Craniocerebral injury mechanism analysis and reasoning system fused with multi-modal large model
By aligning and inferring traumatic brain injury data using a multimodal large model, the problem of information fragmentation in traditional forensic medicine is solved, enabling efficient and interpretable injury mechanism analysis and identification report generation, and supporting natural language queries.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI MEDICAL UNIV
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-26
Smart Images
Figure CN122291008A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and forensic medicine, specifically to a system for analyzing and reasoning about the injury mechanism of craniocerebral injury by integrating a multimodal large model. Background Technology
[0002] In forensic practice, traumatic brain injury is a common and deadly type of injury. Accurate deduction of its injury mechanism (such as the weapon used, direction of impact, and magnitude of force) is crucial for determining the nature of the case and assigning responsibility. Traditional methods primarily rely on forensic experts to make comprehensive judgments based on autopsy reports, imaging data (such as CT / MRI), and crime scene investigation information, which has the following limitations: 1. Medical images, mechanical simulation data (such as finite element stress cloud diagrams), text-based autopsy records, and case descriptions usually belong to different modalities, formats, and semantic systems, making it difficult to achieve cross-modal semantic alignment and joint reasoning; 2. Existing computer-aided systems mostly use image recognition or rule matching to locate the injury area, lacking causal logic modeling of "why this injury morphology is formed", and cannot answer mechanism questions such as "whether a blunt object striking at a 30° angle can cause an occipital fracture with subdural hematoma". 3. Forensic identification reports are highly dependent on individual experience, time-consuming, subjective, and lack automated generation and reliable measurement mechanisms. 4. Existing systems struggle to support complex queries in natural language form, making it impossible to achieve "semantic-driven" case retrieval and scenario simulation.
[0003] In recent years, visual-language large models (such as LLaVA and GPT-4V) have made breakthroughs in cross-modal understanding, but there is still no systematic solution to apply them to the reasoning of forensic injury mechanisms; especially in the area of combining biomechanical simulation data (such as skull stress distribution and strain energy density) with multi-source heterogeneous text / image information for causal reasoning, there is still a gap.
[0004] Therefore, a system for analyzing and reasoning about the injury mechanism of traumatic brain injury that integrates a multimodal large model is proposed to solve the problems mentioned above. Summary of the Invention
[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a system for analyzing and reasoning about the injury mechanism of craniocerebral injury by integrating a multimodal large model. It has advantages such as multimodal deep alignment, causal-level injury mechanism reasoning, automatic generation of structured reports, and semantically driven interactive queries. It solves the technical problems of traditional forensic injury analysis, such as the fragmentation of medical images, biomechanical simulations, and text reports, the reliance on experience for inference of the injury process and the lack of physical interpretability, low identification efficiency, and inability to support complex semantic retrieval.
[0006] (II) Technical Solution To achieve the aforementioned objectives of "multimodal deep alignment and causal-level injury mechanism reasoning," this invention provides the following technical solution: a system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model, comprising, Multimodal data preprocessing and alignment module: used to receive cranial medical images, biomechanical simulation stress cloud maps, forensic autopsy text and case descriptions, perform spatial registration of the cranial medical images and biomechanical simulation stress cloud maps, and map the registered image data and the forensic autopsy text and case descriptions to a shared semantic embedding space through a multimodal encoder; Multimodal large model inference engine: It includes a visual-language large model fine-tuned by forensic traumatic brain injury case data, receives multimodal embedding vectors in the shared semantic embedding space, and outputs injury mechanism inference results in natural language form. The inference results include the type of injury-causing tool, the range of impact angle, the biomechanical mechanism, and the causal logic chain of injury formation. Intelligent report generation module: used to connect to the multimodal large model inference engine and convert the injury mechanism inference results into a structured forensic medical identification report draft containing injury feature summary, injury mode inference basis, mechanical simulation supporting evidence and inference credibility score; Interactive Natural Language Query Interface: This interface receives natural language query statements input by the user, converts them into multidimensional search conditions through a semantic parser, performs scenario matching in a database storing historical multimodal injury cases, and returns medical images, stress cloud maps, and injury parameter sets for the matched cases.
[0007] Preferably, the biomechanical simulation stress cloud map is a principal stress distribution map of the skull surface or interior generated by finite element analysis, and is spatially aligned with the cranial CT image in the standard Talairach coordinate system in the form of a pixel-level heat map.
[0008] Preferably, the multimodal encoder includes an image encoder and a text encoder. The image encoder extracts features from the corresponding regions of the cranial medical image and the stress cloud map, respectively. The text encoder encodes the damage description fragments in the forensic autopsy text. The three achieve cross-modal alignment in a shared semantic space by comparing and learning a loss function.
[0009] Preferably, the visual-language large model is a multimodal basic model based on the Transformer architecture, and its training data includes no less than 500 fully labeled quadruple samples. Each quadruple sample includes a cranial CT image, a corresponding finite element stress cloud map, an autopsy report text, and a conclusion on the injury mechanism confirmed by forensic experts.
[0010] Preferably, the inference credibility score is a weighted comprehensive score, with the weight coefficients corresponding to the simulation-image space overlap, the number of similar cases in the case library, the average confidence of the large model output token, and the matching degree of the preset forensic rule library.
[0011] Preferably, the initial draft of the structured forensic medical report includes a fixed field template, which includes the injury location, fracture type, estimated intracranial hemorrhage volume, inferred geometric features of the injuring tool, and estimated impact direction vector and energy range.
[0012] Preferably, the semantic parser in the interactive natural language query interface is based on a domain-adapted named entity recognition model and a dependency parser, which extracts keywords such as anatomical location, injury type, bleeding threshold, and tool shape from the query statement as structured query parameters.
[0013] Preferably, each case entry in the multimodal injury case database stores registered CT images, stress cloud maps, finite element model input parameters, autopsy text, and injury mechanism tags. Each entry is organized through a vector index structure, supporting joint retrieval based on embedding similarity and parameter constraints.
[0014] Preferably, the causal logic chain is generated by the large model during the reasoning process based on a preset causal prompt template. The causal prompt template limits the output format to a three-part natural language structure: "[Injury-causing action] leads to [mechanical response], which in turn triggers [damage manifestation]".
[0015] (III) Beneficial Effects Compared with existing technologies, this invention provides a system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model, which has the following beneficial effects: 1. This system for analyzing and reasoning about the injury mechanism of craniocerebral injury by integrating a multimodal large model, for the first time, uses a multimodal data preprocessing and alignment module to spatially register biomechanical simulation stress cloud maps as an independent visual modality with cranial CT images in a standard anatomical coordinate system. Combined with forensic autopsy text and case description, it uses a multimodal encoder to achieve deep fusion of heterogeneous data from three sources—image, mechanics, and text—in a shared semantic space. This effectively breaks down the barriers of isolated information and semantic fragmentation of various modalities in traditional forensic analysis, providing a structured and aligned input foundation for subsequent high-precision injury mechanism reasoning.
[0016] 2. This system for analyzing and reasoning about the injury mechanism of traumatic brain injury, which integrates a multimodal large model, introduces a preset causal prompt template based on a visual-language large model finely tuned in the field of forensic medicine. It forces the model to output a three-part causal logic chain of "injury action - mechanical response - injury manifestation" that conforms to biomechanical principles. This not only enables accurate inference of the injury tool, the angle of impact, and the energy range, but also significantly improves the physical interpretability and professional credibility of the reasoning results. At the same time, the system supports complex queries driven by natural language and the automatic generation of structured identification reports, which greatly shortens the forensic identification cycle, reduces subjective bias, and promotes the development of forensic injury analysis towards intelligence, standardization, and traceability. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the structure of the traumatic brain injury mechanism analysis and reasoning system that integrates a multimodal large model, as described in this invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Please see Figure 1 As shown, the system for analyzing and reasoning about the injury mechanism of traumatic brain injury, which integrates a multimodal large model, includes: Multimodal data preprocessing and alignment module: used to receive cranial medical images, biomechanical simulation stress cloud maps, forensic autopsy text and case descriptions, perform spatial registration of the cranial medical images and biomechanical simulation stress cloud maps, and map the registered image data and the forensic autopsy text and case descriptions to a shared semantic embedding space through a multimodal encoder; Multimodal large model inference engine: It includes a visual-language large model fine-tuned by forensic traumatic brain injury case data, receives multimodal embedding vectors in the shared semantic embedding space, and outputs injury mechanism inference results in natural language form. The inference results include the type of injury-causing tool, the range of impact angle, the biomechanical mechanism, and the causal logic chain of injury formation. Intelligent report generation module: used to connect to the multimodal large model inference engine and convert the injury mechanism inference results into a structured forensic medical identification report draft containing injury feature summary, injury mode inference basis, mechanical simulation supporting evidence and inference credibility score; Interactive Natural Language Query Interface: This interface receives natural language query statements input by the user, converts them into multidimensional search conditions through a semantic parser, performs scenario matching in a database storing historical multimodal injury cases, and returns medical images, stress cloud maps, and injury parameter sets for the matched cases.
[0020] This system for analyzing and reasoning about the injury mechanism of traumatic brain injury, which integrates a multimodal large model, includes the following steps: Step 1: Multimodal data input and standardization preprocessing Receive four types of raw data: Cranial medical imaging (such as CT, DICOM format); Biomechanical simulation stress cloud map (principal stress thermal map generated by finite element method, PNG format); Forensic autopsy report; Case description text.
[0021] Rigid spatial registration was performed on CT images and stress contour maps in the standard Talairach coordinate system to ensure consistent anatomical positions (registration error RMSE ≤ 1.5 mm).
[0022] Step 2: Cross-modal semantic alignment and embedding mapping The image encoder extracts the visual features of the registered CT region block and the corresponding stress cloud map region block respectively; A text encoder performs semantic encoding on injury description fragments in autopsy texts; The three elements are mapped to a unified shared semantic embedding space (768 dimensions) through a contrastive learning loss function (such as InfoNCE loss), achieving deep alignment of "image-mechanics-text".
[0023] Step 3: Causal-level injury mechanism reasoning The aligned multimodal embedding vectors are input into a large visual-language model (such as LLaVA-1.6 based on the Transformer architecture) fine-tuned by forensic cases. The model outputs natural language inference results containing the type of injuring tool, the range of the impact angle, the biomechanical mechanism, and the causal logic chain, based on the preset causal prompt template (format: "[injurious action] leads to [mechanical response], which in turn leads to [damage manifestation]").
[0024] Step 4: Generation of Structured Forensic Medical Examination Report The intelligent report generation module parses the reasoning results and automatically generates a draft structured report based on a fixed field template. The content includes: Location of injury, type of fracture; Estimation of intracranial hemorrhage volume; Inference of the geometric features of the injuring tool; Estimation of strike direction vector and energy range; Mechanical simulations support the evidence; Inference credibility score (based on simulation-image overlap, number of similar cases, model confidence, and rule base matching degree weighted calculation).
[0025] Step 5: Interactive semantic search (optional extended functionality) Users input natural language queries (such as "possible causes of depressed parietal fracture with bleeding <20ml"), and the system extracts structured query parameters through a domain-adapted semantic parser (combining named entity recognition and dependency parsing). It then performs a joint retrieval based on vector similarity and numerical constraints in a multimodal injury case database, returning CT images, stress contour maps, and injury parameter sets for matching cases.
[0026] Example 1 This system for analyzing and reasoning about the injury mechanism of traumatic brain injury, which integrates a multimodal large model, has an overall architecture that includes a multimodal data preprocessing and alignment module, a multimodal large model inference engine, an intelligent report generation module, and an interactive natural language query interface. The modules work together to complete the entire process from inputting raw multi-source data to inferring the injury mechanism, generating reports, and performing semantic retrieval.
[0027] I. Multimodal Data Preprocessing and Alignment Module This module first receives four types of input data: (1) cranial CT images (512×512×Z voxels, DICOM format); (2) The principal stress distribution map of the skull (i.e., the biomechanical simulation stress cloud map) generated by finite element simulation software (such as ANSYS or Abaqus) is output in PNG format with the resolution aligned with the CT slices. (3) Structured or unstructured forensic autopsy report text; (4) Case background description text.
[0028] The spatial registration process is as follows: A rigid registration algorithm based on affine transformation is used to align the stress cloud map with the corresponding CT slice in the standard Talairach coordinate system. Registration error is evaluated by root mean square error (RMSE). When RMSE ≤ 1.5mm, it is considered a valid registration; otherwise, manual verification or re-simulation is triggered. After registration, the CT images and stress cloud maps are cropped into 224×224 pixel regions and divided according to anatomical regions (frontal, parietal, temporal, and occipital).
[0029] Multimodal embedding mapping: The image encoder uses ViT-Base (VisionTransformer, embedding dimension d=768) to extract features from the corresponding regions of the CT scan and stress contour maps, respectively, to obtain visual embedding vectors. and ; The text encoder uses Sentence-BERT (also outputting a 768-dimensional vector) to encode injury description fragments in the autopsy text, such as "linear fracture of the right temporal bone" and "epidural hematoma of approximately 25 ml," to obtain the text embedding.
[0030] The three methods are jointly optimized through comparative learning of the loss function (InfoNCE loss) to ensure that the loss function corresponding to the same damage event is optimized. The minimum distance is achieved in the shared semantic space, while the maximum distance is achieved between different events, thus enabling cross-modal alignment.
[0031] II. Multimodal Large Model Inference Engine The engine is based on the LLaVA-1.6 architecture (Vision Encoder+Vicuna-13B LLM) and has been fine-tuned on the dataset dedicated to this invention. The training data contains 512 quadruple samples annotated by forensic experts. Each sample includes: a registered CT image, a stress cloud map, autopsy text, and a conclusion on the injury mechanism (such as "striking with a round-headed hammer, impact angle 35°±5°, energy approximately 60J").
[0032] Cause-and-effect prompt template design and training: During the fine-tuning phase, the outputs of all training samples are organized in a uniform causal format, for example: "The impact of a cylindrical blunt object at a 30° angle on the right temporal region resulted in a local principal stress peak of 120 MPa, which in turn caused a linear fracture of the temporal bone with epidural hematoma." This template forces the model to learn the causal chain structure of "action → mechanical response → damage manifestation"; Fine-tuning employs the LoRA (Low-Rank Adaptation) strategy, which keeps the visual encoder frozen while updating only the low-rank matrix in the language model that is related to causal reasoning, thus avoiding catastrophic forgetting.
[0033] During reasoning, the model receives aligned multimodal embeddings and automatically generates natural language reasoning results that conform to the above three-part structure, ensuring that the output has physical rationality and forensic professionalism.
[0034] III. Intelligent Report Generation Module This module parses the output of large models into structured fields and populates them into a preset report template; the report contains the following fixed fields: Location of injury (e.g., "right temporal bone"); Fracture type (linear / depressed / comminuted); Intracranial hemorrhage volume estimation (based on CT Hounsfield unit segmentation, unit: ml); Inference of the geometric characteristics of the injuring tool (diameter, curvature, contact area); Striking direction vector (represented in spherical coordinates: azimuth φ, elevation θ); Energy range estimate (unit: joules).
[0035] IV. Inference Credibility Score Calculation The system generates a quantitative credibility score for each report. The calculation formula is as follows: ; in: The spatial intersection-over-union ratio between the stress-highlighted area and the actual fracture area reflects the biomechanical-image consistency. This represents the number of the top-10 similar cases in the case library. ; Output the average confidence level of key tokens (such as "fracture" and "stress peak") for large models; The degree of matching between the reasoning result and the preset forensic rule base (0 / 1 Boolean value or fuzzy membership degree); The weighting coefficients were determined through grid search optimization: , , , ,and .
[0036] V. Interactive Natural Language Query Interface When a user inputs a scenario such as "Find all injury scenarios that could lead to depressed parietal fractures with intracranial hemorrhage of less than 20 ml", the system processes it through the following steps: The semantic parser (based on BiLSTM-CRF named entity recognition model + Stanford CoreNLP dependency parser) extracts the following keywords: anatomical location = "parietal bone", injury type = "depressed fracture", and bleeding threshold = "<20ml". Convert keywords into structured query parameters; In the multimodal injury case database, the FAISS vector index is used to perform approximate nearest neighbor retrieval on the embedded vectors, while applying a numerical constraint that the bleeding volume is ≤20ml. Return to the Top-5 matching cases and display their CT images, stress contour maps, wounding tools, impact angles, and energy parameters.
[0037] Each record in the database contains: registered CT scan, stress cloud map, finite element input (material parameters, boundary conditions), autopsy text, and injury label, supporting efficient joint retrieval.
[0038] Example 2 Scenario Example Case Background: In a criminal case, the victim suffered severe traumatic brain injury due to blunt force trauma to the head. The hospital provided the victim's cranial CT scan images, stress cloud maps generated by finite element simulation, a detailed forensic autopsy report, and a case description.
[0039] The application workflow of this system for analyzing and inferring the injury mechanism of traumatic brain injury by integrating a multimodal large model includes the following: I. Multimodal Data Reception and Preprocessing: Cranial CT images (512×512×Z voxels, DICOM format); Biomechanical simulation stress cloud map (PNG format, resolution matched to CT slices); The forensic autopsy report stated: "Linear fracture of the left frontal lobe, subdural hematoma of approximately 40 ml"; The case description mentions that an iron rod, approximately 3 centimeters in diameter, was found at the scene, suspected to be a tool used in the crime.
[0040] Spatial registration and semantic alignment: The stress contour map and the corresponding CT slice were aligned in the standard Talairach coordinate system using a rigid registration algorithm based on affine transformation, with an RMSE evaluation value of 1.3 mm.
[0041] The image encoder (ViT-Base) and the text encoder (Sentence-BERT) perform feature extraction and embedding mapping on CT region blocks, stress cloud maps, and autopsy texts, respectively.
[0042] II. Multimodal Large Model Inference: Using the finely tuned LLaVA-1.6 model to infer the above multimodal embedding vectors, the following injury mechanism analysis results were obtained: "The cylindrical blunt object struck the left forehead at a 45° angle, causing the local principal stress peak to reach 130MPa, which in turn caused a linear fracture of the frontal bone with subdural hematoma." III. Intelligent Report Generation: The automatically generated forensic medical report includes information such as the location of the injury, the type of fracture, the estimated volume of bleeding, and the inference of the weapon used to cause the injury, and is accompanied by a credibility score of S=0.87.
[0043] IV. Application of Interactive Query Interface: The user entered the following query: "Find all injury scenarios that could lead to linear fractures of the frontal bone accompanied by subdural hematoma." The system returns the top 5 similar cases, displaying their CT images, stress contour maps, injury-causing tools, and parameters.
[0044] In summary, this system for analyzing and reasoning about the injury mechanism of traumatic brain injury, which integrates a multimodal large model, achieves accurate inference of the injury-causing tool, angle of impact, and mechanical mechanism by deeply aligning medical images, biomechanical simulation data, and forensic text in a unified semantic space and combining it with a domain-fine-tuned visual-language large model for causal reasoning. It also generates a clear, well-supported, and quantifiable draft forensic medical report. Furthermore, relying on a semantically driven interactive query interface, it supports forensic experts in efficiently retrieving similar historical cases using natural language, significantly improving the objectivity, interpretability, and efficiency of traumatic brain injury mechanism analysis. This effectively overcomes the technical bottlenecks of traditional methods, such as reliance on subjective experience, information fragmentation, and lack of physical logic support.
[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model, characterized in that, include, Multimodal data preprocessing and alignment module: used to receive cranial medical images, biomechanical simulation stress cloud maps, forensic autopsy text and case descriptions, perform spatial registration of the cranial medical images and biomechanical simulation stress cloud maps, and map the registered image data and the forensic autopsy text and case descriptions to a shared semantic embedding space through a multimodal encoder; Multimodal large model inference engine: It includes a visual-language large model fine-tuned by forensic traumatic brain injury case data, receives multimodal embedding vectors in the shared semantic embedding space, and outputs injury mechanism inference results in natural language form. The inference results include the type of injury-causing tool, the range of impact angle, the biomechanical mechanism, and the causal logic chain of injury formation. Intelligent report generation module: used to connect to the multimodal large model inference engine and convert the injury mechanism inference results into a structured forensic medical identification report draft containing injury feature summary, injury mode inference basis, mechanical simulation supporting evidence and inference credibility score; Interactive Natural Language Query Interface: This interface receives natural language query statements input by the user, converts them into multidimensional search conditions through a semantic parser, performs scenario matching in a database storing historical multimodal injury cases, and returns medical images, stress cloud maps, and injury parameter sets for the matched cases.
2. The system for analyzing and reasoning about the injury mechanism of traumatic brain injury by fusing a multimodal large model as described in claim 1, characterized in that: The biomechanical simulation stress cloud map is a principal stress distribution map of the skull surface or interior generated by finite element analysis, which is spatially aligned with the cranial CT image in the standard Talairach coordinate system in the form of a pixel-level heat map.
3. The system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model as described in claim 1, characterized in that: The multimodal encoder includes an image encoder and a text encoder. The image encoder extracts features from the corresponding regions of the cranial medical image and the stress cloud map, respectively. The text encoder encodes the damage description fragments in the forensic autopsy text. The three achieve cross-modal alignment in a shared semantic space by contrastive learning loss function.
4. The system for analyzing and reasoning the traumatic brain injury mechanism by fusing a multimodal large model as described in claim 1, characterized in that: The vision-language big model is a multimodal basic model based on the Transformer architecture. Its training data includes no less than 500 fully labeled quadruplet samples. Each quadruplet sample includes a cranial CT image, a corresponding finite element stress cloud map, an autopsy report text, and a conclusion on the injury mechanism confirmed by forensic experts.
5. The system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model as described in claim 1, characterized in that: The inference credibility score is a weighted comprehensive score, with the weight coefficients corresponding to the simulation-image space overlap, the number of similar cases in the case library, the average confidence score of the large model output token, and the matching degree of the preset forensic rule library.
6. The system for analyzing and reasoning about the injury mechanism of traumatic brain injury by fusing a multimodal large model as described in claim 1, characterized in that: The initial draft of the structured forensic medical examination report includes a fixed field template, which includes the injury location, fracture type, estimated intracranial hemorrhage volume, inference of the geometric features of the injuring tool, and estimation of the impact direction vector and energy range.
7. The system for analyzing and reasoning the traumatic brain injury mechanism by fusing a multimodal large model as described in claim 1, characterized in that: The semantic parser in the interactive natural language query interface is based on a domain-adaptive named entity recognition model and a dependency parser, which extracts keywords such as anatomical location, injury type, bleeding threshold, and tool shape from the query statement into structured query parameters.
8. The system for analyzing and reasoning about the injury mechanism of traumatic brain injury by integrating a multimodal large model as described in claim 1, characterized in that: Each case entry in the multimodal injury case database stores registered CT images, stress cloud maps, finite element model input parameters, autopsy text, and injury mechanism tags. Each entry is organized through a vector index structure, supporting joint retrieval based on embedding similarity and parameter constraints.
9. The system for analyzing and reasoning the traumatic brain injury mechanism by fusing a multimodal large model as described in claim 1, characterized in that: The causal logic chain is generated by the large model during the reasoning process based on a preset causal prompt template. The causal prompt template limits the output format to a three-part natural language structure: "[Injury-causing action] leads to [mechanical response], which in turn causes [damage manifestation]".