative raw multimodal cardiac images from the precise cardiac chamber measurement dataset. The fine structure annotations in the precise cardiac chamber measurement dataset were used as ground truth, and Dice and cross-entropy loss were used to optimize the segmentation effect, resulting in the trained image segmentation and 3D reconstruction module. Freeze the parameters of the image segmentation and 3D reconstruction module after training, and train the spatial distance calculation module based on the 3D heart chamber model output by the image segmentation and 3D reconstruction module to minimize the L1 and / or L2 errors between the quantitative measurement results of each key anatomical structure predicted by the spatial distance calculation module and the accurate measurement dataset of the heart chamber. Based on the quantitative measurement results of key anatomical structures output by the spatial distance calculation module, combined with the intraoperative or postoperative cardiac cycle time series images in the precise measurement dataset of cardiac chambers and the clinical regurgitation grading in the precise measurement dataset of cardiac chambers, the dynamic functional parameter assessment module is trained so that the valvular function quantitative indicators output by the dynamic functional parameter assessment module are consistent with the clinical evaluation. Based on the initial measurement data of the calcified region output by the spatial distance calculation module and the precise measurement data of the calcified region in the fine structural annotation of the precise measurement dataset of the heart chambers, the deviation of the spatial distance calculation module is calculated to obtain the structural size deviation value caused by calcification artifacts. Based on the valvular function quantification index from the dynamic functional parameter assessment module and the clinical true value index of valves in the fine structural annotation of the precise cardiac chamber measurement dataset, the deviation of the dynamic functional parameter assessment module is calculated to obtain the functional assessment deviation value transmitted by the structural size deviation. The structural size deviation value and the functional assessment deviation value are integrated to locate the correspondence between the source of the deviation and the calcified region, resulting in a calcification-related measurement deviation dataset. Calcified regions in preoperative CT images from the original multimodal cardiac images in the precise cardiac chamber measurement dataset are identified to obtain calcification region localization results. The calcification-related measurement deviation dataset and the calcification region localization results are correlated to establish a mapping model between calcification region features and measurement deviation. With the goal of reducing systematic error, the parameters of the calcification artifact compensation module are adjusted. The adjusted parameters of the calcification artifact compensation module are used to compensate for the initial measurement data of the calcified region, verifying whether the deviation between the compensated data and the precise measurement data of the calcified region is reduced. Finally, the optimized calcification artifact compensation module is output.

4. The system for constructing a cardiac chamber measurement model as described in claim 1, characterized in that, The acquisition unit is used to generate initial annotations from the original multimodal cardiac images according to the preset graded anatomical annotation specifications and using a semi-automatic pre-annotation tool. Then, the initial annotations are refined using the doctor's drawing tool to obtain an image annotation draft with preliminary anatomical structure labels; Based on the image annotation draft with preliminary anatomical labels, the Hounsfield threshold slider is used to distinguish soft plaques and calcified tissues in the original multimodal cardiac images and supplement the corresponding pathological tissue annotations. The pathological tissue annotations are integrated into the image annotation draft with preliminary anatomical structure labels to obtain an image annotation draft containing anatomical structure and pathological tissue annotations. Multiple cardiac imaging experts provided image annotation drafts, each based on anatomical structures and pathological tissues. These drafts were used to fully annotate the original multimodal cardiac images, resulting in three independent annotation results. The STAPLE algorithm was used to weightedly fuse overlapping and dissimilar regions from each independent annotation result, generating an initial fused annotation result. The Dice similarity coefficient between the annotations from different experts was calculated to assess the degree of difference, resulting in a Dice coefficient evaluation report for each region. If the Dice coefficient of the dissimilar region in the evaluation report is greater than or equal to a preset threshold, the initial fused annotation result is directly used as the unified annotation draft with consensus. If the Dice coefficient of the dissimilar region in the evaluation report is less than the preset threshold, expert arbitration and negotiation are initiated. Based on the anatomical features and pathological manifestations of the original multimodal cardiac images, the final annotation boundaries and labels in the initial fused annotation result are corrected to obtain a unified annotation draft with consensus. Use preset verification indicators to verify the quality of the unified annotation draft that has reached a consensus; Based on annotation quality, the unified annotation drafts that have passed verification and reached a consensus will be used as fine-structure annotations.

5. A computer-readable storage medium having stored thereon computer-executable instructions, wherein, When the computer-executable instructions are executed by the processor, the processor causes the processor to: execute the system for constructing a cardiac chamber measurement model as described in any one of claims 1 to 4.

Images