Thyroid nodule home intervention system and method based on AR navigation and DRL

Abstract

The invention discloses a thyroid nodule home intervention system and method based on AR navigation and DRL, and relates to the technical field of thyroid nodule home management and intervention.The thyroid nodule home intervention system based on AR navigation and DRL comprises a client module arranged on a user mobile terminal and used for carrying out thyroid nodule home intervention on a thyroid nodule; the guiding module is used for guiding a user to perform self-operation in a home environment, collecting data and receiving feedback; the cloud service module is in communication connection with the client module through the Internet and is used for remote data processing, model operation and personalized intervention scheme management; according to the thyroid nodule home management system, an innovative system of thyroid nodule home management is constructed through cooperation of multiple modules, that is, the maximum acupuncture point positioning error is 3 mm through an AR navigation unit of a client by means of facial feature point fitting and a neck statistical shape model, and the problem that acupuncture points are not found accurately in traditional home operation is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of home management and intervention of thyroid nodules, in particular to a thyroid nodule home intervention system and method based on AR navigation and DRL. BACKGROUND

[0002] Thyroid nodule is one of the most common endocrine diseases. In recent years, with the popularization of high-resolution ultrasound technology and the enhancement of health examination awareness, the detection rate of thyroid nodules has shown a significant upward trend. At present, the standard management of thyroid nodules at home and abroad mainly follows the "Guidelines for Diagnosis and Treatment of Thyroid Nodules and Differentiated Thyroid Carcinoma". For benign nodules, the clinical standard practice is to perform thyroid ultrasound review every 6-12 months, and by comparing the changes in size, shape, boundary and other characteristics of the nodules, the nature and development trend of the nodules are evaluated. This traditional management mode of waiting and observation has obvious defects: first, during the long follow-up period, the patient is in a completely passive state, lacking any active intervention means, which is easy to produce anxiety and panic emotions; second, frequent hospital review brings time cost and economic burden to the patient; finally, ultrasound examination can only provide static morphological information and cannot reflect the dynamic changes of the nodules in the development process.

[0003] In the prior art field, although there are some health management mobile application software that attempts to provide thyroid health related services, the functions are mostly single, mainly focusing on the following aspects: health knowledge popularization and information push, medication reminder and review schedule management, simple symptom recording and tracking. These applications cannot solve the core problem of thyroid nodule management, i.e. lack of evidence-based medical support, quantifiable home intervention means.

[0004] A few studies attempt to apply traditional therapies such as Chinese massage to the adjuvant therapy of thyroid nodules, but there are obvious limitations in actual operation: on the one hand, patients have difficulty in accurately finding neck-related acupoints such as Renying, Tiantu, Futu and foot reflex zones, resulting in large randomness of operation; on the other hand, the intensity, frequency and duration of massage cannot be quantified and standardized, and the effect is difficult to guarantee and evaluate. More importantly, these operations lack objective data recording, and doctors cannot obtain any process data when inquiring, and can only rely on the patient's subjective description and ultrasound image results for judgment.

[0005] In terms of technology, augmented reality technology has been applied in the fields of medical training and anatomy education, but its application in the field of chronic disease home management is still blank. Similarly, the application of artificial intelligence technology in the generation of personalized intervention programs is still in the preliminary exploration stage, and there is no report of combining AR technology and intelligent hardware for thyroid nodule management.

[0006] Therefore, the AR navigation and DRL based thyroid nodule home intervention system and method are proposed to solve the problems in the background. SUMMARY

[0007] The AR navigation and DRL based thyroid nodule home intervention system and method are provided to solve the problem of the application of AR technology in the field of chronic disease home management in the market.

[0008] To achieve the above object, the application provides the following technical scheme. An AR navigation and DRL based thyroid nodule home intervention system comprises: A client module is arranged on a user mobile terminal and is used for guiding the user to perform self-operation, collect data and receive feedback in a home environment. A cloud service module is in communication connection with the client module through the Internet and is used for remote data processing, model operation and individualized intervention scheme management; supports an HL7FHIR standard open interface and is used for integration with a hospital information system and an ultrasonic equipment system; the model is updated incrementally every day for 24 hours, and the client supports edge reasoning and cache update content; A doctor terminal module is in communication connection with the cloud service module and is used for providing a remote monitoring and management interface for medical staff; The client module comprises: An AR navigation unit, which is specifically configured to call a depth perception component of the user mobile terminal, fit based on real-time acquired facial key points and a preset neck statistical shape model, calculate and render and display a virtual mark of a target acupoint on the user's body surface in real time, and guide the user's self-operation; A data collection and interaction unit, which is specifically configured to establish a communication connection with an external intelligent pressure sensing device, receive and display pressure time series data from the device in real time when the user operates, and provide a health care task execution interface; An AI question and answer unit, which is specifically configured to support voice input interaction based on a thyroid field medical knowledge graph and a fine-tuned large model, automatically transfer human customer service for sensitive questions, and provide real-time health consultation; A daily task and punch-in unit, which is specifically configured to generate an individualized massage task list, record the task completion status and support punch-in, issue points based on the completion rate, and use the points to exchange ultrasonic examination coupons.

[0009] The cloud service module comprises: The DRL recommendation engine unit is specifically: a state space is constructed by using user attributes, nodule characteristics, symptom scores and historical operation compliance data, an action space is constructed by using executable massage acupoint combinations, operation force and operation frequency, a reward function is constructed by using predicted nodule volume changes and quality of life improvement, and a user home-executed personalized thyroid health intervention scheme is generated through a trained strategy network.

[0010] Specifically, the state space includes a seven-day massage completion rate of the user; in the reward function, the quality of life improvement weight lambda is dynamically adjusted according to the TI-RADS classification of the user.

[0011] A thyroid nodule home intervention method based on AR navigation and DRL, comprising the following steps: Step 1: remotely receiving user registration information and baseline medical data, analyzing and storing user baseline archives; Step 2: generating an initial personalized thyroid health intervention scheme suitable for home execution based on the initial state of the user through a DRL recommendation model, and issuing the scheme to the user client; Step 3: in the user's home environment, guiding the user to accurately position the acupoints specified in the scheme through the AR navigation function of the client; Step 4: in the process of user self-operation, collecting real-time biomechanical data through external equipment, and processing and extracting features of the data to generate real-time guidance feedback; Step 5: recording the behavior data of each self-operation of the user and uploading it to the cloud server; Step 6: periodically aggregating user data on the cloud server, updating the user state, and optimizing the DRL recommendation model using incremental learning to generate an optimized home intervention scheme for the next period; Step 7: at a preset follow-up node, remotely reminding the user to review, and generating a therapeutic effect analysis report according to the review results obtained remotely and the baseline data.

[0012] Compared with the prior art, the beneficial effects of the present application are: The present application builds an innovative system for home management of thyroid nodules through multi-module cooperation, that is, through the AR navigation unit of the client, the acupoint positioning error is ≤3mm by using facial feature point fitting and neck statistical shape model, solving the problem of inaccurate acupoint positioning in traditional home operation. And the intelligent pressure sensor and data processing unit through filtering, feature extraction and two-classification model, the operation amount such as pressing force and frequency is quantized, so that the effective pressing judgment accuracy is greatly improved, and the experience operation is converted into standardized process.

[0013] The application constructs a complete hospital diagnosis, home intervention and remote supervision closed-loop management system, so that patients can still receive standardized and effective health management after leaving the hospital, changes the traditional mode of only periodic observation of thyroid nodules, and fills the gap in clinical management.

[0014] The application converts traditional empirical operation into standardized and quantifiable digital therapy, adopts a deep reinforcement learning algorithm that can dynamically adjust the intervention scheme according to real-time data of each user, and performs model incremental update every 24 hours, so as to ensure that the recommended scheme always adapts to the latest state of the user.

[0015] The above summary is only for the purpose of the description and is not intended to limit in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the application will be readily apparent to those skilled in the art by reference to the drawings and the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of the method for home intervention of thyroid nodules based on AR navigation and DRL of the application; Figure 2 A module block diagram of the system for home intervention of thyroid nodules based on AR navigation and DRL of the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] Embodiment one

[0019] As shown in the accompanying drawings, Figure 2 A system for home intervention of thyroid nodules based on AR navigation and DRL, comprising: 1. A client module arranged in a user mobile terminal, used for guiding the user to perform self-operation, collect data and receive feedback in a home environment; In the user mobile terminal, the user uploads a thyroid ultrasound report through an App, the system analyzes the report through an OCR and NLP engine, extracts the position, size and TI-RADS classification of the nodule, automatically calculates the volume of the nodule, the user completes a QoL-THY life quality scale, and generates an initial symptom score.

[0020] For the collection of face and neck data, the TrueDepth camera is started to capture RGB-D images, the MediaPipeFaceMesh model is called to extract 468 facial 3D key points, the coordinate system is aligned and rigidly transformed with the nose tip (Landmark1) and the chin center (Landmark152) as stable points, and a pre-trained neck statistical shape model is loaded to generate a personalized 3D neck mesh.

[0021] 2. The cloud service module is in communication connection with the client module through the Internet, and is used for remote data processing, model operation and personalized intervention scheme management; supports the HL7 FHIR standard open interface, and is used for integration with a hospital information system and an ultrasonic equipment system; the model is updated incrementally every day for 24 hours, and the client supports edge reasoning and cache update content; 3. The doctor end module is in communication connection with the cloud service module, and is used for providing a remote monitoring and management interface for medical staff; 4. The external intelligent pressure sensing device collects pressure time series data F_raw(t) at a sampling rate of 50 Hz, and transmits the data to the client through Bluetooth 5.2; a 10 Hz second-order Butterworth low-pass filter is used to smooth F_raw(t) to filter out high-frequency noise, and a filtered signal F_filtered(t) is obtained.

[0022] The sliding window size is 3s, containing 150 data points, and the root mean square value and the peak factor are calculated; The input feature vector [RMS, CF] is input to the pre-trained random forest model, and the training set contains 12,478 labeled samples, of which 8,923 are effective pressing samples; The model contains 500 decision trees, and the depth of a single tree is at least 15 layers, and outputs the "effective pressing" prediction probability P. When P is greater than or equal to 0.85 and the condition is met for 5 consecutive windows, it is determined to be an effective operation. The model AUC=0.92, and the accuracy is 89.7%.

[0023] First, the PPO-Clip algorithm is used to train the DRL for 5000 epochs, and each epoch contains 1000 interaction steps; The strategy network is a 3-layer Transformer encoder with a hidden layer dimension of 256 and 8 attention heads per layer. The value network and the strategy network share the first 2 layers of weights, and the 3rd layer outputs the state value estimate separately. The experience replay pool capacity is 1 million, the batch size is 256 samples per sampling, the clipping parameter ε is 0.2, and the learning rate initial value is 5e-5 which decays with training.

[0024] The client module includes: 1.1, AR navigation unit, extract multiple facial feature points through MediaPipe library, align coordinate system and rigid transformation based on stable feature points of nose tip and chin, load predefined neck statistical shape model SSM, fit to generate personalized neck three-dimensional grid model; According to the pre-stored mapping relationship of acupoint anatomical position, the three-dimensional space coordinates of target acupoints including the carotid acupoint, the tiantu acupoint, the lianquan acupoint, the futu acupoint, the zusanli acupoint and the sanyinjiao acupoint are calculated on the three-dimensional grid model.

[0025] The positioning error of three-dimensional space coordinates is at most 3mm, and the voice broadcast acupoint avoidance area is supported.

[0026] Specifically, the depth perception component of the user's mobile terminal is called, fitting is performed based on the real-time acquired facial key points and the pre-set neck statistical shape model, the virtual identification of the target acupoint is calculated and rendered and displayed on the user's body surface in real time, and the user's self-operation is guided; multiple facial feature points are extracted through the MediaPipe library, coordinate system alignment and rigid transformation are performed based on stable feature points of nose tip and chin, a predefined neck statistical shape model SSM is loaded, and a personalized neck three-dimensional grid model is fitted and generated. In the acupoint positioning algorithm, the input is an RGB-D frame and facial key points, and the process is: MediaPipeFaceMesh, coordinate system alignment, SSM fitting, and output of 3D coordinates of 6 major acupoints.

[0027] The error is controlled to be at most 3mm, and the voice prompt avoids the dangerous area; real-time rendering: using ARKit / ARCore to superimpose virtual identification on the user's body surface.

[0028] 1.2, data acquisition and interaction unit, low-pass filtering is performed on the received original pressure data stream; time domain features including root mean square RMS and peak factor are extracted from the filtered data; the time domain features are input into a pre-trained binary classification machine learning model to determine whether the current operation is an effective pressing, and the determination result is fed back to the user interface in real time.

[0029] Specifically, a communication connection is established with an external intelligent pressure sensing device, pressure time series data from the device during user operation is received and displayed in real time; a health care task execution interface is also provided; 1.3, AI question and answer unit, specifically: based on the thyroid field medical knowledge graph and the fine-tuning large model, voice input interaction is supported, sensitive questions are automatically transferred to artificial customer service, and real-time health consultation is provided; 1.4, daily task and clock-in unit, specifically: generating a personalized massage task list, recording the task completion status and supporting clock-in, issuing points based on the completion rate, and using points to exchange ultrasonic examination coupons.

[0030] The cloud service module comprises: 2.1, DRL recommendation engine unit, the algorithm adopted is proximal policy optimization (PPO) algorithm, the policy network and the value network thereof adopt transformer encoder architecture. The transformer encoder is a 3-layer structure, the policy network and the value network share weights, and the clipping parameter ε=0.2 when the PPO-Clip algorithm is used.

[0031] Specifically: the state space is constructed by user attributes, nodule characteristics, symptom scores and historical operation compliance data, the action space is constructed by executable massage acupoint combinations, operation force and operation frequency, the reward function is constructed by predicted nodule volume change and quality of life improvement, and the personalized thyroid health intervention scheme executed by the user at home is generated through the trained policy network.

[0032] Among them, the state space includes the user's seven-day massage completion rate; in the reward function, the quality of life improvement weight λ is dynamically adjusted according to the user's TI-RADS classification.

[0033] Embodiment two

[0034] 1. When the user starts the AR navigation, the phone camera captures a frame of 1080p RGB image and its corresponding depth map. The frame image is input into the lightweight MediaPipeFaceMesh model, and the model outputs the normalized 468 3D feature point coordinates of the face within 15ms. Take the nose tip (landmark1) and the chin center (landmark152) as stable reference points, and calculate a 4x4 transformation matrix through singular value decomposition algorithm to convert the feature points in the model coordinate system to the camera world coordinate system. The system pre-stores a statistical shape model based on 368 neck CT data, uses the aligned mandibular contour feature points (landmarks334-353) as input, and fits the user's personalized neck 3D grid through principal component analysis model. In the SSM model, the positions of each acupoint have been predefined as specific vertices on the grid. The system directly obtains the 3D world coordinates [x, y, z] = [0.123, -0.045, 0.387] of these vertices. Finally, use the SceneKit rendering engine of ARKit to draw a 3D glowing sphere virtual marker with a diameter of 8mm at this coordinate position, realizing accurate positioning with an error of <2.5mm.

[0035] Specifically, the system's test on 368 volunteers showed that the average positioning error of the Renying and Tiantu points was 2.1 mm. For users with a neck fat thickness of more than 2 cm, the error can be controlled within 2.5 mm by using the SSM model to additionally compensate for the deformation of the subcutaneous tissue. The voice prompt module will detect the carotid artery position in real time and trigger the broadcast of "move up 0.5 cm to avoid the carotid artery" when the user is about to operate.

[0036] 2. The raw pressure signal F_raw(t) from the external intelligent pressure sensing device is first filtered through a second-order Butterworth low-pass filter with a cutoff frequency of 10 Hz to obtain the smoothed signal F_filtered(t). Within a 3-second wide sliding window (150 data points), the root mean square value (RMS = 3.8 N) and the peak factor (CrestFactor = 2.1) of F_filtered(t) are calculated in real time. The feature vector [3.8, 2.1] is input into a pre-trained random forest classifier, which is trained on 12,478 labeled samples containing 500 decision trees. The classifier outputs the prediction probability P(effective) = 0.93. Since P > 0.85 and five consecutive windows are judged to be "effective compression", the system gives continuous green waveform feedback on the UI. When a window is judged to be invalid due to a peak factor CrestFactor = 4.8 > 3.0 (P(effective) = 0.21), the system immediately triggers the linear vibration motor in the device to produce a 100 ms vibration pulse, achieving millisecond-level tactile alarm.

[0037] Specifically, the policy network of the DRL model uses a 3-layer Transformer encoder with a hidden layer dimension of 256 and 8 attention heads per layer. The value network shares the first two layers of weights with the policy network and separately outputs the value estimate in the third layer. During training, the model is initialized with 1 million virtual interaction data, fine-tuned with 2,147 clinical data, and incrementally updated every 24 hours using the user's operation data from that day. The client locally caches the model parameters for the past 7 days to reduce latency.

[0038] More specifically, in the input layer of the acupoint positioning algorithm, the RGB-D frame is captured by the mobile TrueDepth camera; and the 468 facial key points are output synchronously.

[0039] In the processing layer, first, taking the tip of the nose (landmark1) and the center of the chin (landmark152) as the reference points, a 3x3 rotation matrix and a 3x1 translation vector are calculated through singular value decomposition to construct a conversion matrix from the model coordinate system to the camera world coordinate system, realizing the rigid alignment of the face and neck coordinate systems.

[0040] Then, the SSM based on 368 neck CT data, which contains 100 principal components with an explanation rate of ≥95%, is loaded, and the aligned mandibular contour landmark points (landmarks 334-353) are input to fit the user's personalized neck 3D mesh model through least squares iterative optimization.

[0041] Finally, according to the pre-stored acupoint-mesh vertex mapping relationship, the world coordinates of the target acupoints are directly extracted from the fitted 3D mesh, and are converted into screen 2D coordinates through the ARKit rendering engine to superimpose a green virtual circle.

[0042] In the output layer, the 3D spatial coordinates of the 6 major acupoints and the foot thyroid reflex area are output, with a maximum positioning error of 3 mm.

[0043] Example Three As shown in the accompanying Figure 1 The user's process for home intervention of the nodule is as follows: User Zhang, a 38-year-old woman, was found to have a TI-RADS 3 nodule in the right lobe of the thyroid gland during a physical examination, with a size of 6.2 x 4.8 x 5.0 mm. The specific process of using the AR navigation and DRL-based thyroid nodule home intervention system in the home environment is as follows: 1. Zhang downloaded and installed the system APP through the mobile phone application store at home, completed mobile phone registration and identity verification, and uploaded her latest thyroid ultrasound report by taking a photo with her mobile phone according to the system instructions. The system's OCR and NLP analysis engine automatically extracted the key information from the report: nodule location: right lobe, size: 6.2 x 4.8 x 5.0 mm, TI-RADS classification: 3. According to the ellipsoid formula, the initial volume of the nodule was automatically calculated . Subsequently, the thyroid disease-specific quality of life scale assessment built into the system was completed, with a score of 52. All data were uploaded to the cloud after encryption to form her exclusive digital health record.

[0044] 2. The cloud deep reinforcement learning recommendation engine received Zhang's initial state vector s0: {age: 38, BMI: 22.1, TI-RADS: 3, nodule volume: 0.078, QoL score: 52, historical compliance rate: 0}. The engine input s0 into the trained PPO policy network with a 3-layer Transformer encoder structure and a hidden layer dimension of 256. After network inference, the action vector a0 was output, which was decoded to generate her first 14-day personalized health care plan: ① perform 1 time per day, in the evening; ② acupoints and operations: bilateral Renying, each for 2.5 minutes, force 3.8 N, clockwise rubbing and pressing, Tianzhu for 1.5 minutes, force 2.8 N, gentle tapping, bilateral foot thyroid reflex area, each for 3.5 minutes, pushing from bottom to top. The scheme is delivered to her mobile APP end through HTTPS protocol.

[0045] 3. Every night before sleep, open the APP and enter "Today's Task", click "Start AR Navigation". The rear TrueDepth camera of the phone is started, the AR navigation unit detects 468 feature points of the face in real time through the MediaPipeFaceMesh model, calculates the coordinates of the neck acupoints, and displays a virtual green circle on the screen to accurately cover the Renyong and Tiantu points on the neck, with a voice prompt "Please massage in this area". Wear the smart finger pressing device, which is connected to the APP through Bluetooth 5.2. During the operation, the device collects pressure data F(t) at a sampling rate of 50Hz and transmits it to the APP in real time. The APP data processing unit performs 10Hz low-pass filtering on F(t) and calculates the real-time root mean square value. When the pressure is maintained in the range of 3.5-4.2N, the screen displays a green waveform; when the pressing force reaches 5.8N at a certain moment, a red spike appears on the waveform, and the device produces a short vibration to remind that the force is too large.

[0046] 4. After each task is completed, the complete data of the operation is encrypted and uploaded to the cloud. The cloud engine aggregates all users' desensitized data every morning and starts the incremental learning process to fine-tune and optimize the DRL model using new data. After 14 consecutive days of clocking in, 100 points are obtained, and the system automatically pushes the "50 yuan ultrasound coupon" to nearby hospitals, which can be redeemed by clicking on the link to the hospital reservation system.

[0047] 5. After using for 3 months, the system reminds Zhang through the APP and SMS to make an appointment for review. After the review at the local hospital, the new report is uploaded to the system. The system analysis shows that the nodule volume is reduced to 0.062 cm 3 , the volume change rate AV = -20.5%. The QoL score improves to 38. The cloud engine receives this positive feedback and stores it as a high reward value sample in the experience replay pool. The doctor sees the automatically generated follow-up report on the doctor's system, which shows that the nodule volume has significantly decreased and the quality of life has improved significantly, and remotely approves the system's automatically generated next stage optimization plan.

[0048] Specifically, in the operation, the AI question and answer robot replies within 1.2s based on the knowledge graph when asked by voice "Does the nodule enlargement need immediate surgery?" "TI-RADS class 3 nodule increases <20% in the short term and has no malignant characteristics, it is recommended to follow up and observe first", and automatically associates its historical ultrasound data; when she asks "Can I continue to massage during pregnancy", the system identifies it as a sensitive question and jumps to online consultation with a human thyroid specialist.

[0049] Example Four

[0050] In a clinical trial in a certain third-grade hospital, a total of 126 patients with TI-RADS 3 nodules used the system for 3-month home intervention management, and the results showed that: (1) Nodule changes: After intervention, the average volume of the patient's nodule decreased by (15.8±6.3)%, of which the volume of 78.6% (99 / 126) of the patients decreased by more than 10%, the volume of 18.3% (23 / 126) of the patients was stable (within ±10%), and the volume of 3.2% (4 / 126) of the patients increased by more than 10%.

[0051] (2) Quality of life improvement: The QoL-THY score improved by an average of (21.7±8.4) points (from a baseline of 52.3±9.1 points to 30.6±7.2 points).

[0052] (3) Compliance and satisfaction: The average operation compliance rate was 85.7%, and the user satisfaction score was 4.6 / 5.0. 92.1% of the patients indicated that their psychological anxiety was significantly reduced.

[0053] (4) Clinical efficiency improvement: The doctor's follow-up efficiency was improved by about 45%, and the average follow-up time for each patient was shortened from 15.2 minutes to 8.4 minutes.

[0054] The above examples show that the system of the present application can effectively realize the home intervention management of thyroid nodule patients, and has good effects in promoting nodule shrinkage and improving the quality of life, while improving the utilization efficiency of medical resources.

[0055] Example Five

[0056] All data communication between the client and the cloud server is encrypted and transmitted using the TLS1.3 protocol, the data payload is encrypted using the AES-256-GCM algorithm, and two-way certificate authentication is enabled to prevent man-in-the-middle attacks.

[0057] User static data is stored in the cloud database using the AES-256 algorithm, and the encryption key is managed uniformly by a special key management service, realizing the separate storage of data and keys.

[0058] The data set used for model training is protected using differential privacy technology, adding Laplace noise, and the privacy budget is set to ε=0.9, ensuring that no sensitive information about any single user can be inferred from the training data.

[0059] The system implements role-based access control, and all data access operations are recorded in an unalterable audit log, and regular security audits and vulnerability scans are performed.

[0060] The embodiments of the present application are given for illustration and description only, although embodiments of the present application have been shown and described herein, it should be understood by the person skilled in the art that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and the person skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application, which should be included in the protection scope of the present application.

Claims

1. A thyroid nodule home intervention system based on AR navigation and DRL, characterized by: include: The client module is installed on the user's mobile terminal and is used to guide the user to perform self-operation, collect data and receive feedback in a home environment; The cloud service module connects to the client module via the Internet to perform remote data processing, model calculations, and personalized intervention plan management; it supports the HL7 FHIR standard open interface for integration with hospital information systems and ultrasound equipment systems; The doctor-side module communicates with the cloud service module to provide a remote monitoring and management interface for medical staff; The client module includes: The AR navigation unit specifically calls the depth perception component of the user's mobile terminal, fits the real-time acquired facial key points with the preset neck statistical shape model, calculates and renders the virtual markers of the target acupuncture points on the user's body surface in real time, and locates and guides the user's self-operation; The data acquisition and interaction unit is specifically configured to establish a communication connection with an external intelligent pressure sensing device, receive and display in real time the pressure time series data from the device during user operation, and provide a health care task execution interface; The AI ​​question-and-answer unit is specifically designed to: Based on a medical knowledge graph and fine-tuned large-scale model in the thyroid field, it supports voice input interaction, automatically transfers sensitive questions to human customer service, and provides real-time health consultation; The daily task and check-in unit generates a personalized massage task list, records task completion status, supports check-in, and issues points based on completion rate. Points can be redeemed for ultrasound examination coupons. Cloud service modules include: The DRL recommendation engine unit specifically constructs a state space based on user attributes, nodule characteristics, symptom scores, and historical operation compliance data; an action space based on executable massage point combinations, operation intensity, and operation frequency; a reward function based on predicted nodule volume changes and improved quality of life; and a trained policy network to generate a personalized thyroid health intervention plan for users to implement at home. Specifically, the state space includes the user's seven-day massage completion rate; in the reward function, the quality of life improvement weight λ is dynamically adjusted according to the user's TI-RADS grade.

2. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 1 is characterized in that: The algorithm used by the DRL recommendation engine unit is the proximal policy optimization (PPO) algorithm, and its policy network and value network both use the Transformer encoder architecture. The Transformer encoder has a three-layer structure. The policy network and the value network share weights. The clipping parameter ε is 0.2 when the PPO-Clip algorithm is used.

3. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 1 is characterized in that: The cloud service module also includes a data security unit, which specifically: encrypts stored user static data using the AES-256 algorithm, and adds noise to the data set used for model training using differential privacy technology with a privacy budget of ε=1.

0.

4. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 1 is characterized in that: The AR navigation unit extracts multiple facial feature points through the MediaPipe library, performs coordinate system alignment and rigid transformation based on the stable feature points of the nose tip and chin, loads the predefined statistical shape model (SSM) of the neck, and fits and generates a personalized 3D mesh model of the neck. According to the pre-stored mapping relationship of the acupoints' anatomical positions, the three-dimensional spatial coordinates of the target acupoints of Renying, Tiantu, Lianquan, Futu, Zusanli and Sanyinjiao are calculated on the three-dimensional grid model; The positioning error of the three-dimensional space coordinates is ≤3mm, and the acupoint avoidance area is announced by voice.

5. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 1 is characterized in that: The data collection and interaction unit specifically includes: Perform low-pass filtering on the received raw pressure data stream; Extract the time domain features of root mean square (RMS) and crest factor from the filtered data; The time domain features are input into a pre-trained two-classification machine learning model to determine whether the current operation is a valid press, and the judgment result is fed back to the user interface in real time.

6. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 5 is characterized in that: The binary classification machine learning model is a random forest model, the sampling frequency of the received pressure data is 50 Hz, and the cutoff frequency of the low-pass filter is 10 Hz.

7. The thyroid nodule home intervention system based on AR navigation and DRL according to claim 1 is characterized in that: Also includes: The external intelligent pressure sensing device is equipped with a pressure sensor and an inertial measurement unit (IMU) inside, and communicates with the client module via the Bluetooth protocol to collect and upload pressure data during user operation.

8. A method for home intervention of thyroid nodules based on AR navigation and DRL, applied to the home intervention system for thyroid nodules based on AR navigation and DRL as claimed in any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: Remotely receive user registration information and baseline medical data, parse and store user baseline files; Step 2: Based on the user's initial status, the DRL recommendation model generates an initial personalized thyroid health intervention plan suitable for home implementation and sends the plan to the user client; Step 3: In the user's home environment, the client's AR navigation function guides the user to accurately locate the acupuncture points specified in the plan; Step 4: During the user's self-operation, real-time biomechanical data is collected through external equipment, and the data is processed and feature extracted. The effective pressure range is preset to [3N, 5N] as the pressure judgment standard for generating real-time guidance feedback; Step 5: Record the user's behavioral data for each self-operation and upload it to the cloud server; Step 6: The cloud server periodically aggregates user data, updates user status, and uses incremental learning to optimize the DRL recommendation model to generate an optimized home intervention plan for the next cycle; Step 7: At the preset follow-up node, remotely remind the user to review, and generate an efficacy analysis report based on the remotely obtained review results and baseline data.

9. The method for home intervention of thyroid nodules based on AR navigation and DRL according to claim 8, characterized in that: In step 4, real-time guidance feedback is generated, specifically: Compare the processed pressure data with the preset effective pressure range [3N, 5N]; When the pressure data remains within the effective range for more than a preset time threshold, positive visual feedback is provided; when the pressure data exceeds the effective range, a tactile alarm device is triggered to provide a reminder.

Images