Cardiac magnetic resonance report generation method and system based on multi-dimensional tissue characteristic knowledge graph

Abstract

The invention discloses a cardiac magnetic resonance report generation method and system based on a multi-dimensional tissue feature knowledge graph. The method comprises the following steps: performing motion correction and multi-view feature extraction on a cardiac magnetic resonance multi-sequence image to obtain an original numerical attribute of an instance node; constructing a dynamic knowledge graph containing tissue characteristics and medical priori, and generating a target reasoning chain with a clinical basis in combination with neural symbol reasoning; then, in a controlled generation stage, an inference chain is injected into a prefix tree, in a decoding step, mask physical shielding logic conflict description is utilized, and a pointer network is synchronously started to directly copy original numerical value attributes into a text sequence; and finally, executing consistency logic verification through a text restoration technology, and identifying and intercepting factual errors. Through the atlas constraint and numerical unvarnished transmission mechanism, the factual illusion problem of a generative model is effectively solved, and the accuracy and reliability of a heart diagnosis report are remarkably improved.

Description

A Method and System for Generating Cardiac Magnetic Resonance Imaging Reports Based on Multidimensional Tissue Feature Knowledge Graph Technical Field

[0001] This invention belongs to the fields of medical image information processing and medical artificial intelligence, and in particular relates to a method and system for generating cardiac magnetic resonance reports based on a multidimensional tissue feature knowledge graph. Background Technology

[0002] Cardiovascular magnetic resonance (CMR), as the "gold standard" non-invasive imaging technique for assessing the morphology, function, and tissue characteristics of cardiovascular diseases, plays a crucial role in modern clinical diagnosis. Unlike single-modality imaging such as computed tomography (CT) or echocardiography, CMR examinations typically involve highly complex sequences, including cine MRI for assessing wall motion, T2-weighted imaging or T2 mapping for detecting myocardial edema, T1 mapping and extracellular volume (ECV) fractions for assessing diffuse fibrosis or infiltrative lesions, and late gadolinium enhancement (LGE) sequences for identifying focal scarring or necrosis. While this multi-parameter imaging modality provides rich pathophysiological information, it also significantly increases the difficulty of diagnosis. Clinicians must perform high-dimensional logical integration and causal reasoning on heterogeneous data when writing reports. For example, a definitive diagnosis of acute myocarditis requires strict adherence to clinical guidelines, such as the Lake Louise criteria, while simultaneously meeting the logical criteria for T1 and T2 baseline indicators.

[0003] In recent years, automated medical report generation technology based on encoder-decoder architecture has become a research hotspot. However, existing end-to-end pure data-driven models have significant limitations in practical clinical applications. First, these models commonly suffer from the problem of factual illusion. Due to a lack of deep semantic understanding of medical image content, models often generate false reports based on the textual probability distribution of the training data, easily leading to errors in lesion description or confusion of anatomical locations. Second, general deep learning models lack domain-specific logical reasoning capabilities and cannot effectively handle conflicts between multimodal sequences. For example, when LGE shows transmural enhancement but a cinnabar sequence shows normal motion, the model struggles to perform evidence-based logical judgments like a professional physician, resulting in contradictory reports.

[0004] More importantly, existing technologies struggle to guarantee rigorous consistency between quantitative indicators and qualitative descriptions. Because text generation models typically compress image features into high-dimensional latent vectors, precise numerical information is lost during decoding. This results in generated text with low numerical accuracy and frequent logical fallacies such as describing normal T1 values ​​as elevated. Furthermore, medical diagnostic guidelines are frequently updated, and black-box end-to-end models struggle to adapt to changes in logical rules through low-cost fine-tuning, leading to unsatisfactory system flexibility and maintenance costs.

[0005] In summary, how to construct a cardiac magnetic resonance imaging intelligent diagnostic solution that can both perform complex medical logical reasoning and ensure strict consistency between the generated text description and the quantitative image data is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention proposes a cardiac magnetic resonance imaging report generation method based on a multidimensional tissue feature knowledge graph, which addresses the technical problems commonly found in existing end-to-end report generation technologies, such as lack of logical reasoning, the generation of text that is prone to factual illusions, and inconsistencies between key quantitative indicators and qualitative descriptions.

[0007] According to one embodiment of the present invention, a method for generating cardiac magnetic resonance imaging reports based on a multidimensional tissue feature knowledge graph is proposed, comprising: acquiring multimodal cardiac magnetic resonance images of a patient; performing motion correction on the multimodal cardiac magnetic resonance images; identifying anatomical segments in the corrected images; extracting multidimensional tissue feature tensors corresponding to each anatomical segment, wherein the multidimensional tissue feature tensors encapsulate original numerical attributes characterizing anatomical morphology and tissue properties; generating patient-specific image feature instance nodes based on the multidimensional tissue feature tensors; establishing a mapping association between the instance nodes and the original numerical attributes in the multidimensional tissue feature tensors; and mounting the instance nodes to a pre-constructed static ontology graph to complete the process. The construction of a dynamic knowledge graph involves a static ontology graph containing medical prior knowledge and predefined diagnostic conclusion nodes. A graph neural network based on path representation learning is used to transmit information across the dynamic knowledge graph, searching for logical evidence paths from instance nodes to diagnostic conclusion nodes, and determining the target inference chain based on path weights. The target inference chain is then used as context input to a pre-trained large language model, and a prefix tree is dynamically constructed based on the target inference chain. During the decoding phase of the pre-trained large language model, the prefix tree is used to mask and filter the output probability distribution. Simultaneously, a numerical pointer copying operation is used to fill the original numerical attributes in the inference chain into the generated text sequence to obtain a diagnostic report.

[0008] According to one embodiment of the present invention, a cardiac magnetic resonance imaging (MRI) report generation system based on a multidimensional tissue feature knowledge graph is also proposed, comprising: a feature extraction module, used to acquire multimodal cardiac MRI images of a patient, perform motion correction on the multimodal cardiac MRI images, identify anatomical segments in the corrected images, extract multidimensional tissue feature tensors corresponding to each anatomical segment, wherein the multidimensional tissue feature tensors encapsulate original numerical attributes characterizing anatomical morphology and tissue properties; and a dynamic knowledge graph construction module, used to generate patient-specific image feature instance nodes based on the multidimensional tissue feature tensors, establish a mapping association between the instance nodes and the original numerical attributes in the multidimensional tissue feature tensors, and mount the instance nodes to a pre-constructed static ontology graph to complete the dynamic knowledge graph construction. The system comprises: a static ontology graph containing medical prior knowledge and predefined diagnostic conclusion nodes; a logical evidence path reasoning module, which uses a graph neural network based on path representation learning to transmit information on the dynamic knowledge graph, search for logical evidence paths from the instance nodes to the diagnostic conclusion nodes, and determine the target reasoning chain based on the path weights; and a graph constraint report generation module, which takes the target reasoning chain as context input to a pre-trained large language model, dynamically constructs a prefix tree based on the target reasoning chain, uses the prefix tree to mask and filter the output probability distribution during the decoding stage of the pre-trained large language model, and fills the original numerical attributes in the reasoning chain into the generated text sequence through a numerical pointer copying operation to obtain a diagnostic report.

[0009] The proposed cardiac magnetic resonance imaging (MRI) report generation scheme based on a multidimensional tissue feature knowledge graph firstly removes respiratory artifacts and decouples complex image data by performing motion correction on multimodal images and extracting multidimensional tissue feature tensors based on anatomical segments, providing a high-fidelity physical basis for subsequent analysis. Secondly, it constructs a dynamic knowledge graph by establishing a mapping relationship between instance nodes and original numerical attributes, and attaching instance nodes carrying the attributes to a static ontology graph. This overcomes the disconnect between discrete symbolic reasoning and continuous numerical computation, enabling the diagnostic logic to include both graph-based qualitative judgments and original ontology-based qualitative judgments. The initial numerical values ​​provide quantitative basis; furthermore, by using a path-based graph neural network to explicitly search for logical evidence chains, the black-box model becomes interpretable, clearly demonstrating the derivation path from image features to diagnostic conclusions; finally, in the text generation stage, by combining a prefix tree dynamically constructed based on the target inference chain with a numerical pointer copying mechanism, the large language model is forced to strictly follow the inferred logical path during decoding and directly cites the precise numerical values ​​extracted from the image, thereby eliminating the illusion of fact in medical reports and ensuring that the generated diagnostic reports are logically rigorous and numerically accurate, significantly improving the reliability and clinical practical value of the auxiliary diagnostic system. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this specification and, together with the description, serve to explain the principles of this specification.

[0011] Figure 1 shows a flowchart of a cardiac magnetic resonance imaging (MRI) report generation method based on a multidimensional tissue feature knowledge graph according to an embodiment of the present invention; Figure 2 shows an architecture block diagram of a cardiac MRI report generation system based on a multidimensional tissue feature knowledge graph according to an embodiment of the present invention; Figure 3 shows a schematic diagram of the motion correction algorithm based on robust principal component analysis (RPCA) according to an embodiment of the present invention; Figure 4 shows a schematic diagram of the architecture of a multi-view feature extraction network according to an embodiment of the present invention; Figure 5 shows a schematic diagram of the dynamic knowledge graph data structure and neural symbol reasoning logic according to an embodiment of the present invention; Figure 6 shows a schematic diagram of the decoding control based on a prefix tree and numerical pointer copying mechanism according to an embodiment of the present invention; Figure 7 shows a schematic diagram of the consistency logic verification process based on text restoration according to an embodiment of the present invention; Figure 8 is a structural schematic diagram of an electronic device shown in at least one embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0013] Embodiments of this invention can be applied to computer systems / servers that can operate with a wide range of other general-purpose or special-purpose computing system environments or configurations. Examples of well-known computing systems, environments, and / or configurations suitable for use with computer systems / servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments that include any of the above systems, etc.

[0014] Computer systems / servers can be described in the general context of computer system executable instructions (such as program modules) executed by the computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are performed by remote processing devices linked through a communication network. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0015] This invention proposes a cardiac magnetic resonance imaging (CMR) report generation scheme based on a multidimensional tissue feature knowledge graph. It incorporates neural symbolic artificial intelligence processing methods, constructing a dynamic patient-specific knowledge graph as a connection between the underlying raw numerical attributes and the high-level clinical diagnostic logic. This invention extracts multidimensional tissue features from multiple CMR sequences using a deep neural network and synchronously associates them with the raw numerical attributes to instance nodes in the knowledge graph. Subsequently, a graph neural network based on path representation learning is used to perform logical reasoning that conforms to medical guidelines, and this reasoning chain guides the controlled decoding process of a large language model. This invention not only effectively overcomes the factual illusion problem commonly found in end-to-end models of medical image report generation but also ensures that the generated diagnostic conclusions are supported by a rigorous logical evidence chain and are absolutely consistent with quantitative image indicators, achieving highly interpretable intelligent assisted diagnosis.

[0016] Please refer to Figure 1, which shows a flowchart of a method for generating cardiac magnetic resonance reports based on a multidimensional tissue feature knowledge graph according to an embodiment of the present invention. The method mainly includes the following steps S101 to S104.

[0017] Step S101: Acquire the patient's multimodal cardiac magnetic resonance images, perform motion correction on the multimodal cardiac magnetic resonance images, identify anatomical segments in the corrected images, and extract multidimensional tissue feature tensors corresponding to each anatomical segment. The multidimensional tissue feature tensors encapsulate original numerical attributes characterizing anatomical morphology and tissue properties.

[0018] Multimodal cardiac magnetic resonance (CMR) imaging typically refers to a collection of images containing multiple different functional sequences, including cine MRI for assessing wall motion, T2-weighted imaging or T2 mapping for detecting myocardial edema, T1 mapping and extracellular volume (ECV) fractional sequences for assessing diffuse fibrosis, and / or late gadolinium enhancement (LGE) sequences for identifying myocardial scarring. Anatomical segments refer to semantically defined anatomical regions into which cardiac structures are divided according to clinical criteria such as the AHA 17-segment model. Multidimensional tissue feature tensors can be viewed as a structured data carrier that integrates diagnostic indicators from multiple dimensions, including morphology, functional biomechanics, qualitative tissue analysis, and quantitative tissue analysis, along with corresponding anatomical dimensions. Raw numerical attributes characterizing anatomical morphology and tissue properties typically refer to continuous physical values ​​calculated directly from the image pixel level without discretization, such as millimeters of myocardial wall thickness or milliseconds of T1 mapping.

[0019] Raw DICOM data containing the aforementioned sequences can be obtained by interfacing with a Picture Archiving and Transmission System (PACS) compatible with the Medical Digital Imaging and Communications (DICOM) standard, serving as the patient's multimodal CMR images.

[0020] Since heartbeat and respiratory movements can cause spatial misalignment between different frames in pixel-level mapping sequences such as T1 / T2 mapping, in some embodiments of the present invention, motion correction of multimodal CMR images can be performed using matrix factorization algorithms based on low-rank and sparsity constraints. Specifically, the multimodal cardiac magnetic resonance images can be constructed as an image sequence matrix. The image sequence matrix Decomposed into low-rank matrices representing anatomical structures and sparse matrix representing dynamic components By establishing and solving for minimizing the nuclear norm and A convex optimization model based on norm is used to extract background images with respiratory motion artifacts removed. As the corrected image. The specific mathematical model can be expressed as: ,in, It is the sum of the singular values ​​of the matrix (nuclear norm), used to enforce the consistency of the anatomical structure; for Norms are used to induce sparsity in order to capture motion artifact terms; This is the regularization parameter, usually set to... To meet the requirements of clinical computational efficiency, the inexact augmented Lagrange multiplier method (IALM) can be used for iterative solution when solving this convex optimization model.

[0021] Image sequence matrix In this model, each column represents an image vector at a specific moment. According to the principles of magnetic resonance imaging, the anatomical background (e.g., stationary myocardial structures) exhibits high correlation over time, mathematically representing low-rank characteristics; while drastic local changes caused by respiratory motion, contrast agent flow, or random noise exhibit sparsity characteristics. This implementation, by decoupling low-rank and sparsity components, effectively removes spatially consistent anatomical background images, excluding respiratory motion artifacts, from the original images affected by respiratory interference. This avoids spurious increases in tissue mapping values ​​(e.g., T1 values) due to registration errors, thus improving the reliability of subsequent quantitative analysis.

[0022] It should be noted that the motion correction method of this invention is not limited to the matrix factorization method based on low rank and sparse constraints described above. In practical applications, those skilled in the art can also select other image registration techniques to achieve this objective according to specific needs. For example, in some embodiments, an elastic registration technique based on B-spline free-form deformation (FFD) can be used to perform motion correction on the multimodal cardiac magnetic resonance images. In other embodiments, an unsupervised registration network based on deep learning, such as the VoxelMorph architecture, can be used to perform motion correction on the multimodal cardiac magnetic resonance images.

[0023] In some implementations, extracting multidimensional tissue feature tensors corresponding to each anatomical segment includes: using a deep neural network including a multi-view fusion attention module to jointly process long-axis and short-axis images to divide the heart into standard anatomical segments; and extracting morphological indicators, strain dynamics indicators, late gadolinium enhancement pattern classification labels, and mapping sequence quantitative values ​​in parallel for each standard anatomical segment to constitute the multidimensional tissue feature tensor.

[0024] The deep neural network in this embodiment can be a neural network with an architecture such as the Swin UNETR. Considering the resolution limitations of short-axis slices (SAX) at the apex and mitral valve plane, this embodiment introduces a multi-view fusion attention mechanism into the deep neural network. For example, the deep neural network inputs the acquired short-axis slice sequences and long-axis slice sequences (including 2-chamber, 3-chamber, and 4-chamber views) into a feature extraction backbone with shared weights to extract local anatomical features from each view. In the feature fusion stage, this mechanism captures three-dimensional spatial dependencies by establishing a cross-view attention mapping. Specifically, a query matrix can be constructed using the high-resolution features of the long-axis view along the vertical axis, while a key matrix and a value matrix can be constructed using the features of the short-axis view. By calculating the cross-correlation score between the query matrix and the key matrix, the weight ratio of the corresponding anatomical location under different views is dynamically identified, thereby using the longitudinal boundary information of the long-axis view to compensate for the pixel loss or partial volume effect-induced boundary blurring in the short-axis view at the apex or mitral valve annulus. Finally, the fused feature vectors can be fed into the upsampling decoder to accurately predict the contours of the left ventricle, right ventricle, and myocardium, and the segmentation results are automatically mapped onto the AHA 17-segment model based on anatomical morphological features, thereby providing accurate spatial localization labels for each anatomical segment.

[0025] Based on this, for each anatomical segment, key features of multiple dimensions can be extracted from the corresponding motion-corrected sequence to construct the multidimensional tissue feature tensor. Morphological indicators may include, for example, the maximum wall thickness obtained by calculating the Euclidean distance between the endocardium and epicardium at end-diastole, which is a key basis for diagnosing hypertrophic cardiomyopathy; strain dynamic indicators may include, for example, radial and circumferential strain calculated using feature tracking algorithms; late gadolinium enhancement pattern classification labels may include, for example, LGE distribution types identified using classification networks, such as subendocardial, transmural, subepidermal, or patchy; and mapping sequence quantitative values ​​may include, for example, the average ROI calculated on the corrected mapping sequence, including native T1 and T2 values, and ECV scores calculated in conjunction with hematocrit. These features are ultimately encapsulated into a structured tensor, in which original numerical attributes such as specific thickness in millimeters and T1 in milliseconds are fully preserved, rather than simply being converted into discrete labels of "thickened" or "elevated," laying the foundation for subsequent generation of accurate quantitative reports. Table 1 shows a specific example of the AHA 17 segment multidimensional feature definition according to an embodiment of the present invention: Table 1: AHA 17 segment multidimensional feature definition table

[0026] Based on this example, for the target patient, a multidimensional tissue feature tensor of shape (17, D) can be output as the data basis for subsequent atlas construction.

[0027] This implementation method achieves decoupling of image data from the pixel level to the semantic feature level through multi-dimensional parallel extraction, enabling subsequent logical reasoning to be based on a comprehensive, accurate feature set that conforms to clinical guidelines, thereby enhancing the reliability of the diagnostic plan.

[0028] Step S102: Generate patient-specific image feature instance nodes based on the multidimensional tissue feature tensor, establish a mapping association between the instance nodes and the original numerical attributes in the multidimensional tissue feature tensor, and mount the instance nodes to a pre-constructed static ontology graph to complete the construction of a dynamic knowledge graph. The static ontology graph contains medical prior knowledge and predefined diagnostic conclusion nodes.

[0029] This step is used to connect low-level imaging features (also known as raw numerical attributes) with high-level clinical logic.

[0030] The system has a pre-built static ontology graph containing medical prior knowledge and predefined diagnostic conclusion nodes. This atlas can be constructed based on SCMR guidelines and related pathophysiology literature. It can utilize ontology modeling languages ​​such as OWL to define structured medical knowledge in the field of cardiac magnetic resonance imaging. The types of medical entities (sign nodes) in the atlas can include: anatomical structures (e.g., basal anterior wall, left ventricular global features), imaging signs (e.g., late subendocardial gadolinium enhancement, elevated T1 / T2 values, loss of wall motion), pathological diagnoses (e.g., ischemic cardiomyopathy, acute myocarditis, Fabry disease), and diagnostic criteria (e.g., the 2018 Lake Louise criteria). The relationship types between medical entities define the connection rules of medical logic, such as suggestive relationships (e.g., transmural enhancement suggests ischemic cardiomyopathy), dependency relationships (e.g., a diagnosis of myocarditis requires meeting diagnostic criteria), compositional relationships (e.g., diagnostic criteria consist of multiple sub-criteria), and conflicting relationships (e.g., normal wall motion conflicts with signs of transmural infarction).

[0031] The system reads the multidimensional tissue feature tensor output in step S101, for example, the tensor with shape as... The tensor is used to create a corresponding image feature instance node for each anatomical segment (e.g., each examined region in the AHA 17 segment model), which is a digital representation of a specific anatomical location of the target patient.

[0032] In some implementations, the original numerical values ​​in the high-dimensional tensor can be physically bound to graph nodes through an attribute mapping association mechanism. This includes: defining an attribute matrix in the data structure of the dynamic knowledge graph, wherein the row indices of the attribute matrix have a one-to-one correspondence with the instance nodes; and using attribute assignment operations to store the original numerical attributes in the multidimensional organizational feature tensor into the attribute slots in the attribute matrix corresponding to the instance nodes, thereby overcoming the deficiency of traditional symbolic reasoning in handling continuous numerical values.

[0033] In one example, an attribute matrix can be defined within the data structure of a dynamic knowledge graph. ,in, Represents the total number of nodes. Representing the predefined attribute slot dimensions, the row indices of the attribute matrix correspond one-to-one with the instance nodes. Subsequently, an attribute assignment operation can be performed using mechanisms such as Scatter-Update, directly writing the original numerical attributes (e.g., extracted 15mm wall thickness or 1350ms T1 value) from the multidimensional tissue feature tensor into the row vectors (i.e., attribute slots) corresponding to each instance node in the attribute matrix.

[0034] The system can then perform dynamic mounting operations. It can analyze qualitative labels in feature tensors or establish dynamic connections between instance nodes and medical entities in the static ontology graph by thresholding quantitative values. For example, if the feature vector of a patient's 7th segment shows a T2 value exceeding the normal threshold and subepicardial enhancement, the system can create connections from the instance node of that segment to the "T2_High" and "LGE_Subepicardial" nodes in the static ontology graph.

[0035] The dynamic knowledge graph constructed through the above process It can reflect the patient's current pathological state. This atlas transforms unstructured image features into knowledge representations with semantic associations and logical relationships, effectively overcoming the disconnect between discrete symbolic reasoning and continuous numerical computation in end-to-end models, and also providing a structured knowledge foundation for subsequent path-based interpretable reasoning.

[0036] Step 103: Use a graph neural network based on path representation learning to transmit information on the dynamic knowledge graph, search for logical evidence paths from the instance node to the diagnostic conclusion node, and determine the target inference chain based on the path weights.

[0037] In this step, a causal connection between imaging features and clinical diagnosis can be established through explicit graph reasoning. The logical evidence path refers to a semantic path in the dynamic knowledge graph that starts from an instance node representing the imaging features of a specific part of the patient's body, passes through a series of intermediate nodes such as specific imaging signs and pathophysiological indicators, and finally connects to the diagnostic conclusion node. The target reasoning chain refers to the set of paths with the highest confidence or attention weight among all the logical evidence paths searched; it represents the explicit explanation provided by the system for the current diagnostic result.

[0038] To capture the complex nonlinear logical relationships between entities, some implementations may employ a Neural Bellman-Ford Network (NBFNet) as the inference engine, transforming the diagnostic prediction task for a specific patient into a generalized path integral problem on a graph. For a given query... ,in This is the patient instance node (source node). To achieve the target relationship of "having a diagnosis", this invention searches for and calculates all possible nodes that can reach the candidate diagnostic conclusion. Path confidence.

[0039] In some implementations, a graph neural network based on path representation learning is used to transfer information on the dynamic knowledge graph, including modeling the information transfer process of the graph neural network on the dynamic knowledge graph as an iterative solution process of the generalized Bellman-Ford algorithm.

[0040] Specifically, for any path in the graph from the source node to the target node... Its path representation Defined as the generalized product of all edge representations on the path: From the source node To the target node The total logical confidence is the generalized sum of the representations of all possible paths: .

[0041] In the specific iterative solution process, each iteration Nodes in The state update formula can be expressed as: .

[0042] In this model, this implementation maps mathematical operators to medical logic rules. The message passing function (MSG) represents the logical AND operation, used to calculate the representation vector of a single evidence chain, reflecting the joint confidence (or joint probability) of that chain. For example, a path will only generate a high-weight representation vector pointing to the logical determination of "cardiac amyloidosis" if all three conditions—"interventricular septal thickening," "extensive LGE," and "elevated T1 value"—are simultaneously satisfied on a single path. After calculating the representation vector of a single path, a pre-defined scoring mapping operator, such as a fully connected layer or a linear layer, can be used to project the high-dimensional representation vector to a scalar space, and then processed by an activation function to obtain the confidence score of that path. Subsequently, the aggregation function (AGG) can perform a logical OR operation based on the confidence scores of each path to combine the confidence of multiple paths for the same diagnostic conclusion. Evidence representing different sources, such as morphological and histological evidence chains, can independently support the same diagnostic conclusion, thereby enhancing the confidence of the diagnosis. In this implementation, the MSG function ensures that the evidence of all nodes in the path must be valid simultaneously through operator calculation, while the AGG function achieves the complementarity of evidence from different paths through accumulation or maximum value operation, thereby simulating the reasoning process of symbolic logic in the neural network structure.

[0043] By generating path weights from the path representation vectors obtained through the iterative solution process, the system can extract the set of paths with the highest weights as the target inference chain.

[0044] By introducing an explicit reasoning mechanism based on path representation learning, this invention changes the traditional "black box" output method of AI models. This invention not only possesses the powerful feature fitting capabilities of deep learning but also retains the rigor of symbolic logic rules. The system can highlight the most relevant chain of evidence for each diagnostic conclusion, such as patient -> segment 7 -> high T2 value -> edema -> meets Lake Louise criteria -> myocarditis, enabling radiologists to clearly trace the diagnostic basis. This white-box reasoning process effectively avoids the logical contradictions that may arise from purely probabilistic models, significantly improving the accuracy of differential diagnosis of complex cardiomyopathy and the credibility of clinical decision-making.

[0045] Step S104: The target inference chain is used as context input to the pre-trained large language model, and a prefix tree is dynamically constructed based on the target inference chain. During the decoding stage of the pre-trained large language model, the prefix tree is used to mask and filter the output probability distribution. At the same time, the original numerical attributes related to the inference chain are filled into the generated text sequence through the numerical pointer copy operation to obtain a diagnostic report.

[0046] In this step, the Top-K high-weight inference chains output from step S103 are first converted into natural language descriptions through prompting engineering, and then used as context input to the pre-trained large language model. The pre-trained large language model refers to a deep learning model (such as the Llama series or a finely tuned Transformer model) pre-trained on a large-scale medical corpus. By converting structured inference paths, such as "segment 7 → T2 elevation (62ms) → meets myocarditis criteria," into prompts, the system can guide the model to focus on specific imaging evidence. In the subsequent text generation process, the system can use a prefix tree built based on the target inference chain and numerical pointer copying operations to process qualitative descriptions and quantitative values, respectively.

[0047] In some implementations, dynamically constructing a prefix tree based on the target inference chain includes: extracting relevant evidence subgraphs from the dynamic knowledge graph according to the target inference chain; retrieving phrase templates matching the evidence subgraphs from a preset medical phrase library, and removing candidate words that conflict with the attributes of instance nodes in the dynamic knowledge graph to obtain a valid phrase sequence; compiling the filtered valid phrase sequence into the prefix tree, wherein the nodes of the prefix tree represent tokens, and the edges represent the next token sequence allowed to be generated under constraints, thereby realizing lexical-level pruning based on the inference results. The construction process of the prefix tree is the process of mapping semantically validated medical phrases to the model's vocabulary space. Specifically, the valid phrase sequence can be segmented according to the vocabulary of the pre-trained large language model to obtain corresponding token sequences; then, the sequence is mounted into a tree structure according to the prefix sharing principle, so that each node of the tree represents a token that the model is allowed to predict under medical logic constraints in the current decoding step.

[0048] For example, if the inference chain determines the LGE pattern as "subendothelial enhancement," the system will physically remove conflicting terms such as "transmural" or "patchy" when retrieving templates to ensure that these erroneous descriptions do not enter the branches of the prefix tree. This implementation eliminates the possibility of the model generating descriptions that contradict the pathological facts by combining a static phrase library with a dynamic inference pathology.

[0049] After constructing the prefix tree, in some implementations, the constrained lexical prefix tree is used to mask and filter the output probability distribution during the decoding stage of the pre-trained large language model. This includes: at each time step of the decoder, performing path matching in the constrained lexical prefix tree based on the generated historical sequence to locate the current state node, and determining all child nodes of the located current state node as the current valid child node set; setting the probability of lexical units not in the valid child node set to negative infinity to force the model-generated description to maintain consistency with the logic corresponding to the original numerical attributes. This implementation provides a hard-constrained state-dependent intervention method, for each lexical unit in the vocabulary. When it belongs to the set of valid child nodes At that time, its mask vector ;otherwise, Adding this mask vector to the output probability distribution before Softmax ensures that the report text adheres to both natural language fluency and strict medical logic.

[0050] Meanwhile, to further improve the accuracy of quantitative data, in some implementations, the original numerical attributes in the inference chain can be filled into the generated text sequence through a numerical pointer copy operation. This includes: when the decoder of the pre-trained large language model generates a preset numerical placeholder, a copy mode is triggered to calculate the similarity distribution between the current hidden state of the decoder and the original numerical attributes that have been mapped and associated in the dynamic knowledge graph using an attention mechanism, and the original numerical attribute with the highest probability is directly extracted and filled into the text sequence. Specifically, when the model generates, such as...<NUM_SLOT> When a special tag is selected, the model switches from prediction mode to extraction mode, no longer performing probabilistic word generation, but instead relying on attention distribution. The corresponding precise value, such as 1350ms or 40%, is read directly from the numerical registry. This implementation method completely solves the problem of precision loss when processing numerical information in traditional models, ensuring that the quantitative description in the diagnostic report is completely consistent with the image measurement values.

[0051] This invention injects the logical rigor of knowledge graphs into the generation process of large language models by constructing a reasoning-driven controlled generation framework. Through prefix tree constraints and numerical pointer copying operations, this invention eliminates the illusion phenomenon commonly found in medical report generation and ensures that reports meet the requirements of serious medical scenarios in terms of semantic expression, numerical accuracy, and adherence to clinical standards, greatly improving the practical value of automated diagnostic systems.

[0052] In some implementations, the above method further includes a step of performing consistency verification on the diagnostic report. Specifically, the diagnostic report is restored to a structured text graph, a consistency score is calculated between the text graph and the dynamic knowledge graph, and the logical consistency between the diagnostic report and the dynamic knowledge graph is determined based on the score result. The calculation dimensions of the consistency score include the node matching degree between the text graph and the dynamic knowledge graph, the consistency of the relationship between nodes, and the consistency of the negative detection for the lesion description.

[0053] Structured text graphs refer to semantic networks extracted from generated diagnostic report text using natural language processing techniques, representing entities and their logical relationships in graph form. Consistency scoring quantifies the degree of logical matching between the generated diagnostic text and the original image feature graph. Node matching typically refers to whether the anatomical locations mentioned in the verification text (e.g., the inferior lateral wall of the basal segment) spatially overlap with the lesion locations actually detected in the images. Relationship consistency between nodes typically refers to whether the modification relationships between the sign attributes described in the verification text (e.g., subepicardial enhancement) and the anatomical locations are consistent with the original knowledge graph. Negation detection consistency is a security audit mechanism used to ensure that the report text does not erroneously deny objectively existing lesions in the images or falsely affirm signs that do not exist in the images.

[0054] The diagnostic report can be converted into a structured text graph using a pre-trained medical relation extraction model, such as RadGraph. The system first performs named entity recognition on the text generated in step S104 to extract clinical entities, and simultaneously performs relation extraction to determine the structured relationships between entities. For example, for the description "subepicardial LGE visible on the inferior wall of the basal segment," the medical relation extraction model can convert it into a local graph structure containing nodes for "inferior wall of the basal segment (location)," "subepicardial (pattern)," and "LGE (sign)," and establish corresponding modifying relation edges.

[0055] When calculating the consistency score, the system can extract the graph from the text. The dynamic knowledge graph constructed in step S102 Graph matching calculations are performed. For example, in a Fabry disease screening scenario, if the original image atlas records "significantly decreased basal segment T1 value," but the generated report text incorrectly describes it as "increased T1 value," a logical alarm will be triggered, causing the consistency score to fall below a preset threshold. If the score is below the threshold, the system will refuse to publish the report and send feedback to the generation module for recalibration; only when the score meets the logical consistency requirements will the final diagnostic report be output. This implementation introduces a graph-based reverse verification mechanism, adding further security to the automatic report generation process. It effectively avoids logical paradoxes caused by the randomness of model probability sampling, ensuring that the final report delivered to clinicians is rigorous and reliable.

[0056] Referring to Figure 2, Figure 2 shows an architectural block diagram of a cardiac magnetic resonance imaging report generation system based on a multidimensional tissue feature knowledge graph according to an embodiment of the present invention. As shown in the figure, the system can be divided into five main modules.

[0057] The feature extraction module is used to perform refined quantization processing on the input raw image data. It may include a motion correction unit, an anatomical segmentation unit, and a multidimensional feature quantization unit. The motion correction unit performs anti-artifact processing on the multimodal CMR raw image data to ensure spatial consistency; the anatomical segmentation unit identifies anatomical segments in the image based on a preset deep neural network and extracts the myocardial contour; the multidimensional feature quantization unit calculates various morphological, functional biomechanical, and tissue quantification indicators in parallel on a model such as the AHA 17 segment based on the segmentation results, and encapsulates them into multidimensional tissue feature tensors before outputting them to the next module.

[0058] The dynamic knowledge graph construction module is used to transform image data into medical knowledge. This module is connected to a static ontology graph, which stores pre-built ontology logic containing nodes of prior medical knowledge and diagnostic conclusions. The graph instantiation unit generates patient-specific image feature instance nodes based on the extracted feature tensors. The attribute mapping unit establishes a mapping association between the instance nodes and the original numerical attributes in the feature tensors by performing attribute assignment and mounting operations, and mounts the instance nodes to the static ontology graph, ultimately generating a dynamic patient-specific graph, i.e., a dynamic knowledge graph. This module also provides the original numerical attribute mapping signals to subsequent generation modules to ensure the transmission of numerical accuracy.

[0059] The logical evidence path reasoning module simulates the logical judgment process of clinicians within a knowledge space. The path representation learning unit utilizes a graph neural network to transmit information on the dynamic patient-specific atlas, calculating path representations between nodes. The reasoning chain determination unit searches for logical evidence paths based on path weights and extracts the set of paths with the highest confidence as the target reasoning chain. This reasoning chain not only contains the diagnostic conclusion but also carries the anatomical and pathological evidence supporting that conclusion.

[0060] The graph-constrained report generation module implements logically controlled text synthesis. The prefix tree construction unit extracts evidence subgraphs based on the target inference chain and retrieves matching templates from the medical phrase library, dynamically compiling them into a prefix tree. The controlled decoding unit uses the prefix tree to mask and filter the decoding process of the pre-trained Large Language Model (LLM), forcing the model to output a lexical sequence that conforms to medical logic. Simultaneously, the numerical pointer copying unit is activated when the model generates numerical placeholders. Based on the numerical attribute mapping signal from the attribute mapping unit, it directly extracts the original numerical attributes from the graph attributes and fills them into the text sequence, thereby producing a structured diagnostic report.

[0061] The consistency verification module provides quality assurance for the generated report. The text reconstruction unit uses entity relation extraction technology to reconstruct the structured diagnostic report into a text graph. The scoring unit then determines whether the report has logical consistency by calculating a consistency score between the text graph and the original dynamic patient-specific graph. If the score does not meet the standard, the system generates a feedback optimization signal to the controlled decoding unit for regeneration until a final approved diagnostic report is produced, thereby ensuring the rigor and reliability of the system output.

[0062] Referring to Figure 3, Figure 3 shows a schematic diagram of the motion correction algorithm based on robust principal component analysis (RPCA) according to an embodiment of the present invention, which takes the input multimodal image sequence matrix affected by motion artifacts as input. Converted into high-fidelity anatomical background images.

[0063] Specifically, the motion correction unit can first stack the acquired multimodal cardiac magnetic resonance images along the time dimension or sequence dimension to construct an image sequence matrix. Subsequently, the system processes the low-rank matrix. sparse matrix Lagrange multipliers and penalty parameters Initialization is performed. During the iterative solution phase, this unit employs the Inexact Augmented Lagrange Multiplier Method (IALM) for iterative computation. In each iteration, the system alternately updates the multipliers and parameters based on their current states. , , as well as The value of is used to make the objective convex optimization model gradually approach the global optimum.

[0064] After each update, the system automatically executes convergence judgment logic. If the current matrix factorization residual does not reach the preset threshold or does not meet the iteration termination condition, it is fed back to the iteration update stage to continue calculation; if the convergence condition is met, the iterative solution process stops, and the low-rank matrix is ​​extracted synchronously from the final optimization result. With sparse matrices Among them, the low-rank matrix The image representing the spatially consistent cardiac anatomy, i.e., the corrected background image, can be output to subsequent anatomical segmentation units; while the sparse matrix... This represents dynamic interference components such as respiratory movements and heartbeat artifacts. Through this iterative decomposition mechanism, the system can eliminate positioning deviations caused by motion interference during the acquisition process from the physical data, in order to extract high-precision original numerical attributes subsequently.

[0065] Referring to Figure 4, which illustrates a schematic diagram of the architecture of a multi-view feature extraction network according to an embodiment of the present invention, this network significantly improves the segmentation accuracy and feature quantization quality of key cardiac structures by fusing spatial information from different anatomical perspectives.

[0066] Specifically, as shown in Figure 4, the network employs a dual-stream parallel input mode, with its input layer receiving long-axis slice sequences (LAX) and short-axis slice sequences (SAX) from two-chamber or four-chamber views, respectively. During the feature encoding stage, the system utilizes the SwinUNETR encoding stream to extract features from both sequences. Since the long-axis view has higher spatial resolution along the vertical axis, while the short-axis view provides anatomical details of the cross-section, a multi-view fusion attention mechanism can be used to achieve complementarity.

[0067] In the interactive attention computation unit, vectors from the long-axis feature extraction stream are mapped to a query matrix (Query, Q), while vectors from the short-axis feature extraction stream are mapped to a key / value matrix (Key / Value, K / V). By calculating the attention weight distribution between the query matrix and the key matrix, this invention can identify clear longitudinal boundary features in long-axis images and inject their spatial dependencies into short-axis features. This effectively utilizes the high-resolution information of the long axis to compensate for the boundary blurring of short-axis images at the apex or mitral valve plane caused by partial volume effects.

[0068] During the decoding and output stages, the feature vectors incorporating multi-view information are fed into the upsampling decoder for scale restoration. The decoded output is further mapped into the AHA 17-segment model, thus providing subsequent steps with a multi-dimensional organization feature tensor that is highly aligned in spatial location and highly preserved in physical value.

[0069] Referring to Figure 5, Figure 5 illustrates a schematic diagram of the dynamic knowledge graph data structure and neural symbol reasoning logic according to an embodiment of the present invention. Figure 5 uses the clinical diagnostic pathway of myocarditis as an example to demonstrate how the present invention, starting from image features, achieves interpretable auxiliary diagnosis through information transmission on the graph.

[0070] During the inference initiation phase, the system uses patient instance nodes in the dynamic knowledge graph as the search starting point, such as "Patient Instance (Segment 7)" shown in Figure 5. This node encapsulates the extracted multidimensional tissue features of a specific anatomical location and serves as the source of the inference flow, transmitting activation signals to downstream nodes.

[0071] During the information transmission phase, the system executes a message passing mechanism based on path representation learning. As shown in Figure 5, the activation signal starts from instance 7 of segment 7 and retrieves relevant image sign nodes along the preset logical relationship edges. For example, since the original numerical attribute of this instance node shows an abnormally high T2 value, the signal flows to the "high T2 value" node, further stimulating the intermediate pathological manifestation of "myocardial edema"; at the same time, combined with the qualitative analysis of the LGE sequence, the signal flows to the "LGE positive" node. In this process, the path representation learning unit uses a path-based iterative solution algorithm (such as the generalized Bellman-Ford algorithm) to iteratively calculate the representation vector of each path, linking discrete medical signs into a logically meaningful feature chain.

[0072] During the logical judgment phase, the system aggregates multiple evidence paths based on prior medical knowledge, such as the 2018 Lake Louise Standard, as shown in the figure as "compliant with LLC criteria". When searching for logical evidence paths, aggregation functions can be used to represent logical OR operations and combine the confidence levels of multiple paths to the diagnostic conclusion. As shown by the bolded line segment in Figure 5, the path supported by both "edema" and "LGE positive" is identified as a high-weight path, which ultimately converges to the diagnostic conclusion node, such as "acute myocarditis".

[0073] Referring to Figure 6, Figure 6 illustrates a schematic diagram of decoding control for a pre-trained large language model (LLM) based on a prefix tree and numerical pointer copying mechanism according to an embodiment of the present invention. According to the present invention, by introducing external knowledge constraints into the decoding space, it is ensured that the generated report text conforms to medical logic and possesses high numerical accuracy.

[0074] As shown in Figure 6, after receiving the target reasoning chain from the logical evidence path reasoning module, the graph constraint report generation module starts two controlled generation paths in parallel: Path A: qualitative description constraints based on prefix tree mask filtering.

[0075] The prefix tree construction unit first extracts key medical terms and logical predicates from the target inference chain, and then dynamically constructs a prefix tree by combining it with a pre-defined medical phrase library. In each decoding step of the LLM Decoder (Large Language Model Decoder), the controlled decoding unit retrieves this prefix tree in real time to obtain the current set of legal candidate lexical units.

[0076] As shown in Figure 6, if the target inference chain clearly identifies the lesion location as "subepidermal" rather than "transmural," the prefix tree will classify words like "Transmural" that do not conform to imaging evidence as illegal paths. When calculating the Softmax probability distribution, the system uses a masking operation to force the predicted probabilities of these illegal words to be set to a fixed value. This physically shields the possibility of logical fallacies in the output, effectively ensuring the accuracy of the report at the qualitative description level.

[0077] Path B: Quantitative numerical precision mapping based on pointer network copying mechanism.

[0078] To prevent large language models from exhibiting numerical illusions when processing specific clinical indicators such as T1 / T2 values ​​and ejection fraction, this invention introduces numerical pointer duplication. When a predefined numerical placeholder appears in the sequence generated by the large language model, according to this invention, the probabilistic prediction mode is paused, and the pointer network is activated.

[0079] As shown in Figure 6, the numerical pointer copying unit directly locates the original mounted value stored in the feature tensor attribute, such as "T1: 1350ms", based on the attribute mapping signal from the dynamic knowledge graph. By performing the numerical copying operation, the real physical values ​​obtained in the image quantization stage are directly filled into the text sequence without any model resampling or probability transformation, thereby completely overcoming the numerical calculation bias that may occur in the model and ensuring the absolute accuracy of the quantitative description.

[0080] Through the synergistic effect of path A and path B, the system ultimately generates sentences such as "...the T2 value in the subepidermal region is 1350ms...", achieving a unification of qualitative facts and quantitative data. Based on the architecture shown in Figure 6, this invention changes the general large language model's mode of generating content solely based on the probability distribution of the corpus. By injecting hard constraints from the graph into the decoding process, this invention further ensures the medical seriousness of the generated text.

[0081] Referring to Figure 7, Figure 7 illustrates a consistency logic check flowchart based on text reconstruction when diagnosing Fabry Disease according to an embodiment of the present invention. This consistency logic check based on text reconstruction can effectively eliminate logical illusions that may arise from large language models.

[0082] In the input comparison source region, the system simultaneously acquires two parallel equivalent data streams to construct an audit baseline: the first stream is a dynamic graph serving as the objective truth value. The first path contains the original numerical attributes of the instance node, such as the typical quantitative feature of Fabry disease shown in Figure 7: "Basal segment interventricular septum T1 value: significantly reduced"; the second path is the generated report text to be verified and the semantic information it carries, such as the text description containing hallucinations shown in Figure 7: "Basal segment interventricular septum T1 value increased".

[0083] In the text reconstruction and attribute extraction area, the system invokes the text reconstruction unit to extract entity relations from the generated report text, enabling comparison between unstructured text and structured graphs within the same semantic space. This unit remaps the text descriptions into structured nodes, constructing the reconstructed text graph. As shown in Figure 7, the system reverse-engineers the attribute "T1 value increases" predicted by the model from the text graph, completing the transformation from natural language to logical attribute, and providing isomorphic input for subsequent consistency determination.

[0084] In the consistency calculation and decision-making region, the consistency scoring unit simultaneously receives attribute information from both of the aforementioned regions and executes a graph matching scoring algorithm. During this process, the system can identify logical conflicts, such as detecting a discrepancy between the original numerical attribute (physical truth: decreased) and the textual reconstruction attribute (predicted value: increased) of an instance node. Based on this attribute conflict, the consistency score calculated by the scoring unit will be significantly reduced.

[0085] Next, the system executes the scoring criteria judgment logic. If the score is lower than the preset threshold, i.e., "No: Hallucination detected," the system can determine that there is a factual error in the generated content, immediately generate a feedback optimization signal, and refuse to publish the current report. This signal carries conflict point information back to the graph constraint report generation module, forcing it to re-trigger the controlled decoding process.

[0086] If the score is higher than the preset threshold, i.e. "Yes: Logical consistency is achieved", it means that the text description and the original numerical attributes in the image have been aligned, and the system allows the output of the final approved diagnostic report.

[0087] Through the verification architecture shown in Figure 7, this invention overcomes the shortcomings of traditional medical report generation that relies solely on text probability distribution. By introducing hard constraints based on original numerical attributes, it ensures the clinical seriousness and factual accuracy of each output report.

[0088] In summary, the cardiac magnetic resonance imaging (MRI) report generation method and system provided in this invention closely integrate image processing, logical reasoning, and report generation, and incorporate verification rules during the automatic report writing process using a large language model. When generating reports, the system can directly reference the original numerical attributes extracted from the images and, in conjunction with a preset vocabulary filtering mechanism, verify the original numerical attributes in real time during the generation of each description, thereby effectively avoiding discrepancies between the description and the actual situation. Simultaneously, the target reasoning chain provided by the system transforms complex feature calculations into clear derivation steps, allowing doctors to intuitively understand the basis for their conclusions. Through this self-verification design, this invention not only reduces the burden on radiologists in organizing data and writing reports but also provides reliable technical support for the standardization of cardiac MRI diagnosis.

[0089] Figure 8 shows an electronic device provided in at least one embodiment of the present invention. The device includes a memory and a processor. The memory is used to store computer instructions that can be executed on the processor. The processor is used to implement the cardiac magnetic resonance report generation method based on multidimensional tissue feature knowledge graph as described in any embodiment or implementation of the present invention when executing the computer instructions.

[0090] At least one embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cardiac magnetic resonance report generation method based on a multidimensional tissue feature knowledge graph as described in any embodiment or implementation of the present invention.

[0091] It should be emphasized that the descriptions of specific embodiments and the processes shown in the accompanying drawings in this specification are intended to clearly illustrate the technical solutions of the present invention, rather than limiting its scope of protection to the specific structures, sequences, or combinations described therein. Without departing from the core spirit and principles of the present invention, those skilled in the art will understand that the various technical features described in the embodiments can be arbitrarily reorganized or split; the execution order of the steps can also be adjusted; and in some cases, parallel or multi-tasking processing is also feasible. Therefore, any modifications, equivalent substitutions, and improvements made to the above embodiments based on the technical essence of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating cardiac magnetic resonance imaging reports based on a multidimensional tissue feature knowledge graph, characterized in that, include: Multimodal cardiac magnetic resonance imaging (MRI) images of patients are acquired, motion correction is performed on the images, and anatomical segments in the corrected images are identified. Multidimensional tissue feature tensors corresponding to each anatomical segment are extracted. These multidimensional tissue feature tensors encapsulate raw numerical attributes characterizing anatomical morphology and tissue properties. Patient-specific image feature instance nodes are generated based on the multidimensional tissue feature tensors. A mapping relationship is established between these instance nodes and the raw numerical attributes in the multidimensional tissue feature tensors. These instance nodes are then attached to a pre-constructed static ontology graph to complete the construction of a dynamic knowledge graph. The static ontology graph includes medical prior art. The system verifies knowledge and predefined diagnostic conclusion nodes; it uses a graph neural network based on path representation learning to transmit information on the dynamic knowledge graph, searches for logical evidence paths from the instance nodes to the diagnostic conclusion nodes, and determines the target inference chain based on the path weights; it inputs the target inference chain as context into a pre-trained large language model, and dynamically constructs a prefix tree based on the target inference chain; during the decoding stage of the pre-trained large language model, it uses the prefix tree to mask and filter the output probability distribution, and simultaneously fills the original numerical attributes associated with the inference chain into the generated text sequence through a numerical pointer copying operation to obtain a diagnostic report.

2. The method according to claim 1, characterized in that, Motion correction of the multimodal cardiac magnetic resonance images includes: constructing the multimodal cardiac magnetic resonance images into an image sequence matrix. The image sequence matrix Decomposed into low-rank matrices representing anatomical structures and sparse matrix representing dynamic components By establishing and solving for minimizing the nuclear norm and A convex optimization model based on norm is used to extract background images with respiratory motion artifacts removed. As the corrected image.

3. The method according to claim 1, characterized in that, Extracting multidimensional tissue feature tensors corresponding to each anatomical segment includes: using a deep neural network that integrates a multi-view fusion attention mechanism to jointly process long-axis and short-axis images to divide the heart into standard anatomical segments; and extracting morphological indicators, strain dynamic indicators, late gadolinium enhancement pattern classification labels, and mapping sequence quantitative values ​​in parallel for each standard anatomical segment to construct the multidimensional tissue feature tensor.

4. The method according to claim 1, characterized in that, Establishing a mapping association between the instance node and the original numerical attributes in the multidimensional organizational feature tensor includes: defining an attribute matrix in the data structure of the dynamic knowledge graph, wherein the row index of the attribute matrix has a one-to-one correspondence with the instance node; and using attribute assignment operations to store the original numerical attributes in the multidimensional organizational feature tensor into the attribute slots in the attribute matrix corresponding to the instance node.

5. The method according to claim 1, characterized in that, The method utilizes a graph neural network based on path representation learning to perform information transmission on the dynamic knowledge graph, searches for logical evidence paths from the instance node to the diagnostic conclusion node, and determines the target inference chain based on path weights. This includes: modeling the information transmission process of the graph neural network on the dynamic knowledge graph as an iterative solution process based on path algebra; wherein, when searching for logical evidence paths, the message passing function of the graph neural network is used to represent logical AND operations to calculate the representation vector of a single path, and an aggregation function is used to represent logical OR operations to merge the confidence levels of multiple paths to the same diagnostic conclusion node; when determining the target inference chain, path weights are generated based on the path representation vectors calculated through the iterative solution process, and the set of paths with the highest weights is extracted as the target inference chain.

6. The method according to claim 1, characterized in that, The dynamic construction of a prefix tree based on the target inference chain includes: extracting relevant evidence subgraphs from the dynamic knowledge graph according to the target inference chain; retrieving phrase templates that match the evidence subgraphs from a preset medical phrase library, and removing candidate words that conflict with the attributes of instance nodes in the dynamic knowledge graph to obtain a valid phrase sequence; and compiling the filtered valid phrase sequence into the prefix tree, wherein the nodes of the prefix tree represent lexical units, and the edges represent the next lexical unit sequence that is allowed to be generated under constraints.

7. The method according to claim 6, characterized in that, In the decoding stage of the pre-trained large language model, the prefix tree is used to perform masking filtering on the output probability distribution, including: at each time step of the decoder, path matching is performed in the prefix tree according to the generated historical sequence to locate the current state node, and all child nodes of the located current state node are determined as the current effective child node set; the probability of the lexical not in the effective child node set is set to negative infinity to force the description generated by the model to be consistent with the logic corresponding to the original numerical attribute.

8. The method according to claim 1, characterized in that, The process of filling the original numerical attributes in the inference chain into the generated text sequence by copying numerical pointers includes: when the decoder of the pre-trained large language model generates a preset numerical placeholder, a copy mode is triggered to calculate the similarity distribution between the current hidden state of the decoder and the original numerical attributes that have been mapped and associated in the dynamic knowledge graph using an attention mechanism, and the original numerical attribute with the highest probability is directly extracted and filled into the text sequence.

9. The method according to claim 1, characterized in that, The method further includes a step of verifying the consistency of the diagnostic report: restoring the diagnostic report to a structured text graph, calculating the consistency score between the text graph and the dynamic knowledge graph, and determining the logical consistency between the diagnostic report and the dynamic knowledge graph based on the score result. The calculation dimensions of the consistency score include the node matching degree between the text graph and the dynamic knowledge graph, the consistency of the relationship between nodes, and the consistency of the negative detection for the lesion description.

10. A cardiac magnetic resonance imaging report generation system based on a multidimensional tissue feature knowledge graph, characterized in that, include: A feature extraction module is used to acquire multimodal cardiac magnetic resonance images of the patient, perform motion correction on the multimodal cardiac magnetic resonance images, identify anatomical segments in the corrected images, and extract multidimensional tissue feature tensors corresponding to each anatomical segment. These multidimensional tissue feature tensors encapsulate original numerical attributes characterizing anatomical morphology and tissue properties. A dynamic knowledge graph construction module is used to generate patient-specific image feature instance nodes based on the multidimensional tissue feature tensors, establish a mapping relationship between the instance nodes and the original numerical attributes in the multidimensional tissue feature tensors, and attach the instance nodes to a pre-constructed static ontology graph to complete the construction of the dynamic knowledge graph. The static ontology graph contains medical prior knowledge. The system includes: a recognition and predefined diagnostic conclusion node; a logical evidence path reasoning module, which uses a graph neural network based on path representation learning to transmit information on the dynamic knowledge graph, search for logical evidence paths from the instance node to the diagnostic conclusion node, and determine the target reasoning chain based on the path weights; and a graph constraint report generation module, which takes the target reasoning chain as context input to a pre-trained large language model, dynamically constructs a prefix tree based on the target reasoning chain, uses the prefix tree to mask and filter the output probability distribution during the decoding stage of the pre-trained large language model, and fills the original numerical attributes associated with the reasoning chain into the generated text sequence through a numerical pointer copying operation to obtain a diagnostic report.

Images