Extracorporeal shock wave treatment parameter generation method and system based on image semantics

Abstract

The invention discloses an extracorporeal shock wave treatment parameter generation method and system based on image semanteme, and the method comprises the following steps: S1, inputting an ultrasonic and / or X-ray image, and carrying out the preprocessing of the image; s2, image features are extracted, and focus information is analyzed; a lightweight medical image model, a multi-modal model and a clinical reasoning big language model are adopted, image features are converted into structured image features, and structured focus information is reasoned; s3, in-vitro shock wave treatment parameters are generated, and real-time compliance verification is carried out; and S4, visually displaying image labels, treatment parameters and compliance prompts. Therefore, through the collaborative architecture of the lightweight medical image model and the clinical reasoning large language model, the unstructured image is converted into the structured focus feature, the treatment parameter is generated, the treatment accuracy is improved, the double-image information can be fused, and the treatment parameter accuracy is further improved.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to a method and system for generating extracorporeal shock wave therapy parameters based on image semantics. Background Technology

[0002] Extracorporeal shock wave therapy (ESWB) devices are commonly used in orthopedics and rehabilitation departments. They apply pulsed pressure waves to the affected area and can be used to treat conditions such as Achilles tendinitis, lateral epicondylitis, and osteoarthritis. However, their effectiveness is highly dependent on the precise setting of parameters such as the nozzle angle, impact depth, and intensity. Existing technologies suffer from the following core problems: Parameter settings rely on physicians' subjective experience: Existing extracorporeal shock wave therapy devices require physicians to visually observe medical images (ultrasound, X-ray) to determine lesion information, and then manually input parameters based on clinical experience. For example, for calcifications at the Achilles tendon insertion point, physicians need to estimate the depth of the calcification from the skin based on experience, and then set the impact depth. Subjective errors can easily lead to parameter deviations, affecting the therapeutic effect or increasing the risk of adverse reactions. Imaging information is not fully converted into quantitative parameters: Ultrasound can clearly show soft tissue information such as tendon edema and fine calcifications, and X-ray can accurately present bone structure information such as osteophytes and obvious calcifications. However, existing technologies only use images as "visual references" and do not convert the structural features in the images (such as "calcification diameter 3mm" and "osteophyte length 5mm") into quantitative treatment parameters, so the value of imaging is not fully realized. Insufficient dual-image modality coordination: In clinical practice, patients often acquire ultrasound (to view soft tissue) and X-ray (to view bone structure) images simultaneously. However, existing equipment cannot coordinate the two types of image information to generate parameters. Physicians need to interpret the two images separately and then make a comprehensive judgment, which is cumbersome and prone to missing key information. Lack of parameter visualization and compliance verification: After physicians manually set parameters, they can only see isolated values ​​on the screen, which cannot be intuitively associated with the location of the lesion (such as whether the angle of the nozzle is aligned with the lesion). Furthermore, there is no real-time compliance verification mechanism, which may lead to setting parameters that exceed clinical norms due to operational errors. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for generating extracorporeal shock wave therapy parameters based on image semantics, so as to solve one or more of the above-mentioned technical problems.

[0004] To achieve this objective, the present invention adopts the following technical solution: A method for generating extracorporeal shock wave therapy parameters based on image semantics includes the following steps: S1: Input an image and preprocess it, wherein the image is an ultrasound image, or an X-ray image, or the image includes both ultrasound and X-ray images. S2: Extract image features and analyze lesion information; S21: Employs a lightweight medical imaging model to identify key lesion features in images and output structured image features; S22: Employ a multimodal model to transform structured image features into natural language descriptions; S23: Employ a clinical reasoning big language model, combined with an imaging feature and lesion type mapping database, to classify and grade lesions and output structured lesion information; S3: Employs a large clinical reasoning language model, combined with a clinical rule base of lesions and parameters, to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. S4: Visualized display of image annotations, treatment parameters, and compliance prompts.

[0005] In some implementations, in step S1, Gaussian filtering is used to remove artifacts from the ultrasound image, and a threshold segmentation algorithm is used to select and mark the lesion area of ​​the image.

[0006] In some implementations, in step S1, the X-ray image is contrast-processed using standardized grayscale, and an edge detection algorithm is used to identify the distance between the lesion and surrounding tissues or the distance between surrounding tissues.

[0007] In some implementations, in step S21, the key features of the lesion include the location, size, and shape of the lesion; the structured imaging features include the nature of the lesion, its size, and the distance between the lesion and surrounding tissues.

[0008] In some implementations, in step S23, the structured lesion information includes anatomical location, lesion nature, size, and distance between the lesion and surrounding tissues.

[0009] In some implementations, in step S3, the treatment parameters include the nozzle angle, impact depth, impact intensity, impact frequency, and number of pulses. In step S3, if the treatment parameters exceed the safety threshold, a regeneration mechanism is triggered.

[0010] In some implementations, when ultrasound images and X-ray images are input simultaneously in step S1, step S2 extracts the features of the ultrasound images and X-ray images respectively, and outputs the fused structured lesion information; in step S4, the image annotations and treatment parameter association table of the ultrasound images and X-ray images are displayed simultaneously.

[0011] An image semantics-based extracorporeal shock wave therapy parameter generation system for running the above method, the system comprising: The image input and preprocessing module is used to input images and perform image preprocessing. The image semantics and lesion information parsing module is used to extract image features and parse lesion information; The treatment parameter generation and verification module is used to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. The visualization module is used to display image annotations, treatment parameters, and compliance prompts.

[0012] The beneficial effects of this invention are: 1. To achieve automated analysis of lesion information in medical images (ultrasound, X-ray), transforming unstructured images into structured lesion features (location, size, morphology). 2. Establish precise mapping rules between "lesion characteristics and treatment parameters". Based on clinical guidelines and real case data, automatically generate parameters such as nozzle angle, impact depth, and intensity through a large language model. 3. Supports collaborative analysis of dual image modalities of ultrasound and X-ray, automatically adapts analysis logic according to image type, and integrates dual image information to improve parameter accuracy; 4. Enables visualization and real-time compliance verification of parameters, linking image annotations, parameter values, and compliance prompts on the screen to assist physicians in confirmation and prevent parameters from exceeding safety boundaries. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of a method for generating extracorporeal shock wave therapy parameters based on image semantics according to the present invention. Figure 2 This is one of the detailed flowcharts of an extracorporeal shock wave therapy parameter generation method based on image semantics according to the present invention; Figure 3 This is the second detailed flowchart of a method for generating extracorporeal shock wave therapy parameters based on image semantics according to the present invention. Figure 4 This is a comparison image of the ultrasound images before and after preprocessing according to the present invention. Figure 5 This is a structural diagram of the visual display interface of the present invention; Figure 6 This is a structural diagram of an extracorporeal shock wave therapy parameter generation system based on image semantics according to the present invention; The module consists of: 1-Image input and preprocessing module; 2-Image semantics and lesion information analysis module; 3-Treatment parameter generation and verification module; 4-Visualization module. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings.

[0015] refer to Figures 1 to 5A method for generating extracorporeal shock wave therapy parameters based on image semantics includes the following steps: S1: Input image and preprocess the image; The image can be an ultrasound image, an X-ray image, or a combination of ultrasound and X-ray images; that is, an ultrasound image or an X-ray image can be entered alone, or both can be entered simultaneously.

[0016] Among them, reference Figure 5 Preprocessing of ultrasound images: Gaussian filtering is used to remove artifacts from ultrasound images. Gaussian filtering is a linear smoothing filtering algorithm that uses a Gaussian function (such as a normal distribution function) as weights. In ultrasound images, Gaussian filtering can be used to remove artifacts such as random noise and speckle noise, thereby improving the accuracy of image recognition.

[0017] Threshold segmentation algorithm is used to select the lesion area in the marked image and segment or divide the image. Threshold segmentation algorithm is an image segmentation method that is based on the difference in gray value between the target object (such as lesion) and the background in the image. By setting one or more gray value thresholds, the image pixels are divided into two or more categories, thereby achieving region separation and improving the accuracy of recognition and the efficiency of subsequent lesion analysis.

[0018] X-ray image preprocessing: Standardized grayscale is used to perform contrast processing on X-ray images, thereby enhancing contrast and making the target tissue structure more clearly visible, thus improving the accuracy of image recognition. Edge detection algorithms are used to identify the distance between lesions and surrounding tissues or the distance between surrounding tissues. Edge detection algorithms can identify areas in an image where pixel values ​​change drastically, such as the boundary between a lesion and surrounding normal tissue. By calculating the gradient or derivative of the image, discontinuities in grayscale values, colors, or textures can be found. When the gradient value of a pixel exceeds a preset threshold, it is marked as an edge point, providing a basis for subsequent lesion analysis and helping to improve the accuracy and efficiency of subsequent lesion analysis.

[0019] S2: Extract image features and analyze lesion information; for example, use a lightweight medical imaging model and a large language model to frame the key features of the lesion, extract structured image features, convert them into natural language descriptions, and then infer structured lesion information.

[0020] Step S2 specifically includes steps S21 to S23; S21: Employs a lightweight medical imaging model to identify key lesion features in images and output structured image features; The lightweight medical imaging model can use the YOLOv8-seg version, supporting detection and segmentation. The model can be fine-tuned using a database to enable it to automatically map image features to corresponding lesion types.

[0021] Key features of the lesion include its location, size, and shape; Structured imaging features include the nature and size of the lesion, as well as its distance from surrounding tissues. For example: “Calcification at the Achilles tendon insertion point, 3mm in diameter, 4mm deep from the skin surface”; "Osteophytic spur on the lateral epicondyle of the humerus, measuring 5mm × 2mm, located at the tendon attachment point."

[0022] Therefore, key features of lesions are identified from ultrasound and / or X-ray images, and structured image features are output.

[0023] The training data sources for the lightweight medical imaging model YOLOv8-seg include: public datasets (such as DeepLesion, MURA, SIIM-ACR Pneumothorax) and desensitized clinical case data provided by partner hospitals; the total data volume is approximately 9,000 records or more.

[0024] S22: Employ a multimodal model to transform structured image features into natural language descriptions; Among them, the multimodal model can adopt the Qwen-VL version; The structured image features extracted by the lightweight medical imaging model YOLOv8-seg are transformed into natural language descriptions for use in subsequent large-scale model inference.

[0025] For example, the multimodal model Qwen-VL structured image features (such as "diameter:3.5mm", "depth:4.2mm") can also receive image vector features (image patch embedding) from the lightweight medical imaging model YOLOv8-seg, and then translate them into doctor-readable language: "A hyperechoic calcification lesion with a diameter of about 3.5mm is visible at the Achilles tendon insertion point, about 4.2mm from the skin surface, with no acoustic shadow behind it, suggesting old calcification." Therefore, the structured image features are transformed into natural language descriptions, which are convenient for doctors to read and for large language models to perform reasoning.

[0026] S23: Employ a clinical reasoning big language model, combined with an imaging feature and lesion type mapping database, to classify and grade lesions and output structured lesion information; The clinical reasoning language model uses the Qwen3-32b version (a slightly modified version for the medical field).

[0027] Structured lesion information includes anatomical location, lesion nature, size, and distance between the lesion and surrounding tissues.

[0028] The clinical reasoning big language model uses Qwen3-32b to receive natural language descriptions and image types (ultrasound images, X-ray images) from the multimodal model Qwen-VL. Then, it combines the preset "image feature-lesion type" mapping library to classify and grade lesions. Example mapping rules: “Hypoechoic area on ultrasound + tendon thickening → tendinitis”; “High-density nodule on X-ray + tendon course area → calcific tendinitis”.

[0029] Therefore, by using reasoning and mapping databases of image features and lesion types, the classification and grading of lesions can be accurately and quickly determined.

[0030] The "Image Feature-Lesion Type" mapping library includes: ultrasound image feature-lesion type mapping library and X-ray image feature-lesion type mapping library. Different images correspond to different mapping libraries. The mapping library is constructed by forming a structured database through image parameters, case data, classification standards, grading standards, etc.

[0031] Therefore, step S2 uses a lightweight medical imaging model and a large language model to collaboratively frame the key features of lesions, extract structured image features, convert them into natural language descriptions, and then infer structured lesion information to improve the accuracy and efficiency of lesion identification.

[0032] S3: Employs a large clinical reasoning language model, combined with a clinical rule base of lesions and parameters, to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. The clinical reasoning language model uses the Qwen3-32b version (a slightly modified version for the medical field).

[0033] Treatment parameters include nozzle angle, impact depth, impact intensity, impact frequency, and pulse count; If the treatment parameters exceed the safety threshold, a regeneration mechanism is triggered, such as returning to step 2 or issuing an alarm message.

[0034] Therefore, the reasoning big language model uses Qwen3-32b based on structured lesion information and combines it with the "lesion-parameter" clinical rule base to generate extracorporeal shock wave therapy parameters, improve accuracy, and perform real-time compliance verification to improve safety.

[0035] The clinical rule base for lesions and parameters was constructed based on professional medical literature and papers (including but not limited to): "Clinical Application Guidelines for Extracorporeal Shock Wave Therapy of Orthopedic Diseases (2024 Edition)" and "Expert Consensus on Shock Wave Therapy in Rehabilitation Medicine", etc., including data from 1,000+ clinically effective treatment cases. The "lesion-parameter" mapping rules were summarized from actual treatment cases.

[0036] S4: Visualized display of image annotations, treatment parameters, and compliance prompts; For example, visualization can be achieved through a screen, displaying annotations of related images, treatment parameters, and compliance prompts to assist physicians in confirmation and prevent treatment parameters from exceeding safety limits.

[0037] Furthermore, when ultrasound images and X-ray images are input simultaneously in step S1, step S2 extracts the features of the ultrasound images and X-ray images respectively, and outputs the fused structured lesion information; in step S4, the image annotations and treatment parameter association table of the ultrasound images and X-ray images are displayed simultaneously.

[0038] Therefore, this method supports the collaborative analysis of dual-image modalities of ultrasound and X-ray images, automatically adapts the analysis logic according to the image type, and integrates dual-image information to improve the accuracy of treatment parameters and other values.

[0039] refer to Figure 6 An image semantic-based extracorporeal shock wave therapy parameter generation system for running the above method, the system comprising: Image input and preprocessing module 1 is used to input images and preprocess them; Image semantics and lesion information parsing module 2 is used to extract image features and parse lesion information; The treatment parameter generation and verification module 3 is used to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. Visualization module 4 is used to display image annotations, treatment parameters, and compliance prompts, for example, through a screen.

[0040] Beneficial effects: 1. To achieve automated analysis of lesion information in medical images (ultrasound, X-ray), transforming unstructured images into structured lesion features (location, size, morphology). 2. Establish precise mapping rules between "lesion characteristics and treatment parameters". Based on clinical guidelines and real case data, automatically generate parameters such as nozzle angle, impact depth, and intensity through a large language model. 3. Supports collaborative analysis of dual image modalities of ultrasound and X-ray, automatically adapts analysis logic according to image type, and integrates dual image information to improve parameter accuracy; 4. Enables visualization and real-time compliance verification of parameters, linking image annotations, parameter values, and compliance prompts on the screen to assist physicians in confirmation and prevent parameters from exceeding safety boundaries.

[0041] Case 1: Ultrasound image analysis; 1. Image Access and Preprocessing: Physicians acquire ultrasound images of the patient's Achilles tendon using an ultrasound machine. The system imports the images, removes ultrasound artifacts using Gaussian filtering, and uses a threshold segmentation algorithm to select and mark the area of ​​strong echo calcification at the Achilles tendon insertion point. 2. Lesion Information Analysis: The lightweight medical imaging model extracts image features: "Calcification at the Achilles tendon insertion point, diameter 3.5mm, depth from skin surface 4.2mm"; the multimodal model converts structured image features into natural language descriptions; the clinical reasoning big language model combines the "ultrasound-Achilles tendinitis" mapping rules to output the lesion type: "calcific Achilles tendinitis (moderate)". 3. Treatment parameter generation and verification: The clinical reasoning big language model calls the "calcification-parameter" rule base to generate the following parameters: gun tip angle 45° (consistent with the long axis of the Achilles tendon), impact depth 4mm, intensity 1.9 bar, frequency 18Hz, and number of treatments 2000. Combined with the patient's age (45 years old, moderate tolerance), the treatment parameters are verified to be within the safe boundary and marked as "compliant". 4. Visualization and Interaction: Calcifications are marked with red boxes in the screen image area, and dotted arrows indicate the 45° gun tip angle; the parameter table area displays the above parameters; after the physician confirms that there are no objections to the parameters, he / she clicks "Confirm Execution", and the system transmits the parameters to the extracorporeal shock wave main control system.

[0042] Case 2: X-ray image analysis; 1. Image Access and Preprocessing: The X-ray machine outputs X-ray images of the patient's knee joint. After importing the images into the system, the contrast is processed using standardized grayscale. The knee joint space and osteophytes are identified through edge detection algorithms. 2. Lesion Information Analysis: The medical imaging model extracts features: "Knee joint space 2.8mm (narrowing), medial joint margin osteophyte 4mm×2mm"; the multimodal model converts the structured image features into natural language descriptions; the clinical reasoning big language model determines the lesion type: "Knee osteoarthritis (Grade II)"; 3. Treatment parameter generation and verification: The clinical reasoning big language model calls the rule base to generate parameters: gun head angle 60° (along the joint space), impact depth 3.5mm, intensity 1.6bar, frequency 15Hz; combined with the patient's weight (70kg) for verification, the intensity does not exceed the range of 1.0-1.8bar, and is marked "compliant"; 4. Visualization and Interaction: The location of osteophytes and joint spaces are marked with red boxes in the screen imaging area, and the parameter table area displays the values; if the physician thinks the intensity is slightly low, he can drag the slider to adjust the intensity to 1.7 bar. The system prompts "1.7 bar is in line with the recommended range (1.0-1.8 bar) for grade II osteoarthritis". The physician confirms and then executes the adjustment.

[0043] Case 3: Synergistic analysis of ultrasound and X-ray images; 1. Image Access and Preprocessing: The system simultaneously imports the patient's rotator cuff ultrasound image and X-ray image. After preprocessing, the ultrasound image is identified as "hypoechoic area in tendon (edema, depth 2mm)" and the X-ray image is identified as "calcification lesion in tendon with a diameter of 5.5mm". 2. Lesion information analysis: The lightweight medical imaging model extracts the features of the two images respectively; the multimodal model converts the structured image features of the two into natural language descriptions respectively; the clinical reasoning big language model integrates the information of the two to make reasoning judgments: "rotator cuff calcification tendinitis, calcification diameter 5.5mm, accompanied by tendon edema (depth 2mm)". 3. Treatment Parameter Generation and Verification: The clinical reasoning big language model combined with dual image features generates the following parameters: gun head angle 75° (refer to the location of calcifications on X-ray), impact depth 4mm (refer to the distance from ultrasound edema to the skin), and intensity 2.3 bar (calcifications > 5mm); during verification, since the intensity of 2.3 bar is within the "recommended range (2.2-2.5 bar) for calcifications of 5-8mm", it is marked as "compliant". 4. Visualization and Interaction: Dual images are displayed in a split screen, with edema and calcifications marked separately. The parameter table area is associated with the descriptions of the dual images. After the physician confirms, the system executes the command and simultaneously records the source of the dual images and the basis for parameter generation.

[0044] 1. Precise mapping technology of image semantics-parameters: The first collaborative architecture of "lightweight medical image model + clinical reasoning big language model" is used to transform unstructured ultrasound / X-ray images into structured image features, and then generate quantitative treatment parameters based on clinical rule base. The mapping error rate is less than 5%, breaking through the limitation of "images are only used as visual reference". 2. Dual-image modality collaborative analysis technology: It can identify the image type (ultrasound / X-ray) and adapt the corresponding analysis logic. If two images are input at the same time, it will fuse the image features of the two to generate parameters (such as ultrasound to see the depth of edema and X-ray to see the size of calcifications), which solves the problem of incomplete information from a single image. 3. Parameter visualization and compliance interaction technology: The screen displays the associated information of "image annotation - parameter value - compliance prompt", which supports physicians to fine-tune and verify parameter compliance in real time, thus preserving the physician's decision-making authority and avoiding subjective errors.

[0045] The above description only discloses some embodiments of the present invention. For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the invention.

Claims

1. A method for generating extracorporeal shock wave therapy parameters based on image semantics, comprising the following steps: S1: Input image and preprocess the image, where... The image is an ultrasound image, or the image is an X-ray image, or the image includes both ultrasound and X-ray images. S2: Extract image features and analyze lesion information; S21: Employs a lightweight medical imaging model to identify key lesion features in images and output structured image features; S22: Employ a multimodal model to transform structured image features into natural language descriptions; S23: Employ a clinical reasoning big language model, combined with an imaging feature and lesion type mapping database, to classify and grade lesions and output structured lesion information; S3: Employs a large clinical reasoning language model, combined with a clinical rule base of lesions and parameters, to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. S4: Visualized display of image annotations, treatment parameters, and compliance prompts.

2. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, In step S1, Gaussian filtering is used to remove artifacts from the ultrasonic image; A threshold segmentation algorithm was used to select the lesion area in the labeled image.

3. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, In step S1, the X-ray image is contrast-processed using standardized grayscale. Edge detection algorithms are used to identify the distance between the lesion and surrounding tissues or the distance between surrounding tissues.

4. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, In step S21, the key features of the lesion include its location, size, and shape; Structured imaging features include the nature of the lesion, its size, and its distance from surrounding tissues.

5. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, In step S23, the structured lesion information includes anatomical location, lesion nature, size, and distance between the lesion and surrounding tissues.

6. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, In step S3, the treatment parameters include the nozzle angle, impact depth, impact intensity, impact frequency, and number of pulses. In step S3, if the treatment parameters exceed the safety threshold, a regeneration mechanism is triggered.

7. The method for generating extracorporeal shock wave therapy parameters based on image semantics according to claim 1, characterized in that, When ultrasound images and X-ray images are input simultaneously in step S1, step S2 extracts the features of ultrasound images and X-ray images respectively, and outputs the fused structured lesion information; in step S4, the image annotations and treatment parameter association table of ultrasound images and X-ray images are displayed simultaneously.

8. An image semantic-based extracorporeal shock wave therapy parameter generation system for running the method of claim 1, the system comprising: The image input and preprocessing module is used to input images and perform image preprocessing. The image semantics and lesion information parsing module is used to extract image features and parse lesion information; The treatment parameter generation and verification module is used to generate extracorporeal shock wave therapy parameters and perform real-time compliance verification. The visualization module is used to display image annotations, treatment parameters, and compliance prompts.

Images