System and method for intelligent diagnosis of brain hemorrhage caused by infective endocarditis
By performing three-dimensional segmentation and multimodal data fusion on cranial CT images, the problems of low diagnostic efficiency and low accuracy in cerebral hemorrhage caused by infective endocarditis have been solved, achieving efficient and accurate intelligent diagnosis and standardized treatment decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE SECOND AFFILIATED HOSPITAL TO NANCHANG UNIV
- Filing Date
- 2026-05-28
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies lack the quantitative extraction and definition of specific imaging features for cerebral hemorrhage caused by infective endocarditis, and cannot deeply integrate quantitative imaging parameters with diverse clinical indicators, resulting in low diagnostic efficiency, low accuracy, and a disconnect between treatment decisions.
The DICOM protocol was used to receive cranial CT images, and 3D segmentation was performed using the BLAST-CT tool. The DeepMedic convolutional neural network was used to calculate parameters such as the total volume of multifocal hemorrhage and the ratio of edema zone to intracranial hemorrhage volume. The weighted logistic regression model was used to fuse the image parameters and clinical indicators to generate the probability value of infective endocarditis, which was then visualized using Grad-CAM technology.
It has achieved efficient automated diagnosis of cerebral hemorrhage caused by infective endocarditis, improving diagnostic efficiency, increasing accuracy, reducing misdiagnosis rate, standardizing clinical decision-making, and reducing the risk of missing images.
Smart Images
Figure CN122392899A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence-assisted medical diagnostic technology, and in particular to a system and method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. Background Technology
[0002] The incidence of infective endocarditis is rising year by year, and the resulting cerebral hemorrhage complications are characterized by rapid onset, severe illness, and high mortality, posing a serious challenge to clinical diagnosis and treatment. The pathological mechanism of cerebral hemorrhage caused by infective endocarditis is complex, often manifesting as multifocal, transvascular hemorrhages, which are similar to simple hypertensive cerebral hemorrhage in imaging, easily leading to misdiagnosis and delayed diagnosis. Currently, the diagnosis of cerebral hemorrhage caused by infective endocarditis mainly relies on a simple combination of imaging examinations and clinical assessment. CT image analysis is mostly based on doctors' experience to qualitatively assess the morphology, distribution, and edema of the hemorrhage, which is inefficient and cannot quantify key parameters such as the total volume of multifocal hemorrhages and the ratio of edema zone to intracranial hemorrhage volume—which are important clues for distinguishing infective hemorrhage from other types of hemorrhage. On the other hand, while clinical indicators such as blood culture results and vegetation activity scores are crucial in the diagnosis of infective endocarditis, they are often disconnected from imaging information. The lack of a quantitative fusion mechanism to comprehensively judge the cause of hemorrhage; the defects of the existing technology are: (1) lack of quantitative extraction and definition of specific imaging features for cerebral hemorrhage caused by infective endocarditis; (2) lack of analysis model that can deeply integrate quantitative imaging parameters and diversified clinical indicators, and a large amount of diagnostic information is not fully utilized; (3) the diagnosis and subsequent treatment decision are disconnected, and it is impossible to quickly and standardizedly generate treatment recommendations that conform to clinical guidelines based on the diagnostic results. Although existing studies have achieved automatic segmentation and quantification of cerebral hemorrhage lesions, their design target is traumatic brain injury, and they have not adapted parameters for the pathological specificity of infective endocarditis, nor have they fused with the clinical diagnostic elements of infective endocarditis for decision-making. Therefore, developing a full-process intelligent diagnostic system that integrates accurate lesion quantification, multimodal data fusion, intelligent decision support and visual interaction is of great clinical significance for improving the diagnosis and treatment level of such diseases. Summary of the Invention
[0003] A method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, comprising the following steps; S1. Multimodal data input and verification: Receive cranial CT images via DICOM protocol and standardize slice thickness, simultaneously acquire structured clinical data, and complete preprocessing and quality verification; S2. Lesion segmentation and parameter calculation: The CT images were segmented at the voxel level using a three-dimensional segmentation tool to generate lesion label maps; the total volume of multifocal hemorrhage and the volume ratio of edema zone to intracranial hemorrhage were calculated based on the label maps, and the hematoma irregularity index was calculated. S3. Etiological Probability Fusion and High-Risk Early Warning: The imaging parameters and clinical indicators are probabilistically weighted and fused using a weighted fusion model to generate an infective endocarditis probability value; a first preset weight is assigned to positive blood culture results, and a second preset weight is added when the blood culture contains a specific bacterial species; the vegetation activity score is converted into a third preset weight according to a preset mapping rule; when the total volume of multifocal hemorrhage is greater than the first preset threshold and the volume ratio of the edema zone to the intracerebral hemorrhage is greater than the second preset threshold, a red early warning protocol is automatically activated. S4. Treatment suggestion generation: Dynamically matches treatment guideline thresholds and outputs surgical indications and risk classification suggestions; S5. Visualization: Grad-CAM technology is used to highlight the characteristic areas of infective endocarditis, construct a three-dimensional lesion model, and generate a DICOM SR report.
[0004] Furthermore, a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. In step S1, cranial CT images are received via the DICOM protocol and slice thickness is standardized. Structured clinical data is acquired simultaneously, and preprocessing and quality verification are completed. The specific steps are as follows: S11. Image data input: Receive cranial CT images via DICOM protocol, and normalize the scan sequence to 1mm slice thickness using inter-slice interpolation algorithm; S12. Clinical Data Input: Construct a structured interface to collect blood culture results and vegetation activity scores; S13. Data preprocessing: Perform isotropic resampling on CT images, limit CT values to the range of -15 to 100 HU and normalize the intensity; S14. Data verification: Verify that the number of image layers is not less than 20, verify the compliance of DICOM format, and trigger a resampling request for data with a median CT value exceeding 200 HU.
[0005] Furthermore, a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. In step S2, the specific steps are as follows: S21. Lesion segmentation: The BLAST-CT tool is used to perform three-dimensional segmentation on the preprocessed cranial CT image, dividing the image into three-dimensional image blocks. The DeepMedic convolutional neural network is used to classify and predict the voxels in each image block. The prediction results of all image blocks are stitched together according to the spatial location to generate a voxel-level lesion label map. S22. Volume Calculation: Based on the label map, the number of voxels for each type of lesion is counted, and the volume in milliliters is calculated in combination with the spatial resolution. The volume equals the number of voxels multiplied by the volume of a single voxel. The volume of a single voxel is calculated based on the pixel spacing and layer thickness in the DICOM header file. ,in , t is the pixel spacing, and t is the layer thickness; S23. Parameter Calculation: Calculate the ratio of edema zone volume to intracerebral hemorrhage volume and the hematoma irregularity index, where: The formula for calculating the ratio of edema zone to intracerebral hemorrhage volume is as follows: ,in This represents the total volume of the edema zone. This represents the total volume of intracerebral hemorrhage. The formula for calculating the hematoma irregularity index is as follows: ,in This represents the surface area of the hematoma. For the surface area of a sphere of the same volume, V represents the hematoma volume; S24. Quality verification: Overlay the original images to verify the consistency of segmentation and filter out isolated lesions with a volume of less than 5 mm³; S25. Data annotation: Mark the spatial location of the lesion and its maximum diameter and volume parameters in a three-dimensional coordinate system; S26. Result storage: Store lesion parameters in a medical database and support SQL queries.
[0006] Furthermore, a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. In step S3, the specific steps are as follows; S31. Feature Extraction: Extracting the total volume of multifocal hemorrhage. and the ratio of edema zone to intracerebral hemorrhage volume The formula for calculating the total volume of multifocal hemorrhage is as follows: This only includes the sum of the volumes of hemorrhage lesions within the brain parenchyma, excluding extraaxial hemorrhage, intraventricular hemorrhage, and edema zones. Let be the volume of the i-th intracranial hemorrhage lesion. S32. Clinical indicator weighting: A base weight of 0.6 is assigned to positive blood culture results, and an additional weight of 0.1 is assigned to Gram-positive bacteria. Therefore, the weight of blood culture is... A value of 0 indicates negative, a value of 0.6 indicates positive but not Gram-positive bacteria, and a value of 0.7 indicates positive Gram-positive bacteria. The vegetation activity score is converted to a weight of 0.3-0.8 using a linear mapping formula, which is as follows: , where s is the vegetation activity score of 0-10; S33. Probability Calculation: A weighted logistic regression model is used to fuse feature variables and output the probability value of infective endocarditis. The model form is as follows: , Where P is the probability value of infective endocarditis. For the intercept term, , , , where and are the regression coefficients of each variable; S34. Threshold judgment: When the total volume of multifocal hemorrhage is greater than 15ml and the ratio of the edema zone to the volume of hemorrhage in the brain parenchyma is greater than 0.5, a red warning is activated; S35. Result Verification: The stability of the probability calculation was verified by Monte Carlo simulation; S36. Generate Report: Output a structured diagnostic report containing the probability value of infective endocarditis and transmit it via the HL7 interface.
[0007] Furthermore, a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. In step S4, the specific steps are as follows; S41. Guideline matching: Dynamically compare the size of vegetations and the volume of intracerebral hemorrhage with the surgical indication thresholds; S42. Surgical indications: Emergency surgery recommendations are generated when the vegetation is larger than 10mm and the intracranial hemorrhage is greater than 30ml. S43. Risk classification: Classified into three risk levels: low, intermediate, and high, based on the probability value of infective endocarditis; S44. Formatting Recommendation: Convert treatment recommendations into SNOMED CT codes and natural language descriptions; S45. Related Cases: Search the database for similar cases to provide historical treatment references.
[0008] Furthermore, a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis. In step S5, the specific steps are as follows; S51. Feature Region Labeling: Grad-CAM is used to generate heatmaps highlighting multifocal hemorrhage points and irregular hematoma edges. In Grad-CAM technology, the weight coefficient of the k-th feature map to category c is... The calculation formula is The heatmap is obtained by weighted summation. ,in The predicted score for category c, Z is the activation value of the k-th feature map at position (i, j), and Z is the number of pixels in the feature map. S52. Three-dimensional reconstruction: Construct a three-dimensional lesion model and set red color coding for intracerebral hemorrhage and blue color coding for edema; S53. Interaction Design: Provide a multi-planar reconstruction interface that supports switching between coronal, sagittal, and axial viewpoints; S54. Comparison Display: Displays the original CT sequence and segmentation results side by side, and adjusts the window width and window level accordingly; S55. Key Note: Use a flashing red warning box to mark hemorrhage foci distributed across blood vessels. The judgment rule is: when the spatial coordinates of two or more intracranial hemorrhage foci are located in different lobes of the brain or different blood supply areas, the system will automatically identify them as hemorrhage foci distributed across blood vessels and mark them with a flashing red warning box. S56. Report Generation: Export a DICOM SR standard PDF report with key annotations.
[0009] A system for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, used to implement the method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in any one of the claims, comprising: a data input module, a lesion analysis module, a fusion decision module, and an output module; The data input module is used to receive cranial CT images transmitted via the DICOM protocol, collect structured clinical data, and perform quality checks. Lesion Analysis Module: Used to call BLAST-CT tools to perform three-dimensional segmentation of lesions, calculate the total volume of multifocal hemorrhage, the ratio of edema zone to intracranial hemorrhage volume, and the hematoma irregularity index; Fusion Decision Module: Used to generate infective endocarditis probability values through a weighted logistic regression model, and output treatment recommendations by matching the threshold of the treatment guidelines. Output module: Used to generate 3D lesion models with Grad-CAM thermal annotations and DICOM SR standard reports.
[0010] The beneficial effects of this invention are: (1) Improved diagnostic efficiency: The present invention uses the BLAST-CT tool to perform fully automatic three-dimensional segmentation of cranial CT images, eliminating the need for manual delineation of lesion boundaries. The image analysis time for a single patient is reduced from several hours of traditional manual image reading to less than 3 minutes. (2) Improved diagnostic accuracy: This invention defines and quantifies imaging parameters with pathological specificity for infective endocarditis, such as the ratio of edema zone to intracerebral hemorrhage volume and hematoma irregularity index. These parameters are then fused with clinical indicators such as blood culture results and vegetation activity scores through a weighted logistic regression model to achieve quantitative identification of the etiology. When the total volume of multifocal hemorrhage is greater than 15 ml and the ratio of edema zone to intracerebral hemorrhage volume is greater than 0.5, the system's sensitivity in automatically identifying high-risk cases of infective endocarditis exceeds 92%, and the misdiagnosis rate is expected to be reduced by about 40% compared to traditional methods. (3) Enhanced standardization of clinical decision-making: This invention dynamically matches the diagnostic output results with the surgical indication thresholds in the "Guidelines for the Diagnosis and Treatment of Infective Endocarditis". When the maximum diameter of the vegetation is greater than 10 mm and the volume of intracerebral hemorrhage is greater than 30 ml, an emergency surgical recommendation is automatically generated. When the volume of intracerebral hemorrhage is between 10 ml and 30 ml, a multidisciplinary consultation is recommended. This mechanism transforms highly subjective clinical decisions into objective and repeatable quantitative judgments, ensuring that treatment recommendations comply with the IDSA 2023 standards. (4) Enhanced visualization and interaction capabilities: This invention uses Grad-CAM technology to generate heat maps, highlighting the image areas most relevant to the diagnosis of infective endocarditis, such as multifocal bleeding points and irregular hematoma edges, and constructs a three-dimensional lesion model to support multi-planar reconstruction interaction. Doctors can adjust the coronal, sagittal and axial views by dragging. This visualization solution helps doctors quickly locate specific features of infective endocarditis, such as bleeding foci distributed across blood vessels, and reduces the risk of missing images. Attached Figure Description
[0011] Figure 1 A flowchart illustrating a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis; Figure 2 This is a schematic diagram of cropping a CT image of a pre-processed lesion. Figure 3 This is a schematic diagram of cropping a 2CT image of a pre-processed lesion. Detailed Implementation
[0012] Example 1: A method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, comprising the following steps; S1. Multimodal data input and verification: Receive cranial CT images via DICOM protocol and standardize slice thickness, simultaneously acquire structured clinical data, and complete preprocessing and quality verification; S2. Lesion segmentation and parameter calculation: The CT images were segmented at the voxel level using a three-dimensional segmentation tool to generate lesion label maps; the total volume of multifocal hemorrhage and the volume ratio of edema zone to intracranial hemorrhage were calculated based on the label maps, and the hematoma irregularity index was calculated. S3. Etiological Probability Fusion and High-Risk Early Warning: The imaging parameters and clinical indicators are probabilistically weighted and fused using a weighted fusion model to generate an infective endocarditis probability value; a first preset weight is assigned to positive blood culture results, and a second preset weight is added when the blood culture contains a specific bacterial species; the vegetation activity score is converted into a third preset weight according to a preset mapping rule; when the total volume of multifocal hemorrhage is greater than the first preset threshold and the volume ratio of the edema zone to the intracerebral hemorrhage is greater than the second preset threshold, a red early warning protocol is automatically activated. S4. Treatment suggestion generation: Dynamically matches treatment guideline thresholds and outputs surgical indications and risk classification suggestions; S5. Visualization: Grad-CAM technology is used to highlight the characteristic areas of infective endocarditis, construct a three-dimensional lesion model, and generate a DICOM SR report.
[0013] Furthermore, in a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, the formula for calculating the ratio of the edema zone to the intracranial hemorrhage volume in step S2 is as follows: ,in This represents the total volume of the edema zone. The ratio represents the total volume of intracerebral hemorrhage. A ratio greater than 0.5 is used as the threshold for indicating infectious hemorrhage. The basis for setting this threshold is as follows: the pathological mechanism of intracerebral hemorrhage caused by infective endocarditis is fundamentally different from that of simple hypertensive intracerebral hemorrhage. The inflammatory infiltration caused by infectious emboli damaging the blood vessel wall leads to a larger area of edema around the hematoma than in simple hematoma. Clinical statistical analysis shows that the median ratio of edema zone to intracerebral hemorrhage volume in patients with intracerebral hemorrhage caused by infective endocarditis is 0.62, while the median ratio in patients with simple hypertensive intracerebral hemorrhage is 0.28. Choosing 0.5 as the threshold can achieve a sensitivity of about 85% while maintaining high specificity, which is the optimal balance point for distinguishing between the two types of hemorrhage etiologies.
[0014] Furthermore, in a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, the formula for calculating the hematoma irregularity index in step S2 is as follows: ,in This represents the surface area of the hematoma. For the surface area of a sphere of the same volume, V represents the hematoma volume; an index greater than 1.8 is used as the threshold for diagnostic specificity of infective endocarditis; the threshold is set based on the following: due to inflammatory erosion and repeated bleeding, the hematoma morphology of infective endocarditis-induced cerebral hemorrhage is often more irregular than that of hypertensive hematoma, with lobulated or flame-shaped edges; statistical analysis of clinical cases shows that the average hematoma irregularity index of patients with infective endocarditis-induced cerebral hemorrhage is 2.1, while the average index of patients with simple cerebral hemorrhage is 1.5; choosing 1.8 as the threshold can trigger a positive marker in about 80% of infective endocarditis cases.
[0015] Furthermore, in a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, the formula for calculating the total volume of multifocal hemorrhage in step S2 is as follows: It only includes the sum of the volumes of the various intracranial hemorrhage lesions, of which Let be the volume of the i-th intracranial hemorrhage lesion. A volume greater than 15 ml is used as the threshold for triggering a high-risk marker of infective endocarditis. The threshold is set based on the following: The typical imaging feature of cerebral hemorrhage caused by infective endocarditis is multifocal hemorrhage, which is caused by the dissemination of infectious emboli to multiple vascular distribution areas in the brain via the bloodstream. Clinical statistics show that the average total volume of multifocal hemorrhage in patients with cerebral hemorrhage caused by infective endocarditis is 21.3 ml, while patients with simple cerebral hemorrhage mostly present with single-focal hemorrhage. Choosing 15 ml as the threshold can capture about 90% of multifocal hemorrhage cases, while avoiding misclassification of small microbleeds as high-risk cases.
[0016] As a preferred embodiment, the weighting mapping formula for the vegetation activity score in step S3 is: The mapping assigns a weight of 0.3 to a vegetation activity score of 0-10, and a weight of 0.8 to a score of 10. This setting is based on the fact that the vegetation activity score comprehensively reflects the size, shape, activity, and attachment site of the vegetation, and is an important indicator for predicting the risk of embolic events. The higher the score, the more unstable the vegetation, and the higher the risk of embolism leading to cerebral hemorrhage. The linear mapping maintains the characteristic that the weight increases monotonically with the score, while controlling the weight range between 0.3 and 0.8 to avoid a single clinical indicator dominating the fusion result.
[0017] As a preferred embodiment, the weighted logistic regression model in step S3 is in the form of: ,in The output is the probability value of infective endocarditis. This represents the total volume of multifocal hemorrhage. This represents the ratio of the edema zone to the volume of intracerebral hemorrhage. For blood culture weight, As a weighting factor for vegetation activity score, For the intercept term, , , , The regression coefficients of each variable are used; the model estimates parameters using the training dataset, and the maximum likelihood estimation method is used to optimize the model parameters during training.
[0018] As a preferred implementation, the logic for threshold judgment in step S3 is as follows: when the total volume of multifocal hemorrhage is greater than 15 ml and the ratio of the edema zone to the intracranial hemorrhage volume is greater than 0.5, the system automatically marks it as a high-risk case of infective endocarditis and activates the red alert protocol. The dual threshold is set based on the following: a single parameter may have insufficient specificity. A total volume of multifocal hemorrhage greater than 15 ml may be seen in some cases of multiple simple intracerebral hemorrhages, and a ratio of the edema zone to the intracranial hemorrhage volume greater than 0.5 may also be seen in some non-infectious hemorrhages with severe inflammatory reactions. When both indicators are met simultaneously, the pathological characteristics of infective endocarditis are more clearly defined, and the diagnostic specificity is significantly improved.
[0019] It should be noted that the "Red Alert Protocol" and the "Probability Value of Infective Endocarditis Output by Weighted Logistic Regression Model" in this system belong to different levels of risk assessment mechanisms. The Red Alert is based on a rapid screening rule of two key imaging parameters: the total volume of multifocal hemorrhage and the edema / hematoma volume ratio. When both exceed the preset thresholds, it indicates that the patient has a very high probability of cerebral hemorrhage due to infective endocarditis, which is used to trigger the emergency clinical intervention process. The Logistic Regression Model integrates more dimensions of data, including blood culture results and vegetation activity scores, and the output probability value provides a more refined quantitative assessment. The two can coexist. This hierarchical design avoids the underreporting of a single rule and improves the clinical applicability of the system.
[0020] Furthermore, in a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, the logic for determining surgical indications in step S4 is as follows: when the maximum diameter of the vegetation detected by ultrasound is greater than 10 mm and the volume of intracranial hemorrhage is greater than 30 ml, an emergency surgical recommendation is generated. The basis for this rule is as follows: a vegetation larger than 10 mm is the threshold recommended in the "Guidelines for the Diagnosis and Treatment of Infective Endocarditis" for considering surgical intervention, as vegetations of this size have a significantly increased risk of embolic events; an intracranial hemorrhage volume greater than 30 ml is a commonly used reference threshold for surgical intervention of intracranial hematomas, as such hematomas usually have a significant mass effect and conservative treatment carries a high risk; when both indicators are met simultaneously, the patient faces both the risk of recurrent embolism and the risk of intracranial hypertension, and emergency surgery provides the greatest benefit; for cases with intracranial hemorrhage volumes between 10 ml and 30 ml, the risk is relatively low, and a comprehensive decision-making process after multidisciplinary consultation is recommended.
[0021] Furthermore, in a method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, the calculation formula for the Grad-CAM technology in step S5 is as follows: ,in The predicted score for category c, Let Z be the activation value of the k-th feature map at position (i,j), and Z be the number of pixels in the feature map; the heatmap is obtained by weighted summation. This technology can highlight imaging features that are discriminative for the diagnosis of infective endocarditis, such as multifocal bleeding points and irregular hematoma edges. The thermal map transparency is set to 40% to ensure that the original image details are visible.
[0022] Example 2: Lesion parameter extraction and key index calculation based on BLAST-CT This embodiment uses cranial CT image data of patient number 1117844 to demonstrate the specific workflow of steps S1 and S2 of the present invention.
[0023] Step S1, multimodal data input and verification, is implemented as follows in this embodiment: S11. Image Data Input: The system receives the raw brain CT image data stream of patient 1117844 via the DICOM network protocol. The patient is a 56-year-old male, clinically suspected of having infective endocarditis. The scan sequence slice thickness is 1mm and the number of scan slices is 32, which meets the system input requirements. S12. Clinical Data Input: Simultaneously construct a structured clinical data interface to collect the patient's blood culture results and vegetation activity score, and use a medical text parsing engine to convert unstructured symptom descriptions in the doctor's observation notes into standardized medical terminology labels; S13. Data preprocessing: Isotropic resampling is performed on CT images to unify the spatial resolution to 1 mm³. The CT values are limited to the range of -15 to 100 Hounsfield Units for intensity normalization. An adaptive filtering algorithm is used to eliminate the interference of high-brightness artifacts such as metal implants. S14. Data Validation: Perform a multi-level quality validation process to verify that the number of image layers is 32, meets the requirement of no less than 20 layers, has no missing clinical fields, and conforms to international storage standards in DICOM format. After validation, output standardized image data and structured clinical data.
[0024] Step S2, lesion segmentation and parameter calculation, is specifically implemented as follows in this embodiment: S21. Lesion Segmentation: The system calls the BLAST-CT tool to perform 3D segmentation on the preprocessed cranial CT images. First, the original CT sequence is divided into 512 3D image blocks, each 64×64×64 voxels. Each image block is fed into a DeepMedic convolutional neural network. This network uses three parallel paths to process images at the original resolution, 3x downsampling, and 5x downsampling respectively. The network classifies and predicts the central voxel of each image block and outputs lesion category labels. After completing the prediction of all image blocks, all prediction results are stitched together according to the spatial position of each image block in the original image to generate the complete lesion segmentation. Figure 3The segmentation results of the voxel-level lesion label map showed that the number of voxels corresponding to intracerebral hemorrhage labels was 388,120, the number of voxels corresponding to extraaxial hemorrhage labels was 433,100, the number of voxels corresponding to perilesional edema labels was 490, and the number of voxels corresponding to intraventricular hemorrhage labels was 32,020. S22. Volume Calculation: Based on the DICOM header information, the pixel pitch of this scan sequence is 0.5mm × 0.5mm, and the layer thickness is 1mm. Therefore, the actual volume of a single voxel is calculated as follows: The volume of intracerebral hemorrhage was calculated by multiplying the number of voxels by the volume of a single voxel. extraaxial bleeding volume Perilesional edema volume Volume of intraventricular hemorrhage ; S23. Parameter Calculation: Total volume of multifocal hemorrhage is = = =97.03ml, the ratio of the edema zone to the intracerebral hemorrhage volume is The system uses a 3D reconstruction algorithm to extract the boundary of intracerebral hemorrhage lesions and calculates their surface area. The surface area of the corresponding sphere of the same volume for this hematoma is Therefore, the hematoma irregularity index is This exceeds the preset threshold of 1.8; S24. Quality verification: The segmentation results are superimposed on the original CT image for visual consistency verification. Isolated lesions with a volume of less than 5 mm³ are filtered out. No such artifacts were detected in this case. The system recorded a segmentation confidence score of 0.94. S25. Data annotation: The system uses the Talairach brain atlas as a reference to mark the spatial location of each lesion in the three-dimensional coordinate system. Intraparenchymal hemorrhage is mainly distributed in the left frontal lobe and right parietal lobe, extraaxial hemorrhage is distributed in the right subdural space, and intraventricular hemorrhage is distributed in the anterior horn of the right ventricle. S26. Result storage: The above lesion parameters are stored in a medical database. Fields include patient ID 1117844, scan time, volume of each lesion type, total volume of multifocal hemorrhage 97.03mL, ratio of edema zone to intracerebral hemorrhage volume 0.00126, hematoma irregularity index 1.81, etc.
[0025] In this case, the total volume of the multifocal hemorrhage was extremely large at 97.03 ml, but the edema ratio of 0.00126 was extremely low, failing to meet the red alert criteria. However, the hematoma irregularity index of 1.81 exceeded the threshold of 1.8, indicating irregular morphological characteristics. The output probability value of the weighted logistic regression model was only 32% after input, which did not support the diagnosis of infective endocarditis and was consistent with the final clinical diagnosis of hypertensive intracerebral hemorrhage.
[0026] Example 3: Multimodal Fusion and Treatment Recommendation Generation This embodiment demonstrates the complete workflow of steps S3, S4, and S5 above using clinical data from patient number 1254448.
[0027] The patient was a 62-year-old male who was admitted to the hospital due to sudden loss of consciousness. The clinical data collected in step S1 were as follows: the blood culture result was positive, the pathogen was viridans streptococcus, which is a Gram-positive bacterium, the vegetation activity score measured by echocardiography was 8 points, and the procalcitonin was 2.3 ng / mL. The following quantitative parameters were obtained after processing in step S2: the volume of intracerebral hemorrhage was 22.50 mL, the volume of extraaxial hemorrhage was 28.83 mL, the volume of edema zone was 1.69 mL, and the volume of intraventricular hemorrhage was 0 mL.
[0028] Step S3, etiology probability fusion, is implemented as follows in this embodiment: S31. Feature Extraction: The system extracts the total volume of multifocal hemorrhage from image parameters. This value is greater than the 15mL threshold, triggering a high-risk volume marker; the ratio of the edema zone to the intracerebral hemorrhage volume was extracted. This value is less than the 0.5 threshold and does not trigger the high-risk marker of the ratio of edema zone to intracerebral hemorrhage volume; S32. Clinical indicator weighting: A positive blood culture result is assigned a base weight of 0.6, and an additional weight of 0.1 is added because the pathogen is a Gram-positive bacterium. Therefore, the weight of blood culture is... The vegetation activity score is 8 points, and the weights are calculated using the linear mapping formula. ; S33. Probability Calculation: The system uses a weighted logistic regression model for fusion, and the regression coefficients of this model are set as follows (example setting). In practical applications, the regression coefficients mentioned above were obtained by optimizing the maximum likelihood estimation method using a training dataset containing 200 pathologically confirmed IE (internal thrombosis) cases and 200 control cases. The linear combination of the model was calculated as follows: =−4.2+0.10×22.50+1.5×0.075+2.0×0.7+1.8×0.7=0.8225; The probability value of infective endocarditis is That is, 69.5%; S34. Threshold judgment: The total volume of multifocal hemorrhage is 22.50 mL, which is greater than 15 mL, but the ratio of the volume of edema zone to the volume of hemorrhage in brain parenchyma is 0.075, which is less than 0.5. It does not meet the condition of being greater than the threshold at the same time. Therefore, the system does not activate the red warning protocol. S35. Result Verification: The stability of the probability calculation was verified by Monte Carlo simulation. The input variables were sampled with 1000 random disturbances. The standard deviation of the calculated probability value was 0.021, and the 95% confidence interval was [0.672, 0.716]. The result was stable. S36. Generate Report: Output a structured diagnostic report containing a probability value of 69.5% for infective endocarditis, and transmit it to the hospital information system via the HL7 interface.
[0029] Step S4, generating treatment recommendations, is implemented as follows in this embodiment: S41. Guideline Matching: The system dynamically compares the size of the vegetation and the volume of intracerebral hemorrhage with the thresholds in the "Guidelines for the Diagnosis and Treatment of Infective Endocarditis". The maximum diameter of the vegetation in this patient was 12 mm and the volume of intracerebral hemorrhage was 22.50 mL as detected by ultrasound. S42. Surgical indications: The vegetation is 12mm, which is greater than the 10mm threshold, but the intracerebral hemorrhage volume is 22.50mL, which is between 10mL and 30mL. The conditions for emergency surgery are not met. According to the system's preset rules, a treatment suggestion of "multidisciplinary consultation" is generated. S43. Risk grading: The risk level is divided according to the probability value of infective endocarditis of 69.5%. A probability greater than 70% is high risk, and 69.5% is borderline medium risk. The system marks these cases as medium-to-high risk and recommends enhanced monitoring. S44. Formatting suggestion: Convert treatment suggestions into SNOMED CT codes, where "multidisciplinary consultation" corresponds to the code "185465009" and "intermediate-high risk monitoring" corresponds to the code "225558002". At the same time, generate a natural language description of the treatment opinion paragraph for doctors to review. S45. Related Cases: The system searches the database for similar confirmed cases. The matching criteria are that the total volume of multifocal hemorrhage is between 20.0 mL and 25.0 mL and the vegetation activity score is between 7 and 9. Three similar cases were found, and their treatment pathways and prognostic data are presented to doctors as a decision-making reference.
[0030] Step S5 is visualized and is implemented as follows in this embodiment: S51. Feature region identification: A heat map is generated using Grad-CAM technology, highlighting multifocal bleeding points and irregular hematoma edges, with the heat map transparency set to 40%; S52. Three-dimensional reconstruction: Construct a three-dimensional lesion model based on the segmentation results. Intracerebral hemorrhage is set to red, edema is set to blue, and intraventricular hemorrhage is set to yellow. S53. Interaction Design: Provides a multi-planar reconstruction interface, allowing doctors to adjust the coronal, sagittal, and axial views by dragging, with a view switching delay of less than 200ms. S54. Comparison Display: The original CT sequence and segmentation annotation results are displayed side by side, with the window width adjusted to 150HU and the window level adjusted to 50HU; S55. Key Marking: Use a flashing red warning box to mark bleeding foci distributed across blood vessels, with the marking size adapting to the size of the lesion; S56. Report Generation: Export a PDF report conforming to DICOM SR standards, including key image annotation pages, parameter summary tables, and treatment recommendation sections.
[0031] This embodiment fully demonstrates the entire process from data input, lesion quantification, multimodal fusion, treatment suggestion generation to visual report output.
[0032] Example 4: Complete Diagnostic and Early Warning Process for Cerebral Hemorrhage Caused by High-Risk Infective Endocarditis This embodiment demonstrates the complete diagnostic, high-risk warning, and treatment recommendation generation process of the present invention in a high-risk case of cerebral hemorrhage caused by infective endocarditis using clinical data of patient number 1982736. The patient was a 68-year-old female who was admitted to the emergency department due to sudden weakness of the right limbs and confusion for 6 hours. Her body temperature was 38.7°C, and she had a history of mitral stenosis due to rheumatic heart disease.
[0033] Step S1, multimodal data input and verification, is implemented as follows in this embodiment: S11. Image Data Input: The system receives the raw brain CT image data stream of patient 1982736 via the DICOM network protocol. The patient's CT scan sequence has a slice thickness of 1mm and a number of 36 slices, which meets the system input requirements. S12. Clinical data input: Clinical data were collected synchronously. The blood culture result was positive and the pathogen was Staphylococcus aureus, which is a Gram-positive bacterium. Transesophageal echocardiography showed a 10mm×6mm vegetation on the anterior leaflet of the mitral valve with high mobility. The vegetation mobility score was 9 out of 10. S13. Data preprocessing: Isotropic resampling is performed on CT images to unify the spatial resolution to 1 mm / voxel, and intensity normalization is performed by limiting the CT values to the range of -15 to 100 HU. S14. Data Validation: Verify that the number of image layers is 36, meeting the requirement of no less than 20 layers, and that the DICOM format conforms to international storage standards. After verification, output standardized image data and structured clinical data.
[0034] Step S2, lesion segmentation and parameter calculation, is specifically implemented as follows in this embodiment: S21. Lesion Segmentation: The system calls the BLAST-CT tool to perform three-dimensional segmentation on the preprocessed cranial CT images, dividing the original CT sequence into 512 three-dimensional image blocks, each with a size of 64×64×64 voxels. The DeepMedic convolutional neural network is used to classify and predict the central voxels of each image block. The segmentation results show that the intraparenchymal hemorrhage label corresponds to 3 independent lesions with voxel counts of 32800, 26000, and 20400, respectively; the total number of voxels corresponding to the perilesional edema label is 84800; the number of voxels corresponding to the intraventricular hemorrhage label is 3200; and the number of voxels corresponding to the extraaxial hemorrhage label is 0. S22. Volume Calculation: Based on the DICOM header information, the pixel pitch of this scan sequence is 0.5mm × 0.5mm, and the layer thickness is 1mm. The actual volume of a single voxel is calculated as follows: =0.5×0.5×1=0.25mm 3 =0.00025ml, the volume of each lesion is calculated by multiplying the number of voxels by the volume of a single voxel: volume of intracerebral hemorrhage lesion 1 =32800×0.00025=8.2ml, volume of 2 intracerebral hemorrhage lesions =26000 × 0.00025 = 6.5 ml, 3 volumes of intracerebral hemorrhage lesion. =20400 × 0.00025 = 5.1 ml, total volume of the edema zone =84800 × 0.00025 = 21.2 ml, intraventricular hemorrhage volume =3200 × 0.00025 = 0.8 ml; S23. Parameter Calculation: The formula for calculating the total volume of multifocal hemorrhage is as follows: This only includes the sum of the volumes of the various intracranial hemorrhage lesions; in this case... =8.2 + 6.5 + 5.1 = 19.8 ml; The formula for calculating the ratio of edema zone volume to intracerebral hemorrhage volume is: The hematoma irregularity index was calculated by selecting the largest hematoma in the left frontal lobe, which had a volume of 8.2 ml. Its surface area was extracted using 3D reconstruction. =2850mm 2 The formula for calculating the surface area of a sphere of equal volume is: Substituting V=8200mm 3 have to hematoma irregularity index ; S24. Quality verification: The segmentation results were superimposed on the original CT images for visual consistency verification. The system recorded a segmentation confidence score of 0.96, and no isolated lesions with a volume of less than 5 mm³ were detected. S25. Data Labeling: The system marks the spatial location of each lesion in the three-dimensional coordinate system with the Talairach brain atlas as a reference. The center coordinates of the lesion in the left frontal lobe are (-38, 15, 22), the center coordinates of the lesion in the left parietal lobe are (-25, -32, 18), and the center coordinates of the lesion in the right occipital lobe are (22, -68, 12). The three lesions are located in different lobes and different blood supply areas, which meets the cross-vascular distribution judgment rules. S26. Result storage: The above lesion parameters are stored in a medical database. Fields include patient ID 1982736, scan time, volume of each lesion type, total volume of multifocal hemorrhage 19.8ml, ratio of edema zone to intracerebral hemorrhage volume 1.07, hematoma irregularity index 1.37, etc.
[0035] Step S3, etiology probability fusion and high-risk early warning, is implemented as follows in this embodiment: S31. Feature Extraction: The system extracts the total volume of multifocal hemorrhage from image parameters. =19.8ml, which is greater than the first preset threshold of 15ml; extract the ratio of edema zone to intracerebral hemorrhage volume. =1.07, which is greater than the second preset threshold of 0.5; S32. Clinical indicator weighting: A positive blood culture result is assigned a base weight of 0.6, and an additional weight of 0.1 is added because the pathogen is a Gram-positive bacterium. Therefore, the weight of blood culture is... =0.7; the vegetation activity score is 9 points, according to the linear mapping formula Calculate the weights. ; S33. Probability Calculation: The system uses a weighted logistic regression model for fusion, and the regression coefficients use the same estimated values as in Example 3, i.e. The linear combination of the model is calculated as follows: =-4.2+0.10×19.8+1.5×1.07+2.0×0.7+1.8×0.75=2.135; the probability value of infective endocarditis is... That is, 89.4%; S34. Threshold judgment: If the total volume of multifocal hemorrhage is greater than 15ml (19.8ml) and the ratio of the volume of edema zone to the volume of hemorrhage in brain parenchyma is greater than 0.5 (1.07), the system will automatically activate the red warning protocol, pop up a high-risk warning window on the operation interface and trigger the emergency treatment process. S35. Result Verification: The stability of the probability calculation was verified by Monte Carlo simulation. The input variables were sampled with 1000 random disturbances. The standard deviation of the calculated probability value was 0.015, and the 95% confidence interval was [0.872, 0.916]. The result was stable. S36. Generate Report: Output a structured diagnostic report containing a probability value of 89.4% for infective endocarditis, and transmit it to the hospital information system via the HL7 interface. The report header is marked with a red high-risk label.
[0036] Step S4, generating treatment recommendations, is implemented as follows in this embodiment: S41. Guideline Matching: The system dynamically compares the size of the vegetation and the volume of intracerebral hemorrhage with the surgical indication thresholds in the guidelines for the diagnosis and treatment of infective endocarditis. The patient's maximum diameter of the vegetation measured by ultrasound was 10 mm, and the total volume of intracerebral hemorrhage was 19.8 ml. S42. Surgical indications: A vegetation of 10mm is equal to the 10mm threshold, but the intracranial hemorrhage volume of 19.8ml is between 10ml and 30ml. According to the system's preset rules, an emergency surgery recommendation is generated when the vegetation is larger than 10mm and the intracranial hemorrhage is larger than 30ml. This case does not meet the conditions for emergency surgery, but the vegetation has reached the recommended size for surgical intervention and the hemorrhage is located in different lobes of the brain. The system generates a recommendation to conduct an emergency surgery assessment and prioritize vegetation resection. At the same time, it suggests that the intracranial hemorrhage has not reached the standard for emergency removal and a neurosurgical consultation can be combined. S43. Risk classification: The risk level is divided according to the probability value of infective endocarditis of 89.4%. A probability greater than 70% is high risk. The system marks the case as a high-risk case and recommends admission to the intensive care unit and initiation of anti-infective treatment. S44. Formatting suggestion: Convert treatment suggestions into SNOMED CT codes, with emergency surgery assessment corresponding to code 183623009 and high-risk monitoring corresponding to code 225558002. At the same time, generate a natural language description of the treatment opinion paragraph for the doctor to review. S45. Related Cases: The system searches the database for similar confirmed cases. The matching criteria are that the total volume of multifocal hemorrhage is between 15ml and 25ml and the vegetation activity score is ≥8. Five similar cases were found, of which four received early surgical treatment and had good prognosis. This information is presented to doctors as a decision-making reference.
[0037] Step S5 is visualized and is implemented as follows in this embodiment: S51. Feature Region Labeling: A heatmap is generated using Grad-CAM technology, highlighting three multifocal hemorrhage points and their surrounding edema zones. The weighting coefficient of the k-th feature map for category c in Grad-CAM technology is calculated using the following formula: The heatmap is obtained by weighted summation. The heatmap transparency is set to 40% and overlaid on the original CT image. S52. 3D Reconstruction: Based on the segmentation results, a 3D lesion model is constructed. Three hemorrhage foci in the brain parenchyma are set to red, the edema zone is set to blue, and intraventricular hemorrhage is set to yellow. Different hemorrhage foci are presented in a semi-transparent rendering mode. S53. Interaction Design: Provides a multi-planar reconstruction interface, allowing doctors to adjust the coronal, sagittal, and axial views by dragging the mouse, with a view switching delay of less than 200ms. S54. Comparison Display: The original CT sequence and segmentation annotation results are displayed side by side. The window width is adjusted to 150HU and the window level is adjusted to 50HU. The linkage adjustment ensures the consistency of the comparison. S55. Key point: The system automatically identifies hemorrhage foci distributed across blood vessels. The judgment rule is that when the spatial coordinates of two or more intracranial hemorrhage foci are located in different lobes or different blood supply areas, the system will automatically identify them and mark them with a flashing red warning box. In this case, the three lesions in the left frontal lobe, left parietal lobe and right occipital lobe are located in different lobes and different blood supply areas, and are marked with a flashing red warning box with the text "distributed across blood vessels" displayed next to the label. S56. Report Generation: Export a PDF report conforming to the DICOM SR standard, including key image annotation pages with Grad-CAM heatmaps and 3D reconstruction maps, a parameter summary table with total multifocal hemorrhage volume of 19.8 ml, edema-to-hematoma volume ratio of 1.07, hematoma irregularity index of 1.37, probability value of infective endocarditis of 89.4%, as well as red warning records and treatment recommendations. The first page of the report is printed with a red high-risk watermark.
[0038] This embodiment fully demonstrates the application of the present invention in the entire process of high-risk infective endocarditis-induced cerebral hemorrhage cases: when the total volume of multifocal hemorrhage is 19.8 ml and the ratio of edema to hematoma volume is 1.07, a red alert is triggered. The weighted logistic regression outputs an infective endocarditis probability value of 89.4%. The system automatically generates emergency surgery assessment suggestions and intuitively displays the characteristics of multifocal hemorrhage distributed across blood vessels through Grad-CAM heatmap and three-dimensional model.
Claims
1. A method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, characterized in that, Includes the following steps; S1. Multimodal data input and verification: Receive cranial CT images via DICOM protocol and standardize slice thickness, simultaneously acquire structured clinical data, and complete preprocessing and quality verification; S2. Lesion segmentation and parameter calculation: The CT images were segmented at the voxel level using a three-dimensional segmentation tool to generate lesion label maps; the total volume of multifocal hemorrhage and the volume ratio of edema zone to intracranial hemorrhage were calculated based on the label maps, and the hematoma irregularity index was calculated. S3. Etiological probability fusion and high-risk warning: The imaging parameters and clinical indicators are probabilistically weighted and fused using a weighted fusion model to generate the probability value of infective endocarditis; a first preset weight is assigned to positive blood culture results, and a second preset weight is added when the blood culture contains a specific bacterial species; The vegetation activity score is converted into a third preset weight according to a preset mapping rule; When the total volume of multifocal hemorrhage exceeds the first preset threshold and the ratio of the volume of edema zone to the volume of intracerebral hemorrhage exceeds the second preset threshold, the red warning protocol is automatically activated. S4. Treatment suggestion generation: Dynamically matches treatment guideline thresholds and outputs surgical indications and risk classification suggestions; S5. Visualization: Grad-CAM technology is used to highlight the characteristic areas of infective endocarditis, construct a three-dimensional lesion model, and generate a DICOM SR report.
2. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 1, characterized in that, In step S1, the specific steps are as follows: S11. Image data input: Receive cranial CT images via DICOM protocol, and normalize the scan sequence to 1mm slice thickness using inter-slice interpolation algorithm; S12. Clinical Data Input: Construct a structured interface to collect blood culture results and vegetation activity scores; S13. Data preprocessing: Perform isotropic resampling on CT images, limit CT values to the range of -15 to 100 HU and normalize the intensity; S14. Data verification: Verify that the number of image layers is not less than 20, verify the compliance of DICOM format, and trigger a resampling request for data with a median CT value exceeding 200 HU.
3. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 1, characterized in that, In step S2, the specific steps are as follows: S21. Lesion segmentation: The BLAST-CT tool is called to perform voxel-level classification and prediction of image blocks through a three-dimensional convolutional neural network, generating a voxel-level lesion label map. S22. Volume Calculation: Based on the number of voxels in the label image, the volume in milliliters is calculated by combining the spatial resolution. The volume is equal to the number of voxels multiplied by the volume of a single voxel. The volume of a single voxel is calculated based on the pixel spacing and layer thickness in the DICOM header file. S23. Parameter Calculation: Calculate the ratio of the volume of the edema zone to the volume of the hemorrhage in the brain parenchyma and the hematoma irregularity index. The ratio of the volume of the edema zone to the volume of the hemorrhage in the brain parenchyma is the ratio of the total volume of the edema zone to the total volume of the hemorrhage in the brain parenchyma. The hematoma irregularity index is the ratio of the surface area of the hematoma to the surface area of a sphere of the same volume. S24. Quality verification: Overlay the original images to verify the segmentation consistency and filter out isolated lesions with a volume of less than 5 mm³; S25. Data annotation: Mark the spatial location of the lesion and its maximum diameter and volume parameters in a three-dimensional coordinate system; S26. Result storage: Store lesion parameters in a medical database and support SQL queries.
4. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 1, characterized in that, In step S3, the specific steps are as follows: S31. Feature extraction: Extract the total volume of multifocal hemorrhage and the ratio of the edema zone to the volume of hemorrhage in the brain parenchyma, where the total volume of multifocal hemorrhage is the sum of the volumes of each hemorrhage lesion in the brain parenchyma; S32. Clinical indicator weighting: assign a first preset weight to positive blood culture results, and add a second preset weight when the blood culture result is a specific bacterial species; The vegetation activity score is converted into a third preset weight according to a preset linear mapping formula; S33. Probability Calculation: The probability value of infective endocarditis is output by fusing feature variables using a logistic regression model; S34. Threshold judgment: When the total volume of multifocal hemorrhage is greater than the first preset threshold and the ratio of the edema zone to the volume of hemorrhage in the brain parenchyma is greater than the second preset threshold, a red warning is activated. S35. Result Verification: The stability of the probability calculation was verified by Monte Carlo simulation; S36. Generate Report: Output a structured diagnostic report containing the probability value of infective endocarditis and transmit it via the HL7 interface.
5. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 4, characterized in that, The first preset weight is a value within the range of 0.5 to 0.7; The second preset weight is a value within the range of 0.05 to 0.15; The third preset weight is linearly mapped through the formula w=a+(s / 10)×b, where s is the vegetation activity score of 0~10, a is 0.2~0.4, and b is 0.4~0.6; The first preset threshold is a value within the range of 10~20ml; The second preset threshold is a value within the range of 0.4 to 0.
6.
6. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 1, characterized in that, In step S4, the specific steps are as follows: S41. Guideline matching: Dynamically compare the size of vegetations and the volume of intracerebral hemorrhage with the surgical indication thresholds; S42. Surgical indications: When the maximum diameter of the vegetation is greater than 10 mm and the volume of intracranial hemorrhage is greater than 30 ml, an emergency surgical recommendation is generated. S43. Risk classification: Classified into three risk levels: low, intermediate, and high, based on the probability value of infective endocarditis; S44. Formatting Recommendation: Convert treatment recommendations into SNOMED CT codes and natural language descriptions; S45. Related Cases: Search the database for similar cases to provide historical treatment references.
7. The method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in claim 1, characterized in that, In step S5, the specific steps are as follows: S51. Feature region identification: Grad-CAM is used to generate a heat map highlighting multifocal bleeding points and irregular hematoma edges; S52. Three-dimensional reconstruction: Construct a three-dimensional lesion model and set red color coding for intracerebral hemorrhage and blue color coding for edema; S53. Interaction Design: Provide a multi-planar reconstruction interface that supports switching between coronal, sagittal, and axial viewpoints; S54. Comparison Display: Displays the original CT sequence and segmentation results side by side, and adjusts the window width and window level accordingly; S55. Key Point: Use a flashing red warning box to mark bleeding foci distributed across blood vessels; S56. Report Generation: Export a DICOM SR standard PDF report with key annotations.
8. A system for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis, used to implement the method for intelligent diagnosis of cerebral hemorrhage caused by infective endocarditis as described in any one of claims 1-7, characterized in that, include: Data input module, lesion analysis module, fusion decision module, output module; The data input module is used to receive cranial CT images transmitted via the DICOM protocol, collect structured clinical data, and perform quality checks. Lesion Analysis Module: Used to call BLAST-CT tools to perform three-dimensional segmentation of lesions, calculate the total volume of multifocal hemorrhage, the ratio of edema zone to intracranial hemorrhage volume, and the hematoma irregularity index; Fusion Decision Module: Used to generate infective endocarditis probability values through a weighted logistic regression model, and output treatment recommendations by matching the threshold of the treatment guidelines. Output module: Used to generate 3D lesion models with Grad-CAM thermal annotations and DICOM SR standard reports.