A multi-source information fusion and multi-agent collaborative forensic virtual autopsy cause of death inference system and method

Abstract

The application discloses a multi-source information fusion and multi-agent collaborative forensic virtual autopsy cause of death inference system and method, and belongs to the technical field of medical image processing, artificial intelligence and forensic identification. The system comprises preprocessing, feature extraction, chain reasoning, multi-agent collaboration and report generation modules. For the image artifacts of the corpse, a generative adversarial network is used for correction; a three-dimensional hybrid network is used to extract image features, and a domain language model is used to extract text features, which are fused through cross attention. According to the four-level reasoning of 'image anomaly-> injury mechanism-> direct cause of death-> mode of death', and the analysis and report are completed by multiple agents. The application can automatically assist in cause of death inference, provide visual basis, improve the efficiency and objectivity of identification, and is suitable for forensic identification institutions.

Description

Technical Field

[0001] This invention belongs to the fields of medical image processing, artificial intelligence, and forensic identification technology. Specifically, it relates to a system and method that uses a deep learning model to automatically identify and analyze features of forensic CT images and provide auxiliary determination of the cause of death. This invention deeply integrates cutting-edge artificial intelligence technologies such as computer vision, natural language processing, multimodal fusion, and multi-agent collaboration with forensic pathology, and designs a complete intelligent auxiliary identification solution tailored to the unique characteristics of forensic autopsy computed tomography images. Background Technology

[0002] In modern forensic examinations, computed tomography (CT) technology has been increasingly applied to post-mortem examinations due to its non-invasive, rapid, and permanent preservation advantages; this is known as forensic CT or PMCT. However, the interpretation of CT images is highly dependent on professional knowledge and experience, and is typically performed by experienced radiologists. Identifying the cause of death is a crucial step in forensic investigations, and the use of post-mortem CT images offers significant advantages in terms of non-invasiveness, objectivity, and time and cost-effectiveness.

[0003] Currently, the application of this technology in the field of forensic medicine faces the following problems: 1. Mismatch of professional experience. Forensic practitioners in China often lack systematic clinical radiology interpretation experience, making it difficult for them to accurately identify key pathological or injury features from complex CT images.

[0004] 2. High communication costs. If clinical radiologists are invited to assist in interpreting the images, although they are proficient in imaging, they often lack knowledge of forensic pathology and do not understand the forensic connection between injury, disease, and cause of death, resulting in low communication efficiency and interpretation results that may deviate from the needs of forensic identification.

[0005] 3. Highly subjective and lacking standardization. Traditional image interpretation relies on manual methods, resulting in highly subjective outcomes. Furthermore, it is difficult to perform standardized and quantitative analysis on minute features in images, such as minor fractures, air embolisms, and specific patterns of bleeding.

[0006] 4. Postmortem artifacts. After death, a corpse undergoes complex changes such as the formation of livor mortis, putrefaction and gas production, and tissue autolysis. These postmortem processes produce numerous artifacts in images, severely interfering with accurate interpretation.

[0007] 5. Difficulty in integrating heterogeneous data from multiple sources. Forensic cause-of-death analysis requires comprehensive consideration of data from multiple modalities, including autopsy, toxicology analysis, pathology analysis, forensic imaging, and DNA analysis. Each analytical method involves a large amount of data, making integration complex, time-consuming, and prone to errors.

[0008] 6. Low efficiency in generating forensic reports. The writing of forensic reports requires the integration of multi-source data, including case details, examination records, image analysis, and test reports, as well as the selection of terminology and arrangement of conclusions according to judicial appraisal standards. Currently, this process is still mainly done manually, which is inefficient and prone to errors.

[0009] Therefore, there is an urgent need for an intelligent system that can integrate forensic medical expertise, systematically integrate multi-source information, deduce the cause of death through chain reasoning that conforms to forensic logic, and automatically generate identification reports, in order to break through the existing talent bottleneck and improve the objectivity and efficiency of forensic identification. Summary of the Invention

[0010] The purpose of this invention is to provide an automatic CT image recognition and cause-of-death deduction system for forensic scenarios, which integrates multi-source information fusion and multi-agent collaboration. This system constructs and trains a deep neural network model and a multi-agent collaborative architecture, enabling it to learn the intrinsic relationship between image features, case information, examination data and final cause of death in a large number of historical cases, and to perform chain reasoning according to forensic logic. This allows it to comprehensively analyze new cases and output auxiliary judgment evidence and identification reports.

[0011] To achieve the above objectives, the present invention provides the following technical solution: This system mainly includes the following modules: 1. Data Acquisition and Preprocessing Module. Receives raw PMCT image data (DICOM format), performs preprocessing operations such as denoising, normalization, and bone and soft tissue segmentation, and performs specific corrections for postmortem artifacts in cadaver CT images, including putrefaction gas production, blood deposition, and tissue autolysis.

[0012] 2. Multimodal Feature Extraction Module. Utilizing pre-trained or end-to-end trained deep learning models (such as 3D-CNN, Vision Transformer), this module automatically extracts key features from images, such as fracture lines, intracranial hemorrhage, pneumothorax, and fat embolism. Simultaneously, it uses domain-adapted language models to extract semantic features from case text and examination reports.

[0013] 3. AI Model Chain Reasoning Module. Following the forensic logic sequence of "image anomalies → injury mechanism → direct cause of death → manner of death," the module uses a four-level chained reasoning model to infer the cause of death step by step. The first level detects and locates image anomalies; the second level infers the injury mechanism based on combinations of image anomalies; the third level integrates multi-source information to determine the direct cause of death; and the fourth level comprehensively outputs a classification of the manner of death.

[0014] 4. Multi-agent collaboration module. This module constructs a system architecture that integrates a large language model, image analysis model, forensic knowledge base, and report generation agent to enable task invocation, decision collaboration, and result verification among the agents.

[0015] 5. Results Output and Report Generation Module. Based on the fusion features and chain-based reasoning results, the module outputs the probability of cause of death category and key imaging evidence, and displays the reasoning basis through heatmap visualization technology. At the same time, it automatically generates an appraisal report according to forensic appraisal standards.

[0016] System Architecture The overall architecture of this invention includes five layers: First layer: Multi-source data acquisition and preprocessing layer.

[0017] This layer includes a PMCT / MRI image acquisition unit for importing raw DICOM files from cadaver CT / MRI scans, supporting unified access to multimodal images from CT and MRI; an unstructured data acquisition unit for inputting or importing text data such as case summaries, crime scene investigation records, and post-mortem examination records; and a structured data acquisition unit for inputting structured data such as toxicology test results, individual information, and postmortem time intervals.

[0018] The image preprocessing module adjusts the window width and level of the DICOM file, optimizing the bone window, lung window, and soft tissue window, performing voxel normalization, and multi-sequence registration. Simultaneously, it invokes the postmortem artifact correction module, utilizing a GAN network to fuse PMI and environmental parameters to correct artifacts related to putrefaction gas production, blood deposition, and tissue autolysis. The text preprocessing module performs word segmentation on the case text and identifies forensic-related entities, such as keywords related to injury location and manner of death.

[0019] The second layer: multimodal feature extraction layer.

[0020] (1) Image feature extraction model: 3D Hybrid CNN-Transformer (3D ResFormer). A hybrid architecture was designed to address the three-dimensional characteristics of forensic CT images and the dual requirements of global context and local details in forensic tasks. The input is a 3D CT image block, 128×128×128×1; Stage 1 is shallow feature extraction, with 7×7×7 3D convolutional layers, stride=2, 64 channels + BN + ReLU; 3D max pooling 3×3×3, stride=2; Stages 2-4 are Hybrid coding blocks ×4, each coding block contains a 3D residual convolutional module for local feature extraction and a 3D windowed self-attention module for global dependency modeling; Stage 5 is global feature aggregation, 3D global average pooling; Flatten + FC layer outputs a 1024-dimensional image feature vector. The core innovation lies in the dual-path design: CNN is responsible for extracting local texture and edge information, and is sensitive to fractures and bleeding, while Transformer is responsible for modeling long-distance dependencies, and is sensitive to the extent of damage and organ deformation; 3D windowed attention is used to reduce the computational complexity from O(n²) to O(n); residual connections ensure the flow of gradients in deep networks.

[0021] (2) Text Feature Extraction Model: BioMed-Forensic BERT. Considering the professional nature and forensic characteristics of the case text, a domain-adapted pre-trained language model was adopted. The input was the case text, such as "Male, 45 years old, found at a construction site with head injuries"; after WordPiece segmentation, [CLS] and [SEP] tags were added; embedding was performed using Token Embeddings + Position Embeddings + Segment Embeddings; after 12 layers of Transformer encoding, 12 heads, a 768-dimensional multi-head self-attention network with hidden layers + a 3072-dimensional hidden layer feedforward network + LayerNorm + residual connections; the [CLS] tag output was 768-dimensional, plus the mean pooling of all token outputs to 768-dimensional; after domain adaptation: forensic entity recognition head + FC layer mapping to 1024 dimensions. In terms of training strategy, domain-adaptive pre-training was first performed on PubMed summaries and MIMIC clinical reports in the biomedical corpus, and then fine-tuned using forensic identification reports.

[0022] (3) Multimodal data fusion module: A two-stage cross-attention fusion module (DSCMAF) is adopted to allow the two modalities to guide and enhance each other during the interaction process. Stage 1 is text-guided image enhancement: using text features as queries to guide the model to focus on areas in the image related to the case description. Stage 2 is collaborative cross-attention: realizing bidirectional interaction and aligning the two modalities in a common semantic space: image → text attention and text → image attention. Feature fusion adopts a multi-scale fusion strategy: F_fusion=concat([F_v2t,F_t2v,elementwise_max(F_v2t,F_t2v), elementwise_mean(F_v2t, F_t2v)]); F_compact = FC_layer(F_fusion) (2048 → 512 dimensions); output = softmax(FC_layer(F_compact)) (512 → number of cause of death categories).

[0023] The third layer: AI model chain inference layer (core layer).

[0024] This layer adopts a chain-like reasoning framework with four levels of models connected sequentially, and makes inferences step by step according to the analytical logic of forensic experts: "abnormal imaging → injury mechanism → direct cause of death → manner of death".

[0025] Level 1 – Image Anomaly Detection and Localization Model: Input image feature vectors, output multi-label classification results: fracture, intracranial hemorrhage, pneumothorax, pericardial hemorrhage, ascites, and 3D coordinate localization of abnormal areas. Level 2 – Injury Mechanism Inference Model: Integrates the image anomaly detection results from Level 1, case text features, and post-mortem examination features. Through a Transformer-based multimodal fusion encoder, outputs the probability distribution of injury mechanisms: blunt force trauma, stab wound, gunshot wound, asphyxiation, electric shock, and burns. Level 3 – Direct Cause of Death Determination Model: Integrates the injury mechanism inference results from Level 2, image anomaly features, toxicology detection results, and individual information. Through a multimodal gating fusion network, outputs the probability distribution of direct causes of death: traumatic brain injury, hemorrhagic shock, cardiac injury, asphyxiation, poisoning, and disease. Level 4 – Manner of Death Classification Model: Integrates the direct cause of death determination results from Level 3, injury mechanisms, case information, and scene information. Based on a large language model, performs classification reasoning and outputs the manner of death: homicide, suicide, accident, natural death, and cause of death unknown. In chained reasoning, the output of the previous model serves as one of the input features of the next model, thereby accumulating and transmitting information along the reasoning chain.

[0026] Fourth layer: Multi-agent collaboration layer.

[0027] A forensic intelligent workstation system based on multi-agent collaboration is constructed, with the following division of labor among the agents: The scheduling agent is responsible for task decomposition and distribution, assigning tasks to appropriate specialized agents based on user-inputted analysis requirements. The image analysis agent uses the 3D ResFormer model to automatically analyze CT / MRI images and outputs anomaly detection results. The knowledge retrieval agent performs case retrieval and knowledge matching based on a forensic knowledge base, providing reference for the reasoning process. The reasoning and decision-making agent synthesizes the outputs of all agents and makes a comprehensive judgment according to chain-like reasoning logic. The report generation agent automatically generates an identification report based on the reasoning results and forensic identification standards. The verification agent performs consistency verification and confidence assessment on the output results.

[0028] The calling, decision-making, and verification mechanism among the various intelligent agents is as follows: After receiving a user request, the scheduling intelligent agent calls the image analysis intelligent agent to perform image analysis; the analysis results are transmitted to the reasoning and decision-making intelligent agent, while the knowledge retrieval intelligent agent retrieves relevant cases and knowledge from the forensic knowledge base; the reasoning and decision-making intelligent agent integrates all information to infer the cause of death; the verification intelligent agent evaluates the confidence level and verifies the consistency of the inference results; finally, the report generation intelligent agent outputs an identification report.

[0029] Fifth layer: Result output and report generation layer.

[0030] The image annotation and visualization unit highlights key areas where models at each level make judgments using heatmaps (such as Grad-CAM) on raw CT / MRI images. The chain reasoning display unit presents the complete reasoning chain of "image abnormality → injury mechanism → direct cause of death → manner of death" and the confidence level of each step in the form of cards or flowcharts. The forensic report generation unit integrates case information, image analysis results, reasoning conclusions at each level, and multi-agent collaborative output to automatically generate forensic reports that include terminology selection and conclusion arrangement according to forensic standards, and reserves interfaces for manual review and editing.

[0031] Model training methods The core of this system lies in model training, which includes the following steps: Step A: Construct a multi-level training dataset. Collect historical closed forensic cases, including PMCT / MRI DICOM image data, case texts, post-mortem / autopsy records, toxicology reports, and final cause-of-death determinations. Forensic experts annotate abnormal areas in the images, including the type, location, and extent of the abnormality. For each case, label each level of a four-level inference chain: image abnormality → injury mechanism → direct cause of death → manner of death.

[0032] Step B: Data Augmentation. Randomly rotate, translate, scale, and flip the images in the training set to increase the amount of data and prevent the model from overfitting.

[0033] Step C: Pre-training and Fine-tuning. First, the backbone network is pre-trained using a large, publicly available medical image dataset, such as NIH ChestX-ray, to enable it to extract general medical image features. Then, the entire network is fine-tuned end-to-end using the forensic-specific dataset constructed in Step A to optimize model parameters and make it accurately suited for forensic cause-of-death deduction tasks.

[0034] Step D: Model Validation. Test the trained model using an independent validation set to evaluate its accuracy, recall, and AUC metrics, ensuring that the model performance meets the usage requirements.

[0035] Step E: Chained Joint Optimization. A course-based learning strategy is employed, first training each level of the model separately, then performing end-to-end joint fine-tuning. A chained loss function is designed to backpropagate the errors of each level of the model, achieving overall optimization.

[0036] Multi-task joint damage time inference method This system also supports multi-task joint injury time inference: the input is shoulder joint MRI multi-sequence images: T1WI, T2WI, PDWI; the network structure is based on a multi-branch parallel feature analysis architecture; the multi-task outputs include: task 1 is whether the tendon is ruptured / damaged, task 2 is the nature of the injury, and task 3 is whether there is a coexisting underlying lesion; the training strategy adopts joint optimization of multi-task loss function, each task shares the underlying feature extraction network, and the upper layer branches into dedicated classification heads for each task.

[0037] Automatic generation method of medical imaging reports for forensic identification This system also supports the automatic generation of medical imaging reports for forensic identification: It integrates multi-source data including case information, imaging examination records, image analysis results, and laboratory reports. The report generation process is as follows: Step 1, retrieve the target identification report template and obtain the target chapter titles from the template; Step 2, extract the required content for each chapter from the multi-source data; Step 3, select terminology and arrange conclusions according to forensic identification standards; Step 4, generate the initial version of the identification report. After report generation, a manual review and editing interface is provided, allowing forensic personnel to modify, annotate, and confirm the report content.

[0038] The beneficial effects of this invention are as follows: This invention's system incorporates forensic expertise and chain-like reasoning logic, enabling ordinary forensic personnel to access expert-level image analysis support without relying on external radiologists. It automatically completes preliminary screening, feature annotation, and chain-like reasoning, significantly reducing image interpretation time. The data-driven analysis eliminates the subjectivity of personal experience, achieving standardized and quantifiable feature descriptions.

[0039] It can not only analyze CT images, but also combine case texts and test data to achieve a deep correlation between images, cases, and tests, making the auxiliary inference results more in line with reality.

[0040] Through the division of labor and collaboration among multiple intelligent agents, the entire process of automatic task decomposition, professional model invocation, knowledge retrieval enhancement, and automatic report generation is made intelligent, which significantly improves the system's processing efficiency and decision-making accuracy.

[0041] Specialized corrections to artifacts such as gas production from putrefaction and blood deposition in cadaver CT images effectively eliminate interfering factors and improve the model's recognition accuracy on cadaver images. It supports multi-dimensional inference of injury time, providing more comprehensive quantitative evidence for forensic identification. With the continuous accumulation of new cases, the system can undergo periodic incremental training to continuously optimize model accuracy, forming a virtuous cycle.

[0042] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a detailed flowchart of the image preprocessing module in this invention; Figure 3 This is a schematic diagram of the training and verification process of the deep learning model in this invention. Detailed Implementation

[0044] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0045] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. Example

[0048] A forensic virtual autopsy intelligent cause-of-death deduction system based on multi-source information fusion and multi-agent collaboration.

[0049] This system comprises a multi-source data acquisition and preprocessing layer, a multi-modal feature extraction layer, an AI model chain inference layer, a multi-agent collaboration layer, and a result output and report generation layer, which are connected in sequence.

[0050] 1. Multi-source data acquisition and preprocessing layer: The forensic image acquisition unit is used to import raw DICOM files from cadaver CT / MRI scans, supporting unified access to multimodal images from CT and MRI. The unstructured data acquisition unit is used to input or import text data such as case summaries, crime scene investigation records, and post-mortem examination records. The structured data acquisition unit is used to input structured data such as toxicology test results, individual information, and postmortem interval (PMI).

[0051] The image preprocessing module adjusts the window width and level of the DICOM files, optimizing the bone window, lung window, and soft tissue window, performing voxel normalization, and multi-sequence registration. Simultaneously, it invokes the postmortem artifact correction module, using a GAN network to fuse PMI and environmental parameters to correct artifacts related to putrefaction gas production, blood deposition, and tissue autolysis. The text preprocessing module performs word segmentation on the case text and identifies forensic-related entities, including injury sites and manner of death.

[0052] 2. Multimodal feature extraction layer: The image feature extraction model employs a 3D Hybrid CNN-Transformer (3D ResFormer). The input is a 3D CT image patch (128×128×128×1). Stage 1 is shallow feature extraction, with 7×7×7 3D convolutional layers, stride=2, 64 channels + BN + ReLU; 3D max pooling 3×3×3, stride=2; Stages 2-4 are Hybrid coding blocks ×4, each coding block containing a 3D residual convolutional module for local feature extraction and a 3D windowed self-attention module for global dependency modeling; Stage 5 is global feature aggregation, with 3D global average pooling; the Flatten + FC layer outputs a 1024-dimensional image feature vector.

[0053] The text feature extraction model uses BioMed-Forensic BERT. The input is the case text; it is segmented by WordPiece and marked with [CLS] and [SEP] tags; it is embedded through Token Embeddings + Position Embeddings + Segment Embeddings; it is encoded by a 12-layer Transformer with 12 heads and a 768-dimensional hidden layer multi-head self-attention; a feedforward network with a 3072-dimensional hidden layer; LayerNorm + residual connections; the [CLS] tag output is 768-dimensional + the average pooling of all token outputs is 768-dimensional; it is then mapped to 1024 dimensions through a domain adaptation layer, a forensic entity recognition head + FC layer.

[0054] The multimodal data fusion module employs a two-stage cross-attention fusion module (DSCMAF). Stage 1 is text-guided image enhancement: using text features as queries, it guides the model to focus on areas in the image relevant to the case description. Stage 2 is collaborative cross-attention: enabling bidirectional interaction and aligning the two modalities in a shared semantic space. Feature fusion uses a multi-scale fusion strategy, outputting 512-dimensional fused features.

[0055] 3. AI Model Chain Inference Layer: The first level is an image anomaly detection and localization model. It takes image feature vectors as input and outputs multi-label classification results: fracture, intracranial hemorrhage, pneumothorax, pericardial hemorrhage, ascites; and 3D coordinate localization of the abnormal area. The second level is an injury mechanism inference model. It integrates the image anomaly detection results from the first level, case text features, and post-mortem examination features. Through a Transformer-based multimodal fusion encoder, it outputs the probability distribution of injury mechanisms: blunt force trauma, stab wound, gunshot wound, asphyxiation, electric shock, and burns. The third level is a direct cause-of-death determination model. It integrates the injury mechanism inference results from the second level, image anomaly features, toxicology detection results, and individual information. Through a multimodal gating fusion network, it outputs the probability distribution of direct cause of death: traumatic brain injury, hemorrhagic shock, cardiac injury, asphyxiation, poisoning, and disease. The fourth level is a mode of death classification model. It integrates the direct cause of death determination results from the third level, injury mechanisms, case information, and scene information. Based on a large language model, it performs classification reasoning and outputs the mode of death: homicide, suicide, accident, natural death, and cause of death unknown. The output of the previous model serves as one of the input features of the next model, enabling the accumulation and transmission of information along the inference chain.

[0056] 4. Multi-agent collaboration layer: The scheduling agent is responsible for task decomposition and distribution. The image analysis agent uses the 3D ResFormer model to automatically analyze CT / MRI images and outputs image anomaly detection results. The knowledge retrieval agent performs case retrieval and knowledge matching based on the forensic knowledge base. The reasoning and decision-making agent integrates the output results of each agent and makes a comprehensive judgment according to chain reasoning logic. The report generation agent automatically generates an appraisal report based on the reasoning results and forensic appraisal standards. The verification agent performs consistency verification and confidence assessment on the output results.

[0057] After receiving a user request, the scheduling agent calls the image analysis agent to perform image analysis; the analysis results are transmitted to the reasoning and decision-making agent, while the knowledge retrieval agent retrieves relevant cases and knowledge from the forensic knowledge base; the reasoning and decision-making agent integrates all information to infer the cause of death; the verification agent evaluates the confidence level and verifies the consistency of the inference results; finally, the report generation agent outputs an identification report.

[0058] 5. Results Output and Report Generation Layer: The image annotation and visualization unit highlights key areas for judgment by each level of the model on the original CT / MRI images using heatmaps. The chain reasoning display unit presents the complete reasoning chain of "image abnormality → injury mechanism → direct cause of death → manner of death" and the confidence level of each step in the form of cards or flowcharts. The forensic report generation unit integrates case information, image analysis results, reasoning conclusions at each level, and multi-agent collaborative output to automatically generate an forensic report that includes terminology selection and conclusion arrangement according to forensic standards, and reserves interfaces for manual review and editing. Example

[0059] Model training methods based on historical big data.

[0060] Step A: Construct a multi-level training dataset. Collect historical closed forensic cases, including PMCT / MRI DICOM image data, case texts, post-mortem / autopsy records, toxicology reports, and final cause-of-death determinations. Forensic experts annotate abnormal areas in the images, including: abnormality type, location, and extent. For each case, label each level of a four-level inference chain: image abnormality → injury mechanism → direct cause of death → manner of death.

[0061] Step B: Data Augmentation. Randomly rotate, translate, scale, and flip the images in the training set to increase the amount of data and prevent the model from overfitting.

[0062] Step C: Pre-training and Fine-tuning. First, the backbone network is pre-trained using a large, publicly available medical image dataset to enable it to extract general medical image features. Then, the entire network is fine-tuned end-to-end using the forensic-specific dataset constructed in Step A to optimize model parameters and make it accurately suited for forensic cause-of-death deduction tasks.

[0063] Step D: Model Validation. Test the trained model using an independent validation set to evaluate its accuracy, recall, and AUC metrics, ensuring that the model performance meets the usage requirements.

[0064] Step E: Chained Joint Optimization. A course-based learning strategy is employed, first training each level of the model separately, then performing end-to-end joint fine-tuning. A chained loss function is designed to backpropagate the errors of each level of the model, achieving overall optimization. Example

[0065] A multi-task collaborative method for forensic injury time estimation.

[0066] The input consists of multiple MRI sequences of the shoulder joint: T1WI, T2WI, and PDWI. The network structure is based on a multi-branch parallel feature analysis architecture. Multi-task outputs include: Task 1: whether the tendon is ruptured / damaged; Task 2: the nature of the injury; and Task 3: whether there are underlying lesions. The training strategy employs joint optimization of a multi-task loss function, with each task sharing the underlying feature extraction network, and the upper layers branching into dedicated classification heads for each task. Example

[0067] A method for automatically generating medical imaging reports for forensic identification.

[0068] The input integrates multi-source data including case information, imaging examination records, image analysis results, and test reports. The report generation process is as follows: Step 1, call the target forensic report template and obtain the target chapter titles from the template; Step 2, extract the required content for each chapter from the multi-source data; Step 3, select terminology and arrange conclusions according to forensic identification standards; Step 4, generate the initial version of the forensic report. After the report is generated, a manual review and editing interface is provided, allowing forensic personnel to modify, annotate, and confirm the report content.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A forensic virtual autopsy cause-of-death deduction system based on multi-source information fusion and multi-agent collaborative forensic methods, characterized in that, include: The multi-source data acquisition and preprocessing module is used to acquire CT / MRI image data, case text data, post-mortem / autopsy examination data, and toxicology detection data of the target object, and to perform standardized preprocessing. The multimodal feature extraction module is used to extract visual features of images, semantic features of text, and encoded features of structured data based on deep learning networks, and to perform multimodal fusion through a cross-attention mechanism. The AI ​​model chain reasoning module is used to infer the cause of death step by step through a four-level chain reasoning model, following the logical order of "image abnormalities → injury mechanism → direct cause of death → manner of death". The multi-agent collaboration module is used to achieve task decomposition, model invocation, knowledge retrieval, comprehensive judgment, report generation, and result verification by coordinating the work of scheduling agents, image analysis agents, knowledge retrieval agents, reasoning and decision-making agents, report generation agents, and verification agents. The results output and report generation module is used to output the probability of cause of death category and key imaging evidence, and automatically generate an appraisal report according to the judicial appraisal standards.

2. The system according to claim 1, characterized in that, The image preprocessing in the multi-source data acquisition and preprocessing module is configured to perform the following operations: window width and window level adjustment, voxel normalization, multi-sequence registration, and postmortem artifact correction based on generative adversarial networks; the postmortem artifact correction integrates postmortem interval time and environmental parameters to correct for putrefactive gas production, blood deposition, and tissue autolysis artifacts.

3. The system according to claim 1, characterized in that, The multimodal feature extraction module includes: an image encoder based on a 3D hybrid CNN-Transformer architecture, wherein the CNN path extracts local texture and edge information, and the Transformer path models long-distance dependencies; a text encoder based on a domain-adaptive language model; and a structured feature encoder; the multimodal fusion adopts a two-stage cross-attention fusion mechanism.

4. The system according to claim 1, characterized in that, The AI ​​model chain inference module includes: Level 1: Imaging anomaly detection model, outputting the abnormality types and three-dimensional localization of fractures, bleeding, and pneumothorax; Level 2: Damage mechanism inference model, outputting the probability distribution of damage mechanisms such as blunt impact, sharp object puncture, and asphyxiation; Level 3: Direct cause of death determination model, outputting the probability distribution of direct causes of death for traumatic brain injury and hemorrhagic shock; Level 4: Death Method Classification Model, outputting classification results for homicide, suicide, accident, natural causes, and unexplained death; The output of the previous model serves as one of the input features of the next model, enabling the accumulation and transmission of information along the inference chain.

5. The system according to claim 1, characterized in that, The multi-agent collaborative module includes: The scheduling agent is responsible for task decomposition and distribution. Image analysis intelligent agents are used to call image analysis models for automatic analysis. A knowledge retrieval intelligent agent is used for case retrieval and knowledge matching based on a forensic knowledge base. The reasoning and decision-making agent is used to synthesize the outputs of various agents for comprehensive judgment. The report-generating agent is used to automatically generate appraisal reports based on reasoning results and forensic appraisal standards. The verification agent is used to evaluate the confidence level and verify the consistency of the output results.

6. The system according to claim 1, characterized in that, The result output and report generation module also includes a multi-task joint injury time inference function, which automatically identifies whether the rotator cuff tendon is ruptured, the rupture location, and the degree of injury based on shoulder joint MRI multi-sequence fusion, and simultaneously outputs multi-dimensional results of the nature of the injury and whether it is accompanied by underlying lesions.

7. The system according to claim 1, characterized in that, The output of results and the generation of the appraisal report in the report generation module are configured to: integrate multi-source data such as case information, image examination records, image analysis results, and test reports; and call the target appraisal report template; The terminology and conclusions were selected and arranged in accordance with judicial appraisal standards; a preliminary appraisal report was generated and an interface for manual review and editing was reserved.

8. A method for intelligent cause-of-death deduction based on the system according to any one of claims 1 to 7, characterized in that, Includes the following steps: S1. Collect PMCT / MRI images, case texts, post-mortem / autopsy examination data, and toxicology test data of the target subject; S2. Preprocess the multi-source data, including image preprocessing, text preprocessing, and structured data encoding; S3. Using a multimodal feature extraction model, extract image features, text features, and structured features respectively, and then perform multimodal fusion. S4. Input the multimodal features into the four-level chain inference model in sequence, and output the image anomaly detection results, damage mechanism inference results, direct cause of death determination results and death mode classification results step by step; S5. Task scheduling, knowledge retrieval, comprehensive judgment and result verification are performed through a multi-agent collaborative module; S6. Based on the chain reasoning results and the multi-agent collaborative output, generate an output result that includes a display of the reasoning chain and an identification report.

9. A computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the method of claim 8.

Images