A subcutaneous injection probe insulin pen, smart annotation system and method

Abstract

This invention provides a subcutaneous insulin pen, an intelligent annotation system, and a method. It improves upon traditional insulin pens by installing a subcutaneous detection module at the pen tip. This module uses non-contact detection technologies, such as ultrasound, to detect abnormalities in subcutaneous tissue and nodules. The module then uploads the analysis results to a backend server via a signal analysis module. Nurses, patients, and family members can directly view skin data via a mobile app connected to the insulin pen before injection to assess whether the skin at that location is suitable for injection. This enables precise in-situ detection, automated assessment, safety management, intelligent interpretation, and full-cycle health management before insulin injection, fundamentally addressing the safety concerns of home insulin injection and overcoming all the shortcomings of existing technologies.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical device manufacturing technology, and in particular to a subcutaneous insulin probe pen, an intelligent annotation system and its application control method, electronic equipment and computer-readable storage medium. Background Technology

[0002] Insulin pens (also known as NovoPen) have become the mainstream tool for home insulin injection due to their ease of use, precise dosage, and portability, accounting for more than 90% of clinical insulin injection devices. However, long-term subcutaneous insulin injection carries clear clinical risks: repeated injections at the same site can easily lead to subcutaneous fat hyperplasia, induration, and nodule formation, with an incidence rate as high as 60%-78%. If the injection site is not chosen properly, injecting insulin into the muscle layer, blood vessels, or nodule areas can cause abnormal insulin absorption rates, drastic fluctuations in blood sugar, and even complications such as severe hypoglycemia, infection, and subcutaneous hematoma, seriously affecting the efficacy of insulin therapy and the patient's quality of life.

[0003] Traditional insulin injection site selection relies entirely on the patient's visual observation, tactile sensation, and experience. It cannot effectively identify tiny nodules with a diameter of <5mm, abnormal fat layer thickness, or subcutaneous microvascular / nerve distribution, resulting in a high rate of mis-injection and is a core cause of insulin injection-related complications.

[0004] Therefore, current technologies for assisting insulin injection all have unavoidable technical shortcomings, failing to meet the needs for safe, convenient, and accurate injection management in home settings. Specific deficiencies are as follows: 1. Traditional insulin pens lack in-situ detection capabilities and safety control mechanisms: Commercially available insulin pens only have dose adjustment and injection execution functions, but no subcutaneous tissue detection capabilities. They cannot objectively assess the suitability of the injection site, rely entirely on the user's subjective judgment, and lack hardware-level safety protection mechanisms, thus failing to eliminate the risk of accidental injection at the source.

[0005] 2. Separate detection devices suffer from site offset and cannot achieve detection-injection synchronization: Existing injection auxiliary devices with subcutaneous detection function all adopt a design that separates the ultrasound detector from the injection device. After detection is completed, the injection device is moved to inject. This cannot guarantee the consistency between the detection site and the injection site, and site offset is very likely to occur, causing the detection results to lose their guiding significance. At the same time, the device is large in size and complex to operate, requiring professional personnel to operate, and cannot be adapted to home use scenarios.

[0006] 3. Poor adaptability of contact-based detection technology and lack of dedicated scenario-specific algorithms: Existing medical ultrasound detection is all based on a contact design, requiring the application of coupling gel to achieve effective imaging. This results in poor hygiene and cumbersome operation for home use, leading to low user acceptance. At the same time, general ultrasound imaging algorithms are designed for deep organs and have insufficient resolution for superficial subcutaneous tissue of 1-10mm. They cannot identify tiny nodules or small blood vessels and can only achieve imaging functions. They cannot automatically complete the quantitative assessment of the suitability of injection sites, requiring interpretation by a professional ultrasound physician and thus cannot achieve automated home applications.

[0007] 4. Lack of standardized intelligent annotation and multi-terminal collaborative management system: Existing technologies have not built an intelligent annotation system for insulin injection scenarios, making it impossible to automatically and easily interpret the detection results. Ordinary patients cannot understand the clinical significance of ultrasound images. At the same time, there is a lack of a multi-terminal collaborative mechanism among patients, medical staff, and family members, which makes it impossible to achieve remote monitoring, professional guidance, and long-term injection management, and a complete injection safety closed loop has not been formed. Summary of the Invention

[0008] To address the technical problems existing in the prior art, the present invention provides the following technical solution: On one hand, a subcutaneous insulin detection pen is provided, including a pen tip, wherein a subcutaneous detection module is disposed on the pen tip, the subcutaneous detection module being used for: Signal detection is performed on subcutaneous tissue / nodule abnormalities using non-contact detection technology. The analyzed skin condition data is then uploaded to a backend server to assess whether the skin at that location is suitable for injection.

[0009] On the other hand, an intelligent annotation system is provided, characterized by comprising the aforementioned subcutaneous insulin probe pen, cloud-based intelligent analysis and annotation engine, and multi-terminal collaborative intelligent annotation terminal, wherein: The integrated subcutaneous injection probe insulin pen has a built-in insulin pen basic execution unit, a high-frequency non-contact ultrasound subcutaneous detection module, an embedded edge computing and preprocessing unit, and a pen-end wireless communication module. It is used to realize non-contact in-situ detection of subcutaneous tissue, execution of insulin injection, edge preprocessing of echo signals, and injection safety interlock control. The needle adapter interface of the high-frequency non-contact ultrasound subcutaneous detection module and the insulin pen basic execution unit are designed coaxially. The cloud-based intelligent analysis and annotation engine adopts a distributed cloud-native architecture and communicates with multi-terminal collaborative intelligent annotation terminals. It has built-in subcutaneous tissue ultrasound image segmentation module, multi-dimensional feature extraction module, injection site suitability intelligent assessment module, and multi-terminal intelligent annotation generation module to complete intelligent analysis of ultrasound images, quantitative assessment of injection site suitability, intelligent annotation information generation, and multi-terminal data collaboration. The multi-terminal collaborative intelligent annotation terminal includes a patient-side intelligent terminal, a medical staff-side management terminal, and a family-side monitoring terminal, which are used to realize user-system interaction, visualization of detection results, multi-terminal remote collaborative monitoring and health management.

[0010] On the other hand, an application control method for the aforementioned intelligent annotation system is provided, characterized by comprising the following steps: System initialization and user pairing: Complete the full hardware self-test of the insulin pen, and the user completes account registration, real-name authentication and device pairing and binding through the patient-side smart annotation App, establishes a personal electronic health record and stores it in the cloud; Pre-injection preparation and detection activation: After the user completes the installation of the insulin pen cartridge, needle, and dosage adjustment, and aligns the insulin pen tip with the intended injection site on the skin, when the infrared distance sensor detects that the distance between the pen tip and the skin is within the effective detection range of 0.5mm-5mm, an operation reminder is issued. After the user triggers the detection command, the high-frequency non-contact ultrasound subcutaneous detection module is activated to complete the acquisition of subcutaneous tissue echo signals. Echo signal preprocessing and data transmission: The FPGA at the pen end performs parallel delay summation beamforming, Hilbert transform envelope detection, logarithmic compression, and digital filtering preprocessing on the acquired echo signal to generate standardized subcutaneous tissue ultrasound images. After compression and encryption, the images are transmitted to the patient's app via Bluetooth link and then uploaded to the cloud-based intelligent analysis and annotation engine. Cloud-based intelligent analysis and injection site suitability assessment: The cloud-based system completes pixel-level segmentation of subcutaneous tissue using an improved U-Net segmentation network. Based on the segmentation results, five core assessment features are extracted. The system then uses the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method to complete a three-level graded assessment of injection site suitability, generating assessment results and corresponding safety interlock control commands. Intelligent annotation generation and multi-terminal synchronization: Based on the evaluation results, feature values ​​and segmentation results, the cloud generates standardized text annotations and visual annotation images. At the same time, it generates injection site recommendation schemes through personalized recommendation algorithms and synchronizes relevant information to the patient, medical staff and family members through WebSocket and MQTT protocols. Safety interlock control and injection permission management: The patient-side App encrypts and transmits the safety interlock control commands sent from the cloud to the insulin pen MCU. The MCU drives the safety interlock module to perform graded unlocking actions according to the commands: the appropriate injection level unlocks the injection permission directly, the cautious injection level requires the user to manually confirm the risk before unlocking, and the inappropriate injection level remains permanently locked and the manual unlocking channel is blocked. Injection execution and data recording and archiving: After the user unlocks injection privileges, the MCU completes the insulin injection. The MCU uses sensors to detect the injection dosage and execution status in real time, automatically records all data of this injection, synchronizes it to the App and the cloud, and archives it to the user's personal health record. Long-term data management and personalized health management: The cloud performs statistical analysis on the user's full detection, injection, and blood glucose data according to a preset cycle, generates intelligent annotated health reports and personalized health management plans, and synchronizes them to various terminals to realize full-cycle health management for home injections for diabetes.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. Pioneering a coaxial integrated non-contact detection hardware architecture, enabling in-situ synchronization of detection and injection, fundamentally solving the problem of site offset. This innovative solution coaxially integrates a high-frequency non-contact ultrasound detection module with the insulin pen tip, placing the needle at the center of the detection array. The detection area 100% overlaps with the injection site, completely resolving the site offset problem of existing separate detection devices and ensuring the clinical guidance significance of the detection results. Employing high-frequency air-coupled ultrasound technology, it achieves non-contact detection of 0.5mm-5mm without any coupling agent, completely addressing the inconvenience and poor hygiene of traditional contact ultrasound devices for home use. The integrated hardware-level safety interlock mechanism ensures that the injection button is only unlocked at sites deemed suitable for injection, eliminating the risk of accidental injection at the hardware level—a core safety innovation not found in any existing insulin pens or detection devices.

[0012] 2. A dedicated ultrasound algorithm system optimized for subcutaneous injection scenarios enables ultra-high resolution imaging of superficial tissues and precise identification of minute abnormalities. This solution optimizes the delayed summation beamforming algorithm for superficial subcutaneous tissue ranging from 1 to 10 mm, achieving a lateral resolution of 0.1 mm and an axial resolution of 0.05 mm. This significantly surpasses the superficial imaging capabilities of general ultrasound imaging algorithms, enabling the identification of minute nodules, microvessels, and nerves with diameters ≥ 0.2 mm. This addresses the limitation of insufficient resolution in superficial tissue imaging by traditional ultrasound. The preprocessing algorithm is implemented in parallel using FPGA, with a processing latency of < 100 ms, enabling real-time imaging and smooth user operation. The improved attention mechanism U-Net segmentation network, optimized for subcutaneous ultrasound images, achieves a tissue segmentation accuracy of ≥ 99%, providing a precise foundation for subsequent evaluation.

[0013] 3. A multi-dimensional, hierarchical suitability assessment algorithm enables objective and accurate automated assessment, with an accuracy rate far exceeding that of manual judgment. This solution employs a combination of analytic hierarchy process (AHP) and fuzzy comprehensive evaluation to construct a hierarchical assessment system comprising five core indicators across three dimensions: tissue anatomy, abnormal tissue, and historical injection history. This system achieves a three-tiered assessment of injection sites, balancing safety and flexibility, unlike existing technologies that rely on a simple binary judgment of the presence or absence of abnormalities. The assessment accuracy is ≥98%, significantly higher than the 60% accuracy of manual visual and tactile assessments. It effectively identifies minute abnormalities that are undetectable by humans, avoiding the risk of accidental injection. The assessment process is fully automated with a latency of <500ms, requiring no professional intervention; ordinary patients can complete the assessment at home, addressing the pain point of traditional ultrasound requiring interpretation by a professional physician.

[0014] 4. A multi-terminal collaborative intelligent annotation system throughout the entire process enables the easy-to-understand interpretation of test results, breaking down information barriers between doctors and patients. This solution pioneers an intelligent annotation system specifically for insulin injection scenarios. It employs template matching and NLG technology to achieve standardized text annotation while simultaneously providing visual annotation of ultrasound images. No professional ultrasound knowledge is required; ordinary patients and their families can interpret the images with 100% accuracy, completely solving the interpretation challenges of home ultrasound scans. The system utilizes WebSocket and MQTT protocols to achieve real-time synchronization across multiple terminals: patient, healthcare provider, and family member. Healthcare professionals can remotely view the patient's pre-injection scan and provide professional guidance, while family members can remotely monitor the process, enhancing the safety of home injections. Furthermore, it supports manual annotation by healthcare professionals, achieving a dual guarantee of automation and manual intervention, constructing a complete remote healthcare closed loop, unlike existing technologies that are single-function designs without annotation or collaboration.

[0015] 5. A fully closed-loop personalized injection management system, achieving full coverage from pre-injection to long-term management. This solution constructs a complete closed loop encompassing "pre-injection detection and assessment - in-injection safety control - post-injection data recording - long-term health management," unlike existing technologies that only focus on single dose control or detection functions and lack a comprehensive management system. The personalized recommendation algorithm, based on users' long-term data, generates personalized injection site rotation plans and health management plans. As user data accumulates, the accuracy of recommendations continuously improves, achieving precise management of diabetes and other chronic diseases. Intelligent analysis of long-term data can detect potential risks such as lipomatosis and improper injection in advance, allowing for timely intervention and effectively reducing the occurrence of insulin injection-related complications, thereby improving patients' glycemic control rates and quality of life. Attached Figure Description

[0016] Figure 1 This is an application system architecture diagram provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system composition provided in an embodiment of the present invention; Figure 3This is a schematic diagram of the internal communication system of the pen terminal hardware provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the communication link between the insulin pen and multiple terminals provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] Terminology Explanation: The core hardware is a subcutaneous insulin probe pen; the core detection unit is a high-frequency non-contact ultrasound subcutaneous detection module; the edge processing unit is an embedded dedicated signal preprocessing unit; the backend core is a cloud-based intelligent analysis and annotation engine; the user terminal is a multi-terminal collaborative intelligent annotation App terminal; and the core evaluation system is a subcutaneous injection site suitability grading evaluation system. All technical terms are kept consistent throughout the process.

[0019] I. This solution develops an insulin pen system that integrates coaxial non-contact subcutaneous detection, dedicated ultrasound imaging algorithm, graded suitability assessment, hardware-level safety interlock, multi-terminal intelligent annotation, and closed-loop injection management. It achieves in-situ accurate detection, automated assessment, safety control, intelligent interpretation, and full-cycle health management before insulin injection, fundamentally solving the safety pain points of home insulin injection and avoiding all the application shortcomings of existing technologies.

[0020] II. System Introduction like Figure 1 As shown, this application improves upon traditional insulin pens by installing a subcutaneous detection module (the module's composition is described below, and its size, specific installation location, and structure are not limited) in the pen tip (referencing the structure and installation of existing insulin pens). This module can detect subcutaneous tissue abnormalities such as nodules using contactless detection technologies such as ultrasound. The module uploads the analysis results to a backend server via a signal analysis module. Nurses, patients, and family members can directly view skin condition data via a mobile app connected to the insulin pen before injection to assess whether the skin at that location is suitable for injection.

[0021] The improved subcutaneous insulin probe pen, along with its backend system, forms an intelligent annotation system for subcutaneous insulin detection and injection. The system components will be described in detail below.

[0022] like Figure 2 As shown, the hardware architecture of the intelligent annotation system consists of the following components: 2.1 (Integrated) Subcutaneous Insulin Probe Pen The insulin pen itself (refer to the main structure of existing insulin pens; some hardware is interchangeable) is the core terminal hardware of the system, integrating subcutaneous tissue detection, injection execution, edge preprocessing, and safety control. It uses a medical-grade PC / ABS shell, is compatible with commercially available insulin cartridges and disposable needles, and has the same overall size as traditional insulin pens, making it highly portable. It consists of four core units: 2.1.1 Basic Execution Unit of Insulin Pen This unit is the basic actuator for insulin injection. It is adapted and optimized based on the structure of traditional insulin pens to achieve coaxial integration and status linkage with the detection module. The core components and working principle are as follows: Core structural components (refer to existing insulin pens for reference): cartridge compartment, spiral propulsion mechanism, dose adjustment knob, injection trigger button, piston rod, needle adapter interface, dose detection sensor, and injection status detection sensor.

[0023] Working principle: The dose adjustment knob is linked to the spiral piston rod through a precision gear transmission mechanism to achieve precise adjustment of the injection dose, with an adjustment accuracy of ±1 IU and a dose adjustment range of 0-60 IU, compatible with all clinical insulin injection dose requirements; the dose detection sensor adopts a Hall displacement sensor to detect the position of the piston rod in real time, accurately obtain the adjusted injection dose, and transmit it to the MCU main control unit through the GPIO interface; the injection trigger button is linked to the piston rod advancement mechanism and is mechanically coupled to the safety interlock mechanism. The trigger button can only be pressed to complete the injection when the safety interlock mechanism is unlocked; the needle adapter interface and the detection module are coaxially designed to ensure that the center axis of the needle is completely coincident with the imaging center axis of the detection module, with a coaxiality tolerance of <0.1mm, ensuring that the detection area is 100% consistent with the injection site, and avoiding the site offset problem from the root.

[0024] This unit is an improvement on existing insulin pens, fully compatible with current clinical insulin injection practices, allowing users to use it without changing their procedures; it has 100% dose detection accuracy, enabling precise recording of injection doses; and its coaxial design ensures in-situ synchronization between detection and injection, completely resolving the site offset issue of separate devices.

[0025] 2.1.2 High-frequency non-contact ultrasound subcutaneous detection module This module is the core detection unit of the system and represents a key innovation that avoids existing contact ultrasound technology. It employs high-frequency air-coupled ultrasound technology to achieve contactless and coupling agent-free detection of superficial subcutaneous tissue. Its core components and working principle are as follows: The core structure consists of: a 32-element high-frequency PMUT transducer array, a high-voltage pulse transmitting circuit, a low-noise echo receiving circuit, a nanosecond-level timing control circuit, a miniature acoustic matching lens, a coaxial fixed housing, and an infrared distance sensor.

[0026] How the core components work: High-frequency PMUT transducer array: Employing 20MHz-50MHz high-frequency piezoelectric micromechanical ultrasonic transducers, a 32-element linear array is arranged in a ring, with the needle located at the center of the array and an element spacing of 0.15mm. Based on the inverse piezoelectric effect, high-frequency mechanical vibration is generated under high-voltage pulse excitation, emitting ultrasonic waves. The ultrasonic waves penetrate the air and skin epidermis to enter the subcutaneous tissue. Tissues with different acoustic impedances (epidermis, dermis, fat layer, muscle layer, nodules, blood vessels, nerves) will generate reflected echoes. After the echoes are received by the transducer, they are converted into weak electrical signals based on the direct piezoelectric effect, enabling signal detection of subcutaneous tissue. The transducer front end is designed with a double-layer acoustic matching layer to adapt to the acoustic impedance difference between air and skin, eliminating echo loss at the air interface, achieving non-contact detection of 0.5mm-5mm, requiring no coupling agent, and completely solving the pain points of traditional contact ultrasound for home use.

[0027] High-voltage pulse transmission circuit: Employing a MOSFET push-pull topology, the high-voltage pulse generator outputs an adjustable pulse amplitude of ±30V, an adjustable pulse width of 50ns-200ns, and an adjustable pulse repetition frequency of 1kHz-10kHz, synchronized with the timing control circuit. It provides precisely timed high-voltage excitation pulses to each transducer element. By adjusting the transmission timing of different elements, it achieves focused ultrasonic electronic beams with a focusing depth of 1mm-10mm, fully covering the effective depth range for subcutaneous insulin injection.

[0028] The low-noise echo receiver circuit consists of a cascaded low-noise amplifier (LNA), a programmable gain amplifier (PGA), a bandpass filter (BPF), and a high-speed analog-to-digital converter (ADC). The weak μV-level echo signal output from the transducer is first pre-amplified by the LNA, achieving a noise figure <2dB to ensure effective extraction of the weak signal. Then, the signal gain is adjusted by the PGA, with a dynamic range of 60dB to adapt to echo signal amplitudes of different depths. Next, out-of-band noise is filtered out by the BPF, with a passband of 15MHz-55MHz, perfectly matching the transducer's operating frequency. Finally, the analog echo signal is converted into a digital signal by a 12-bit, 100MSPS high-speed ADC and transmitted to the FPGA co-processing unit via the LVDS interface, ensuring accurate and real-time signal sampling.

[0029] The timing control circuit is implemented using a low-power micro FPGA chip, achieving a timing control accuracy of 1 ns. It synchronously controls the pulse transmission timing, echo reception gain, and ADC sampling timing of 32 array elements, realizing ultrasonic beamforming and dynamic reception focusing. This ensures high-resolution imaging of subcutaneous tissues at different depths while avoiding signal interference between different array elements.

[0030] Miniature acoustic matching lens: Made of high-polymer acoustic material, it is mounted at the front end of the transducer array and coaxially with the needle tip. It acoustically focuses the emitted ultrasound waves, reducing the beam diameter and improving lateral resolution, ensuring a spatial resolution of 0.1mm in the detection area. It can identify subcutaneous micronodules, microvessels, and nerves with a diameter ≥0.2mm, far exceeding the superficial resolution of traditional general ultrasound.

[0031] Infrared distance sensor: Integrated into the front of the coaxial fixed housing, it detects the distance between the pen tip and the skin surface in real time with a detection accuracy of ±0.1mm. When the distance is within the effective detection range of 0.5mm-5mm, it sends a valid signal to the MCU and issues an operation reminder to the user via the App, ensuring that the detection is within the effective working distance and avoiding blurry images and inaccurate assessments caused by improper distance.

[0032] This module enables contactless and coupling agent-free subcutaneous tissue detection, making it hygienic and convenient for home use with an extremely low operating threshold. Its superficial imaging resolution far exceeds that of traditional ultrasound, accurately identifying minute abnormal tissues that cannot be detected manually. The coaxial design ensures that the detection area and injection site completely overlap, and the detection results have direct clinical guidance significance.

[0033] 2.1.3 Embedded Edge Computing and Preprocessing Unit This unit, installed inside the pen body, serves as the core computing and control center at the pen tip. It enables real-time preprocessing of echo signals, system status management, safety interlock control, and data scheduling. Its core components and working principle are as follows: Core structural components: FPGA coprocessor chip, MCU main control chip, storage module, power management module, and safety interlock driver module.

[0034] like Figure 3 As shown, the working principle of the core components is as follows: FPGA coprocessor chip: Utilizing the Microsemi IGLOO2 low-power FPGA chip, it possesses parallel computing capabilities and is responsible for real-time preprocessing of the echo signal (the specific preprocessing algorithm is detailed later; the same applies below). It receives the digital echo signal transmitted from the ADC and executes beamforming, envelope detection, logarithmic compression, and digital filtering algorithms in parallel, with a processing delay of <100ms, achieving real-time imaging. The preprocessed ultrasound image data is transmitted to the MCU main control chip via the SPI bus, solving the problem of insufficient MCU computing power and inability to achieve real-time signal processing.

[0035] MCU main control chip: Utilizing an ARM Cortex-M4F core microcontroller with a main frequency of 120MHz and floating-point arithmetic capabilities, it serves as the main controller for the pen system. It is responsible for the overall system scheduling of the pen, including start / stop control of the detection module, dose and injection status detection, FPGA preprocessing scheduling, wireless communication control, and drive control of the safety interlock mechanism; it also performs data compression, verification, and temporary storage, enabling bidirectional data interaction with the App terminal.

[0036] Storage module: Employs an 8MB SPI Flash storage chip, ensuring data integrity even after power loss. It stores pen firmware, basic user information, pre-processed temporary data, injection history records, and device configuration parameters, guaranteeing data integrity even after power failure. It can store ≥1000 complete injection and detection records.

[0037] Power Management Module: Employs a 3.7V / 500mAh rechargeable medical-grade lithium battery, paired with a dedicated charging management chip, a multi-channel LDO voltage regulator chip, and a Type-C charging interface. It provides stable power to all hardware units at the pen tip, achieving power isolation between different modules through the LDO, reducing interference from digital circuits to analog detection circuits. On a full charge, it can complete ≥500 probes + injection cycles, with a standby time of ≥6 months, meeting the needs of long-term home use. It is compatible with ordinary mobile phone chargers, offering convenient charging.

[0038] The safety interlock drive module consists of a miniature electromagnetic lock, a drive circuit, and a status feedback sensor, mechanically coupled to the injection trigger button. It receives control commands from the MCU and controls the engagement and disengagement of the electromagnetic lock via the drive circuit, thus unlocking and locking the injection trigger button. The status feedback sensor monitors the status of the electromagnetic lock in real time and sends feedback to the MCU for real-time status monitoring. The core control logic is as follows: the MCU unlocks the electromagnetic lock only when the injection site is assessed as "suitable for injection," allowing the user to trigger the injection; when assessed as "unsuitable for injection," the electromagnetic lock is permanently locked, with no manual unlocking option, eliminating the risk of accidental injection at the hardware level. This is a core safety design not found in any existing insulin pens.

[0039] This unit can perform real-time edge preprocessing of echo signals to ensure real-time imaging and reduce cloud computing pressure; it enables precise control of the entire pen system with 100% accuracy in status detection; and it features a hardware-level safety interlock mechanism to fundamentally avoid the risk of accidental injection and improve injection safety.

[0040] 2.1.4 Pen-end wireless communication module It adopts a Bluetooth 5.3 BLE low-power module with an integrated miniature FPC antenna installed inside the pen body. It connects to the MCU main control chip via a UART serial port, achieving a communication rate of 921600bps. This enables secure short-range communication between the insulin pen and a mobile app, with a communication distance of ≤10m, suitable for home use. It employs a master-slave communication mode, with the insulin pen as the slave device and the mobile app as the master device. After pairing and binding, an encrypted communication link is established, and AES-128 encryption algorithm is used to fully encrypt transmitted data, ensuring the security of medical data transmission. It supports a transmission rate of 2Mbps, enabling the transmission of pre-processed ultrasound image data within 1 second, while simultaneously receiving evaluation results and control commands from the app / cloud.

[0041] The low-power design does not affect the pen's battery life, the communication latency is less than 100ms, and the data transmission accuracy is 100%; the fully encrypted transmission meets the requirements for medical data privacy protection and avoids data leakage.

[0042] 2.2 Cloud-based intelligent analysis and annotation engine hardware cluster The cloud-based hardware cluster serves as the core computing, storage, and analysis hub of the system, enabling intelligent segmentation, feature extraction, suitability assessment, intelligent annotation generation, data storage, and multi-terminal collaboration of ultrasound images. It adopts a distributed cloud-native architecture, supporting high concurrency, high availability, and high security. Its core components and working principles are as follows: 1. Edge Computing Node Cluster: Edge computing nodes are deployed across multiple regions nationwide, employing 2U rack-mount servers configured with dual Intel Xeon Gold processors, 512GB of memory, and NVIDIA A10 GPU accelerator cards. They receive pre-processed ultrasound image data uploaded from nearby terminals and perform real-time algorithm inference, including image segmentation, feature extraction, and suitability assessment. End-to-end processing latency is <500ms, ensuring real-time user operations and reducing the computational burden on the central cluster.

[0043] 2. Core Computing Service Cluster: Utilizing a CPU+GPU heterogeneous distributed architecture, configured with high-performance Intel Xeon processors and NVIDIA A30 GPU accelerator cards, and employing Kubernetes containerized orchestration. It is responsible for training, iterating, and updating intelligent annotation algorithms, image segmentation models, and suitability assessment models; performing batch analysis of large-scale user data, calculating personalized recommendation algorithms, and generating user health reports; and scheduling the system's core business logic to ensure system stability under high concurrency scenarios, supporting ≥100,000 users simultaneously online.

[0044] 3. Distributed Storage Service Cluster: Utilizing a distributed object storage architecture with a RAID 6 redundant array and multiple off-site backups, it stores all user data, including raw detection data, ultrasound image data, injection records, assessment results, annotation information, personal health records, and blood glucose data. The data is stored using hierarchical encryption and access control is tiered, complying with the requirements of the Personal Information Protection Law and the Medical Data Security Management Standard, ensuring the integrity, security, and traceability of user medical data.

[0045] 4. Web and API Service Cluster: Employing a load-balanced architecture, it is configured with high-performance web servers and API gateways. It is responsible for interface communication with multiple terminals, handling data uploads, result distributions, and request / response processes; implementing interface authentication, rate limiting, and encryption to ensure the security and availability of the system interfaces.

[0046] 5. Security Protection Unit: Equipped with a next-generation firewall, intrusion detection system (IDS), data encryption gateway, and security audit system. This provides network security protection to defend against hacker attacks and malicious intrusions; enables end-to-end security auditing of data transmission, storage, and access to prevent data leakage; and meets the Level 3 security compliance requirements for medical data, ensuring system compliance.

[0047] 2.3 Multi-terminal Collaborative Intelligent Annotation Terminal This layer serves as the system's user interaction terminal, enabling data display, intelligent annotation viewing, operation control, remote monitoring, and health management functions. It primarily comprises three types of terminals: Patient-side smart terminal: Supports smartphones and tablets running iOS / Android systems. It communicates with insulin pens and cloud engines through the installation of a dedicated app. It is the core operating terminal for users, enabling functions such as device pairing, detection control, image viewing, annotation interpretation, injection record viewing, personalized recommendation reception, and health report viewing.

[0048] Healthcare management terminal: Supports Windows / macOS computers and tablets. Access to the system is achieved through a web management platform or a dedicated client. Healthcare staff can view the probe data, injection records, and health reports of bound patients, add professional human annotations, adjust patients' injection plans, and achieve remote guidance and chronic disease management.

[0049] Family monitoring terminal: Supports iOS / Android smartphones, and can access the system through a dedicated app. It can view the injection completion status, assessment results, and blood glucose data of the bound patient, receive injection reminders and alarms for abnormal situations, and realize remote monitoring of the patient. It is especially suitable for the use scenarios of elderly people living alone and patients with limited mobility.

[0050] III. Data Stream Communication Transmission Mechanism This system adopts a layered and hierarchical communication architecture to achieve secure, stable, and real-time data transmission across the entire link. It employs encrypted communication mechanisms throughout, meeting medical data security requirements. Details are as follows: 3.1 Internal communication of pen terminal hardware like Figure 3 As shown, the various modules within the pen terminal employ highly reliable, low-latency communication interfaces to achieve high-speed data exchange, as detailed below: 1. Detection module and FPGA co-processing unit: Communicate using an LVDS (Low Voltage Differential Signaling) interface with a communication rate of 1Gbps. The LVDS interface features strong anti-interference capability, low transmission delay, and low power consumption, enabling lossless transmission of digital echo signals after high-speed ADC sampling, ensuring the real-time performance and integrity of echo data, and avoiding image quality degradation caused by signal interference.

[0051] 2. FPGA coprocessor unit and MCU main control unit: Communication is achieved through an SPI serial peripheral interface with a communication rate of 50Mbps. The SPI interface features low pin usage, high transmission rate, and full-duplex communication, enabling the transmission of preprocessed image data from the FPGA to the MCU, while simultaneously allowing the MCU to send control commands and configuration parameters to the FPGA, ensuring the collaborative operation of the edge computing units.

[0052] 3. MCU Main Control Unit and Peripheral Modules: Communication is achieved through a GPIO general-purpose input / output interface. Peripheral modules include a safety interlock driver module, a dose detection sensor, an injection status detection sensor, a distance sensor, and a detection button. The MCU outputs high and low levels via the GPIO interface to control the unlocking / locking of the electromagnetic lock. Simultaneously, it reads the status data of each sensor in real time through the GPIO interface, enabling real-time detection and precise control of the system status with a response latency of <1ms.

[0053] 4. MCU Main Control Unit and Wireless Communication Module: Utilizing a UART asynchronous serial interface with a communication rate of 921600bps, this module enables bidirectional data transmission between the MCU and the Bluetooth module. This includes uplink probe data, device status, and injection records, and downlink control commands, evaluation results, and configuration parameters. Simultaneously, it performs data stream packetization, reassembly, and verification to ensure data transmission accuracy.

[0054] like Figure 3 The diagram shows the system communication link mechanism.

[0055] 3.2 Short-range communication between the pen and the patient's app The Bluetooth 5.3 BLE protocol is used to achieve short-range wireless communication between the pen and the mobile app. The specific communication mechanism is as follows: Pairing and Binding Mechanism: Upon first use, the insulin pen enters broadcast mode after power-on, broadcasting the device's unique hardware ID, device type, and service UUID. The user scans the broadcast signal through the App, selects the corresponding device, and initiates a pairing request. During the pairing process, the device and the App generate a unique AES-128 session key using the Diffie-Hellman key exchange algorithm, establish an encrypted communication link, and complete the pairing and binding. Each device can only be bound to one main user account to prevent unauthorized device connections and ensure the security of device use.

[0056] Data transmission mechanism: After pairing, the device and the App establish a persistent BLE connection, using the ATT protocol to achieve bidirectional data transmission; Uplink data includes preprocessed ultrasound image data, device status data, dosage data, and injection records, compressed using the LZ4 lossless compression algorithm with a compression ratio ≥5:1, transmitted in packets, with each data packet having a 16-bit CRC checksum added to ensure data transmission integrity; Downlink data includes evaluation results, safety interlock control commands, device configuration parameters, and user information sent from the cloud, transmitted encrypted, and verified and decrypted upon receipt to ensure the accuracy and security of commands.

[0057] 3.3 Wide Area Network Communication between Patient-Side App and Cloud Engine Combined with appendix Figure 4 As shown, a dual-protocol architecture using HTTPS and WebSocket protocols is employed to achieve secure wide-area network communication between the App and the cloud engine, as detailed below: HTTPS protocol: Employs TLS 1.3 encryption for non-real-time large data transmission, including uploading and downloading pre-processed ultrasound image data, user personal information, historical injection records, and health reports; it enables interface authentication, user identity verification, and encrypted data transmission to prevent man-in-the-middle attacks and ensure data transmission security.

[0058] WebSocket protocol: used for real-time bidirectional communication, establishing a persistent connection between the client and the cloud with a transmission latency of <200ms; used for the transmission of real-time imaging images, real-time distribution of evaluation results, synchronization of intelligent annotation information, and real-time data push for remote monitoring, ensuring the real-time nature of user operations and enabling synchronized updates of data across multiple devices.

[0059] 3.4 Collaborative Communication between Cloud Engine and Multiple Terminals The MQTT 3.1.1 protocol is used to achieve multi-terminal collaborative communication between the cloud engine and patients, medical staff, and family members. It is based on a lightweight publish / subscribe communication model (e.g., establishing an independent topic space for each user, including sub-topics such as "Assessment Results," "Injection Records," "Abnormal Alarms," ​​"Comment Information," and "Health Reports"; patients subscribe to all topics, medical staff subscribe to all topics bound to the patient, and family members subscribe only to monitoring-related topics, achieving hierarchical data push and ensuring data privacy), adapting to the low-bandwidth, high-reliability communication requirements of multiple terminals. The specific mechanism is as follows: IV. System Chip Layer Execution Algorithm Description 4.1 Pen-embedded echo signal preprocessing algorithm This algorithm runs in parallel on the FPGA at the pen end. Its core is a delay-and-sum (DAS) beamforming algorithm optimized for superficial subcutaneous tissue, which enables real-time processing of ultrasound echo signals and generates high-resolution ultrasound images of subcutaneous tissue, which form the basis for all subsequent analyses.

[0060] 4.1.1 Delay Summation Beamforming Core Formula and Definition The core of beamforming is to compensate for the path difference between echoes from different array elements to achieve in-phase superposition of echo signals, thereby improving the imaging signal-to-noise ratio and resolution. The core formula is as follows: Detailed definitions of formula symbols and operation rules: Character parameter definition: The output signal after beamforming, measured in volts (V), represents time. and the horizontal coordinates of the imaging point The function corresponding to the imaging point The tissue echo amplitude is the core data for generating ultrasound images; Time variable, in seconds, corresponds to the propagation time of ultrasound waves in tissue, and is directly related to the depth of the imaging point; The horizontal coordinate of the imaging point, in meters, is based on the transducer array center (needle center) as the origin, with the horizontal axis as the x-axis and the depth axis as the z-axis. : Depth coordinates of the imaging point, in meters, i.e., the depth of the subcutaneous tissue, calculated from the skin surface. The effective range of this algorithm is 1mm-10mm. The total number of array elements of the transducer array is fixed in this scheme. ; : No. The amplitude weighting coefficients for each array element are dimensionless and used to reduce beam sidelobes and improve imaging contrast. This scheme uses Hanning window weighting, and the calculation formula is as follows: , ; : No. The original digital echo signal received by each array element, in V, is a time-domain discrete signal after ADC sampling. : No. Each element relative to the imaging point The delay compensation amount, measured in seconds, is the core parameter of this algorithm. It is used to compensate for the acoustic path difference between different array elements and the imaging point, thereby achieving in-phase superposition of echo signals.

[0061] 4.1.2 Calculation Formula and Definition of Delay Compensation The delay compensation is determined by the array element position, the imaging point position, and the propagation speed of ultrasound waves in the tissue. The calculation formula is as follows:

[0062] Detailed definitions of formula symbols: : No. The horizontal coordinates of each array element, in meters, and the element spacing of the 32 array elements in this scheme. ,therefore This ensures that the center of the array is completely aligned with the center of the needle, perfectly matching the coaxial design of the hardware. The average speed of ultrasound propagation in human subcutaneous soft tissue, measured in m / s, is fixed in this scheme. Clinically verified, this value has a matching error of less than 1% with the sound velocity of human subcutaneous fat and dermal tissue, ensuring the accuracy of depth calculation; The remaining symbols are consistent with the definition in formula (1).

[0063] 4.1.3 Envelope Detection and Logarithmic Compression Algorithm After beamforming, the radio frequency signal needs to undergo envelope detection to extract tissue amplitude information. The dynamic range is then displayed through logarithmic compression and adaptation. The core formula is as follows: Hilbert transform envelope detection: , , Symbol definition: For Hilbert transform operators, Beamforming signal orthogonal components, The envelope signal of the echo signal, in V, reflects the acoustic impedance difference of subcutaneous tissue and the corresponding structural information of the tissue. For convolution integral operations, the Hilbert transform is essentially a signal-to-interaction (SI) operation. The convolution operation.

[0064] Logarithmic compression algorithm:

[0065] Symbol definition: The compressed image grayscale value is dimensionless, ranging from 0 to 255, and is adapted for 8-bit grayscale image display. envelope signal The maximum value, in V; The base-10 logarithmic operation is used to compress the echo signal with a 60dB dynamic range to the range that the display is adapted to; +1 is used to avoid the meaningless case where the independent variable of the logarithmic operation is 0.

[0066] 4.1.4 Operating Mechanism ① The timing control circuit controls 32 array elements to emit high-voltage pulses according to a preset timing sequence, which excites the transducer to emit ultrasonic waves and focus them on the target imaging point; ② When ultrasound waves propagate in subcutaneous tissue, they generate echoes at tissue interfaces with different acoustic impedances. These echoes are received by 32 array elements, converted into electrical signals, and then sampled by an ADC to obtain the original echo signal. ; ③ The FPGA calculates the distance from each array element to the imaging point in parallel according to formula (2). Delay compensation amount To compensate for the path difference between different array elements; ④ The FPGA performs a weighted summation of the delayed echo signals of each array element according to formula (1) to obtain the beamforming output signal of the imaging point. ; ⑤ Repeat the above steps for all imaging points within the detection area (lateral range -2.5mm +2.5mm, depth range 1mm +10mm) to obtain the radio frequency echo data of the entire detection area; ⑥ According to formulas (3)-(5), envelope detection, logarithmic compression and digital filtering are performed on the radio frequency data to obtain a preprocessed 500×500 pixel grayscale ultrasound image, which is then transmitted to the MCU and uploaded to the App and the cloud via Bluetooth.

[0067] This algorithm is perfectly matched with the timing control circuit of the pen tip hardware and the FPGA parallel computing architecture, achieving high-resolution real-time imaging of superficial subcutaneous tissue. It provides high-quality foundational data for subsequent cloud-based image segmentation, feature extraction, and suitability assessment, serving as the signal processing basis for the entire system. The beamforming algorithm optimized for superficial subcutaneous tissue achieves a lateral resolution of 0.1mm and an axial resolution of 0.05mm, with a sidelobe suppression ratio >40dB, far exceeding the superficial imaging capabilities of general ultrasound algorithms. The FPGA parallel processing latency is <100ms, enabling real-time imaging with smooth user operation and allowing for real-time adjustment of the pen tip position to find suitable injection sites.

[0068] 4.2 Cloud-based subcutaneous tissue ultrasound image segmentation and feature extraction algorithm This algorithm runs on cloud edge computing nodes. Its core consists of an improved attention mechanism U-Net image segmentation network and a multi-dimensional feature extraction algorithm, which enables accurate stratification of subcutaneous tissue and identification of abnormal tissue. It extracts core quantitative features related to injection suitability and provides input for evaluating the algorithm.

[0069] 4.2.1 Improved U-Net Subcutaneous Tissue Segmentation Network To address the characteristics of subcutaneous ultrasound images, the traditional U-Net network was optimized by adding a CBAM convolutional attention module to each convolutional block of the encoder to enhance the extraction of effective features and suppress background noise. The network adopts an encoder-decoder architecture with 4 downsampling layers and 4 upsampling layers. Skip connections are used to fuse shallow detail features with deep semantic features, outputting pixel-level segmentation results for 7 categories: epidermis, dermis, fat layer, muscle layer, nodule / induration region, vascular region, and nerve region, with a segmentation accuracy of ≥99%.

[0070] 4.2.2 Core Feature Extraction Formula and Definition Based on the segmentation results, three major categories and five core evaluation features are extracted. All features are quantitatively calculated to provide objective input for suitability assessment. The core formulas are as follows: Core feature of tissue layering: thickness of the fat layer Fat layer thickness is a key assessment indicator for subcutaneous injection. Insulin needs to be injected into the fat layer to ensure stable absorption. Insufficient thickness may lead to injection into the muscle layer, while excessive thickness may result in incomplete injection. The calculation formula is as follows:

[0071] Symbol definition: The thickness of the fat layer at the detection site is expressed in mm. The depth of the upper boundary of the fat layer (the interface between the dermis and the fat layer) is measured in mm. The depth of the lower boundary of the fat layer (the interface between the fat layer and the muscle layer) is in mm, and is calculated from the segmentation network results.

[0072] Core characteristics of abnormal organizations ① Nodule area percentage: This reflects the severity of fat hyperplasia and induration within the detected area. The calculation formula is as follows:

[0073] Symbol definition: The percentage of the nodule area is expressed as a percentage (%). The total area of ​​the nodule / hardened region obtained by segmentation is expressed in mm². The area of ​​interest for detection is the circular region with a diameter of 5mm directly below the probe, with a fixed value. .

[0074] ② Maximum vessel diameter: Reflects the distribution of blood vessels within the detection area, avoiding severe hypoglycemia caused by injection into blood vessels. The calculation formula is as follows:

[0075] Symbol definition: The maximum diameter of blood vessels within the detection area is expressed in mm. For the first The diameter of each blood vessel, in mm; To detect the total number of blood vessels in the area; This is a maximum value extraction operation.

[0076] ③ Neural distribution characteristics: Binary feature. If the segmentation result shows that there is neural tissue in the detection area, the feature value is 1; otherwise, it is 0. Injection is absolutely prohibited at sites where neural tissue is present.

[0077] Key characteristics of historical injections: injection frequency at injection sites This reflects the repeat injection status at this site and helps avoid fat hyperplasia caused by repeated injections. The calculation formula is as follows:

[0078] Symbol definition: This represents the injection frequency at this site, expressed as injections per day. For the past Total number of injections at this site within one day; The statistical time window is fixed at 30 days.

[0079] 4.2.3 Operating Mechanism ① The cloud edge computing node receives the preprocessed ultrasound image uploaded by the App, inputs it into the improved U-Net segmentation network, and obtains a pixel-level tissue segmentation mask; ② Based on the segmentation results, the quantitative values ​​of the five core evaluation features are calculated in parallel according to formulas (6)-(9); ③ Store the feature values ​​and segmentation results in the user's personal health record, and input them into the injection site suitability assessment algorithm.

[0080] The segmentation results of this algorithm directly determine the accuracy of feature extraction, and the feature values ​​are the sole input to the suitability assessment algorithm. This enables the conversion from image to quantitative assessment indicators, serving as the core bridge connecting imaging and assessment. It is fully integrated with subsequent assessment algorithms, annotation algorithms, and control logic. The accuracy of tissue layering and abnormal tissue segmentation is ≥99%, with feature calculation precision ±0.1mm. It achieves fully automated extraction of injection site suitability-related features without manual intervention, providing a precise quantitative basis for subsequent objective assessment.

[0081] 4.3 Intelligent Algorithm for Injection Site Suitability Assessment This algorithm is the core decision-making algorithm of the system. It adopts the analytic hierarchy process (AHP) + fuzzy comprehensive evaluation method and combines clinical expert experience to build a multi-dimensional hierarchical evaluation system to realize the automated and objective hierarchical evaluation of the suitability of the injection site. The output evaluation results directly link the control logic of the pen tip safety interlock mechanism.

[0082] 4.3.1 Evaluation Index System and Weight Calculation Based on clinical guidelines and expert surveys, a three-tiered assessment indicator system was constructed. The weight of each indicator was determined using the analytic hierarchy process (AHP). The core of the system is as follows: Primary indicators: Tissue anatomical structure indicators (weight) ), abnormal organization indicators (weight) ), historical injection indicators (weight) ); Secondary indicators: Anatomical structural parameters: Fat layer thickness (weight) Total weight ); Abnormal tissue indicators: Nodule percentage (weight) Total weight ), Maximum diameter of blood vessels (weight) Total weight ), neural distribution (weights) Total weight ); Historical injection metrics: Site injection frequency (weighted) Total weight ).

[0083] The weights are calculated using the analytic hierarchy process (AHP) judgment matrix, and the consistency check formula is as follows:

[0084] , Symbol definition: As a consistency indicator, To determine the largest eigenvalue of a matrix, Let be the order of the matrix. For consistency ratio, The average random consistency index; when At this point, the judgment matrix exhibits satisfactory consistency, and the weights are effective. In this scheme, the judgment matrix... The weights meet the consistency requirements.

[0085] 4.3.2 Membership Function and Fuzzy Comprehensive Evaluation The fuzzy comprehensive evaluation method is adopted to construct the membership function of each indicator to three evaluation levels (suitable for injection, cautious injection, and unsuitable for injection), convert the quantitative features into membership degrees, and obtain the comprehensive evaluation result through weighted synthesis.

[0086] (1) Evaluation level and membership function The three evaluation levels are: suitable for injection (Level 1), cautious for injection (Level 2), and unsuitable for injection (Level 3). Based on clinical guidelines, a piecewise membership function for each indicator is constructed. Taking the core indicator of fat layer thickness as an example, the membership function is as follows: Suitable injection membership function:

[0087] Inject membership functions with caution:

[0088] (2) Injection of membership functions is not suitable:

[0089] Symbol definition: These represent the degree of membership of fat layer thickness to three levels: suitable, cautious, and unsuitable. The values ​​range from 0 to 1, with larger values ​​indicating a higher degree of belonging to that level. The thickness of the fat layer is calculated using formula (6).

[0090] Similarly, membership functions were constructed for nodule percentage, maximum blood vessel diameter, nerve distribution, and injection frequency. The core rules are as follows: nodule percentage <5% is suitable, 5%-20% is cautious, and >20% is unsuitable; maximum blood vessel diameter <0.3mm is suitable, 0.3-0.8mm is cautious, and >0.8mm is unsuitable; presence of nerve distribution is unsuitable, absence is suitable; injection frequency <0.1 times / day is suitable, 0.1-0.3 times / day is cautious, and >0.3 times / day is unsuitable.

[0091] (3) Fuzzy comprehensive evaluation synthesis formula Based on the weight vector and membership matrix, a fuzzy synthesis is performed using a weighted average operator to obtain the comprehensive evaluation result, as shown in the following formula: ,

[0092] Symbol definition: This is a vector representing the comprehensive evaluation results. These represent the comprehensive membership degree to the three levels: suitable, cautious, and unsuitable. This is the total weight vector for the five secondary indicators; It is a 5×3 membership matrix, with each row corresponding to the membership degree of an indicator to three levels; As a fuzzy synthesis operator, this scheme adopts a weighted average operator to ensure that the impact of all indicators is included in the evaluation.

[0093] (4) Rules for judging evaluation results Based on the comprehensive evaluation result vector, the level is determined according to the principle of maximum membership, and the safety interlock control logic is activated simultaneously. like and If the system determines that the injection is appropriate, an unlocking command is sent to the pen tip, the electromagnetic lock unlocks, and injection is permitted. like If the injection is deemed to be cautious, the electromagnetic lock will remain locked, the App will display a risk warning, and after the user manually confirms, an unlocking command will be issued to allow the injection. like or If the injection is deemed unsuitable, the electromagnetic lock should be permanently locked, with no manual unlocking method available, and injection is prohibited.

[0094] 4.3.3 Algorithm Operation Mechanism ① The cloud receives the quantified values ​​of the five core indicators output by the feature extraction module; ② The membership calculation module calculates the membership degree of each indicator to the three levels based on the membership function of each indicator, and constructs a membership matrix. ; ③ The fuzzy comprehensive evaluation module obtains the comprehensive evaluation result vector by weighted synthesis of the weight vector and the membership matrix according to formulas (15)-(16). ; ④ Based on the judgment rules, the final suitability assessment level is obtained, and the corresponding safety interlock control command is generated; ⑤ The evaluation results, control commands, and feature values ​​are sent to the App and the pen, and simultaneously input into the intelligent annotation algorithm.

[0095] This algorithm serves as the system's decision-making center, directly determining the action of the pen-end safety interlock mechanism. It is also the core input to the intelligent annotation algorithm, connecting the entire process of front-end detection, edge computing, cloud analysis, and terminal execution, and is the core of the system's security logic. Its multi-dimensional comprehensive evaluation accuracy far exceeds that of manual judgment; the evaluation latency is <500ms, imperceptible to the user; the tiered evaluation balances security and flexibility, and hardware-level linkage eliminates the risk of accidental injection at its source—a core innovation not found in existing technologies.

[0096] 4.4 Multi-terminal Intelligent Annotation and Personalized Recommendation Algorithm This algorithm is the core of the intelligent annotation system, enabling automated, standardized, and visualized annotation of detection results, as well as personalized injection plan recommendations. It solves the pain point of home users being unable to interpret ultrasound images and achieves multi-terminal collaborative health management.

[0097] 4.4.1 Real-time detection result intelligent annotation algorithm A template matching + natural language generation (NLG) technical architecture is adopted. By combining evaluation results, feature values, and segmentation results, standardized text annotations and visual image annotations are generated. The core mechanism is as follows: Text annotation generation: For the three assessment levels, standardized clinical annotation templates are constructed. The template content is automatically populated based on feature values ​​to generate easy-to-understand annotation text that can be comprehended without requiring specialized knowledge. Suitable Injection Site: [Injection Site Assessment Result: Suitable for Injection] The fat layer thickness at this site is X.X mm, within the optimal injection range (3-6 mm); there are no obvious nodules, large blood vessels, or nerve distributions within the detection area; the injection frequency at this site over the past 30 days has been XX times / day, with no risk of repeated injections. It is recommended to inject at this site, with a suggested injection depth of X.X mm.

[0098] Cautionary Injection Template:

Injection Site Assessment Result: Cautionary Injection

[0099] Injection Site Not Suitable: [Injection Site Assessment Result: Not Suitable for Injection] The fat layer thickness at this site is X.X mm, which does not meet the requirements for subcutaneous injection. Nerve distribution is present in the detection area. Nodules account for XX% of the area. The maximum vessel diameter is X.X mm, posing a serious injection risk. Injection at this site is prohibited. Please choose another injection site.

[0100] Visual image annotation: Based on the segmentation results, pixel-level annotations are performed on the ultrasound images, including: green lines to mark the upper and lower boundaries of the fat layer and label the thickness values; red areas to mark nodules, blood vessels, and nerves and label their size / diameter; green boxes to mark suitable injection areas and red boxes to mark unsuitable areas; and the evaluation level and core parameters are marked in the upper right corner of the image to achieve intuitive visual interpretation.

[0101] 4.4.2 Personalized Injection Site Recommendation Algorithm Based on users' historical data, physical characteristics, and blood glucose levels, a personalized recommendation algorithm is constructed to calculate the recommendation priority of optional injection sites and generate an injection site rotation plan. The core formula is as follows:

[0102] Symbol definition: The higher the value, the higher the recommendation level for the injection site. The site's anatomical structure was scored (0-100 points). The score for the site of abnormal tissue is 0-100. Historical injection scores for the site (0-100 points). The site-specific blood glucose response score (0-100 points); These are weighting coefficients, which maintain logical consistency with the weights in the evaluation algorithm.

[0103] Algorithm operation mechanism: The system divides the user's injection area (abdomen, thigh, upper arm, buttock) into 20 independent small blocks. Based on the user's historical detection data, injection records, and blood glucose data, the system calculates the recommendation priority of each block according to formula (17), sorts them by priority to generate a list of recommended injection sites, and generates annotation information explaining the reasons for the recommendation. This information is displayed on the App homepage to guide users to rotate injection sites and avoid repeated injections.

[0104] 4.4.3 Algorithm Operation Mechanism ① The cloud receives the evaluation results, feature values, and segmentation results, and inputs them into the intelligent annotation algorithm; ② The text annotation module generates standardized annotation text based on template matching, and the visualization annotation module generates annotated ultrasound images; ③ The personalized recommendation module calculates the site recommendation priority based on the user's historical data according to formula (17), and generates a recommendation list and annotation information; ④ Synchronize the annotation information, recommendation list, and labeled images to the patient's, medical staff's, and family's devices via the MQTT protocol; ⑤ Medical staff can add professional annotations, which are synchronized to all terminals, achieving a combination of automated and manual annotation.

[0105] This algorithm transforms technical data into user-understandable information, serving as the core of the system's user interaction. It also enables multi-terminal collaboration among patients, medical staff, and family members, constructing a complete closed loop for remote guidance and monitoring. The annotations conform to clinical standards, are easy to understand, and achieve 100% accuracy for ordinary users, addressing the pain points of interpreting home ultrasound scans. Real-time synchronization across multiple terminals enables remote monitoring and professional guidance, while personalized recommendations effectively reduce the risk of lipomatosis caused by repeated injections and improve insulin therapy efficacy.

[0106] Example 2 Based on the above control logic, the application control method of this system is described in detail below: Step 1: System initialization and user pairing / binding 1.1 When the user turns on the power switch of the insulin pen, the MCU main control chip at the pen tip starts a full hardware self-test, including the function and status detection of the detection module, communication module, safety interlock module, power module, and sensor module. If the self-test passes, the Bluetooth module enters broadcast mode and broadcasts the device's unique hardware ID and service UUID. If the self-test fails, the App will pop up a fault reminder and disable the detection function.

[0107] 1.2 Users install and open the dedicated smart annotation app on their smartphones, complete account registration and real-name authentication, fill in basic personal information (name, gender, age, type of diabetes, type of insulin, injection plan, blood glucose target, etc.), establish personal electronic health records, and store them in a cloud distributed storage cluster.

[0108] 1.3 Users click “Add Device” in the App, scan the Bluetooth broadcast of the insulin pen, select the corresponding device to initiate a pairing request; during the pairing process, the device and the App complete the key exchange, establish an AES-128 encrypted communication link, and complete the pairing and binding. Each device can only be bound to one main user account. After binding, the user's basic information is automatically synchronized to the pen storage module.

[0109] 1.4 After pairing is complete, the App automatically detects the pen's firmware version. If a new version is available, the App prompts the user to complete the OTA firmware upgrade. At the same time, the App synchronizes the user's injection plan and evaluation parameter configuration to the pen, completing the entire system initialization process.

[0110] The initial pairing and binding process takes less than 30 seconds, is simple to operate, and requires no professional knowledge; the hardware self-test has a 100% fault detection rate, preventing the device from operating with faults; end-to-end encryption complies with medical data compliance requirements, ensuring user privacy and security.

[0111] Step 2: Pre-injection preparation and detection initiation 2.1 Following the clinical insulin injection guidelines, the user completes the installation of the insulin cartridge, needle, and air venting, and rotates the dose adjustment knob to set the injection dose; the dose detection sensor detects the adjusted dose value in real time and transmits it to the MCU through the GPIO interface, which is then displayed synchronously on the App interface, allowing the user to verify the dose accuracy.

[0112] 2.2 The user aligns the insulin pen tip with the skin site to be injected, keeping the pen tip perpendicular to the skin surface. The infrared distance sensor detects the distance between the pen tip and the skin in real time. When the distance is within the effective detection range of 0.5mm-5mm, the App issues a voice and text reminder saying "The distance is appropriate, detection can be started"; if the distance exceeds the range, the App issues an adjustment reminder to ensure that the detection is in an effective working state.

[0113] 2.3 When the user presses the detection button on the pen tip or the "Start Detection" button on the App interface, the MCU receives the detection start command and sends a start command to the timing control circuit. The high-frequency non-contact ultrasound detection module starts, beginning continuous ultrasound transmission and echo reception. The FPGA synchronously starts real-time echo signal preprocessing, and the App interface displays subcutaneous tissue ultrasound images in real time. This achieves non-contact detection, is hygienic and convenient, avoids cross-infection, and allows the user to adjust the pen tip position in real time to find a suitable site.

[0114] Step 3: Echo signal preprocessing and data transmission 3.1 The echo signals received by the 32 array elements of the detection module are amplified, filtered, and sampled by the echo receiving circuit and then converted into digital echo signals, which are then transmitted at high speed to the FPGA coprocessing chip through the LVDS interface.

[0115] 3.2 The FPGA performs delay summation beamforming, Hilbert transform envelope detection, logarithmic compression, and digital filtering on the digital echo signal in parallel according to formulas (1)-(5) to obtain the preprocessed 500×500 pixel grayscale ultrasound image data.

[0116] 3.3 The FPGA transmits the preprocessed image data to the MCU main control chip via the SPI bus. The MCU performs LZ4 lossless compression on the data, adds a 16-bit CRC checksum, and transmits it to the Bluetooth module via the UART interface.

[0117] 3.4 The Bluetooth module transmits compressed image data, device status, dose data, and operation data to the mobile app via Bluetooth 5.3 BLE protocol. After receiving the data, the app decompresses and performs CRC verification. Once the verification is successful, the ultrasound image is displayed on the interface in real time. At the same time, the data is encrypted and uploaded to the cloud-based intelligent analysis and annotation engine via HTTPS protocol.

[0118] Step 4: Cloud-based intelligent analysis and injection site suitability assessment 4.1 The cloud web server receives the data uploaded by the App, completes decryption, identity authentication, and data verification. After successful verification, the data is distributed to the nearest edge computing node.

[0119] 4.2 Edge computing nodes input the preprocessed ultrasound image into the improved U-Net segmentation network to complete pixel-level tissue segmentation and obtain segmentation masks for the epidermis, dermis, fat layer, muscle layer, nodules, blood vessels, and nerves.

[0120] 4.3 Based on the segmentation results, the feature extraction module calculates the quantitative values ​​of five core evaluation indicators, namely, fat layer thickness, nodule ratio, maximum blood vessel diameter, nerve distribution, and site injection frequency, according to formulas (6)-(9).

[0121] 4.4 The suitability assessment module calculates the membership matrix based on the membership function of each indicator, performs weight calculation and fuzzy comprehensive evaluation according to formulas (10)-(16), obtains the comprehensive evaluation result vector, and determines the final suitability assessment level according to the judgment rules.

[0122] 4.5 The cloud archives and stores the evaluation results, segmentation results, feature values, and evaluation process data in the user's personal health record, while inputting the evaluation results and control instructions into the intelligent annotation algorithm and terminal control module.

[0123] By improving the accuracy of assessment through evaluation and analysis, it can effectively identify minute anomalies that cannot be detected manually; the end-to-end processing latency is less than 500ms, so users are unaware of the process; the entire assessment process is fully traceable, meeting the requirements of medical data management, and providing accurate decision-making basis for subsequent security control and annotation.

[0124] Step 5: Intelligent annotation generation and multi-device synchronization 5.1 The cloud-based intelligent annotation module generates standardized text annotation information based on the evaluation results, feature values, and segmentation results through template matching + NLG technology. At the same time, it completes visual annotation on the ultrasound image, generating an annotated ultrasound image.

[0125] 5.2 The personalized recommendation module calculates the recommendation priority of all available injection sites according to formula (17) based on the user's historical injection data and current site evaluation results, and generates a list of recommended injection sites and corresponding annotations.

[0126] 5.3 The cloud uses the WebSocket protocol to send the assessment results, text annotations, visual annotation images, recommended site list, and safety interlock control commands to the patient's App in real time; at the same time, it uses the MQTT protocol to synchronize relevant information to the bound medical staff management platform and family monitoring terminal.

[0127] 5.4 After receiving the data, the patient-side App displays annotated images, text annotations, assessment results, and recommended sites on the interface, and issues corresponding voice prompts based on the assessment level. Medical staff can view the patient's detection data and assessment results on the medical staff's end, add professional manual annotations, adjust the injection plan, and synchronize the data to the patient's end and the family's end. The family's end can view relevant information in real time, realizing remote monitoring.

[0128] Intelligent annotation eliminates the need for professional ultrasound knowledge, enabling accurate interpretation for ordinary users and resolving the pain points of home ultrasound interpretation. Real-time synchronization across multiple devices breaks down the barriers between home and hospital, enabling remote guidance from medical staff and remote monitoring by family members, thus improving the safety of home injections.

[0129] Step 6: Safety Interlock Control and Injection Access Management 6.1 After receiving the security interlock control command from the cloud, the patient-side App transmits the command to the MCU main control chip of the insulin pen via an encrypted Bluetooth link.

[0130] 6.2 The MCU drives the safety interlock module to perform corresponding actions according to the control instructions, strictly following the hierarchical control rules: If the evaluation result is "suitable for injection": the MCU outputs a high level to the electromagnetic lock drive circuit, the electromagnetic lock engages and unlocks, the mechanical limit of the injection trigger button is released, the user can press the button to complete the injection, and the App interface displays "Injection permission unlocked"; If the assessment result is "Inject with caution": The MCU keeps the electromagnetic lock locked, and a risk confirmation window pops up on the App interface, showing the reason for cautious injection and potential risks. After the user reads and clicks "Confirm Injection", the App will transmit the unlocking command to the MCU, and the MCU will control the electromagnetic lock to unlock, allowing injection; If the assessment result is "not suitable for injection": the MCU keeps the electromagnetic lock permanently locked, the App interface pops up a reminder prohibiting injection, explaining the reasons for unsuitability, and at the same time, all manual unlocking options are blocked, and there is no way to unlock the injection button. Injection at this site is absolutely prohibited.

[0131] 6.3 The status of the electromagnetic lock is transmitted to the MCU in real time through the feedback sensor. The MCU synchronizes the status data to the App and the cloud. All terminals display the current injection permission status in real time, realizing full monitoring and traceability.

[0132] Hardware-level safety interlocks eliminate the risk of injection at inappropriate sites, avoiding complications such as hypoglycemia and malabsorption caused by injection into the muscle layer, blood vessels, or nodules; the absolute prohibition mechanism prevents user misoperation and wishful thinking, which is a core safety function that existing insulin pens do not have.

[0133] Step 7: Injection Execution and Data Recording and Archiving 7.1 After the user unlocks the injection privileges, they insert the needle into the skin according to clinical guidelines, press the injection trigger button, and the piston rod advances to complete the insulin injection. During the injection, the MCU uses a Hall sensor to detect the advancement status of the piston rod in real time, confirm the completion of the injection dose, and ensure that the insulin is completely injected.

[0134] 7.2 After the injection is completed, the user removes the needle, and the MCU automatically records all the data of this injection, including: injection time, user ID, device ID, injection dose, injection site, original detection image, evaluation results, annotation information, and data of the entire injection operation process, and stores it in the pen-end Flash storage module to ensure that it is not lost when power is off.

[0135] 7.3 The MCU synchronizes all data from this injection to the App via Bluetooth, and the App uploads it to the cloud engine via HTTPS protocol. The cloud archives the data to the user's personal health record and updates the user's injection site history data, injection frequency statistics, and blood glucose data correlation model.

[0136] 7.4 The App interface displays confirmation information for the completion of this injection, provides precautions such as post-injection dwell time, needle handling, and skin care, and updates the recommended injection site and time reminder for the next injection.

[0137] Injection data is recorded fully automatically, eliminating the need for manual input by the user. The recording accuracy is 100%, avoiding errors and omissions caused by manual recording. Multiple copies of the data ensure that no data is lost, and even if the pen tip is damaged, the data can still be fully traced. Standardized data archiving provides medical staff with complete clinical evidence for adjusting treatment plans.

[0138] Step 8: Long-term data management and personalized health management 8.1 The cloud engine performs statistical analysis on the user's full detection data, injection data, and blood glucose data on a daily, weekly, and monthly basis, generating corresponding intelligent annotated health reports, including: injection site distribution analysis, nodule detection rate analysis, fat layer thickness change analysis, injection operation standardization analysis, blood glucose control effect analysis, insulin absorption effect analysis, etc., while providing corresponding risk warnings and improvement suggestions.

[0139] 8.2 The personalized recommendation module optimizes the weight coefficients of the recommendation algorithm based on long-term user data, updates personalized injection site rotation plans, injection time suggestions, and injection depth suggestions, and provides personalized health guidance plans for lipoma prevention, skin care, blood sugar management, diet and exercise.

[0140] 8.3 Users can view health reports, historical injection records, historical detection images and annotation information for different periods on the App, and can share the reports with the supervising medical staff with one click; medical staff can view all the data and reports of the bound patients on the medical staff terminal, add professional treatment guidance annotations, adjust the patient's insulin injection plan, and synchronize it to the patient's App terminal.

[0141] 8.4 Family members can view the patient's injection completion status, blood glucose data, and health reports in real time, and receive injection reminders and alarms for abnormal situations, enabling long-term remote monitoring of the patient, especially suitable for use scenarios of elderly people living alone and patients with limited mobility.

[0142] Intelligent analysis of long-term data can detect potential risks such as lipomatosis, improper injection, and poor blood sugar control in advance, allowing for timely intervention and reducing the occurrence of diabetic complications; personalized health management programs enable precise management of diabetes chronic diseases, improving the effectiveness of insulin therapy and patients' quality of life; multi-terminal collaboration constructs a complete closed loop for home-based diabetes management, improving the efficiency and accessibility of chronic disease management.

[0143] Therefore, this approach can be widely applied in scenarios such as home insulin injection for diabetic patients, community health service centers, hospital endocrinology departments, and rehabilitation departments. It can effectively improve the safety and standardization of insulin injection, reduce injection-related complications, and improve the blood glucose control of diabetic patients, thus possessing significant clinical value and social significance.

[0144] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A subcutaneous insulin detection pen, comprising a pen tip, characterized in that, The pen tip is equipped with a subcutaneous detection module, which is used for: Signal detection is performed on subcutaneous tissue / nodule abnormalities using non-contact detection technology. The analyzed skin condition data is then uploaded to a backend server to assess whether the skin at that location is suitable for injection.

2. An intelligent annotation system, characterized in that, Including the subcutaneous insulin probe pen, cloud-based intelligent analysis and annotation engine, and multi-terminal collaborative intelligent annotation terminal as described in claim 1, wherein: The integrated subcutaneous injection probe insulin pen has a built-in insulin pen basic execution unit, a high-frequency non-contact ultrasound subcutaneous detection module, an embedded edge computing and preprocessing unit, and a pen-end wireless communication module. It is used to realize non-contact in-situ detection of subcutaneous tissue, execution of insulin injection, edge preprocessing of echo signals, and injection safety interlock control. The needle adapter interface of the high-frequency non-contact ultrasound subcutaneous detection module and the insulin pen basic execution unit are designed coaxially. The cloud-based intelligent analysis and annotation engine adopts a distributed cloud-native architecture and communicates with multi-terminal collaborative intelligent annotation terminals. It has built-in subcutaneous tissue ultrasound image segmentation module, multi-dimensional feature extraction module, injection site suitability intelligent assessment module, and multi-terminal intelligent annotation generation module to complete intelligent analysis of ultrasound images, quantitative assessment of injection site suitability, intelligent annotation information generation, and multi-terminal data collaboration. The multi-terminal collaborative intelligent annotation terminal includes a patient-side intelligent terminal, a medical staff-side management terminal, and a family-side monitoring terminal, which are used to realize user-system interaction, visualization of detection results, multi-terminal remote collaborative monitoring and health management.

3. The intelligent annotation system according to claim 1, characterized in that, The high-frequency non-contact ultrasonic subcutaneous detection module includes a 32-element high-frequency PMUT transducer array, a high-voltage pulse transmitting circuit, a low-noise echo receiving circuit, a nanosecond-level timing control circuit, a miniature acoustic matching lens, and an infrared distance sensor. The 32-element high-frequency PMUT transducer array uses 20MHz-50MHz high-frequency piezoelectric micromechanical ultrasonic transducers, arranged in a ring around the needle adapter interface, with the needle located at the center of the array. A double-layer acoustic matching layer is set at the front end of the transducer array to adapt to the acoustic impedance difference between air and skin, achieving a coupling agent-free non-contact detection of 0.5mm-5mm.

4. The intelligent annotation system according to claim 1, characterized in that, The embedded edge computing and preprocessing unit includes an FPGA coprocessor chip, an MCU main control chip, a storage module, a power management module, and a safety interlock drive module. The safety interlock drive module includes a miniature electromagnetic lock, a drive circuit, and a status feedback sensor. The miniature electromagnetic lock is mechanically coupled to the injection trigger button of the insulin pen's basic execution unit. The MCU main control chip is electrically connected to the safety interlock drive module. It controls the electromagnetic lock to engage and unlock the injection trigger button only when the injection site is assessed as suitable for injection. When the site is assessed as unsuitable for injection, the electromagnetic lock remains permanently locked.

5. The intelligent annotation system according to claim 1, characterized in that, The subcutaneous tissue ultrasound image segmentation module of the cloud-based intelligent analysis and annotation engine incorporates an improved attention mechanism, the U-Net image segmentation network. This network adds a CBAM convolutional attention module to each convolutional block of the encoder and adopts a 4-layer downsampling + 4-layer upsampling encoder-decoder architecture to output pixel-level segmentation results for seven categories: epidermis, dermis, fat layer, muscle layer, nodule region, vascular region, and nerve region. Based on the segmentation results, the multi-dimensional feature extraction module extracts five core evaluation features: fat layer thickness, nodule region proportion, maximum vascular diameter, nerve distribution characteristics, and injection site frequency.

6. The intelligent annotation system according to claim 4, characterized in that, The intelligent assessment module for injection site suitability uses the analytic hierarchy process (AHP) combined with fuzzy comprehensive evaluation to construct a three-level assessment index system. It calculates the membership degree of each index to the three levels of suitable injection, cautious injection, and unsuitable injection through a preset membership function. The comprehensive evaluation result is obtained by weighted synthesis and the classification is completed according to the principle of maximum membership degree. The assessment result directly links the control logic of the pen tip safety interlock drive module, and there is no manual unlocking channel for the unsuitable injection level.

7. The intelligent annotation system according to claim 1, characterized in that, The multi-terminal collaborative intelligent annotation terminal connects the patient end, medical staff end, and family end via MQTT. 3.1.1 The protocol communicates with the cloud-based intelligent analysis and annotation engine. Based on the publish / subscribe model, an independent topic space is established for each user. Patients subscribe to all topics, medical staff subscribe to all topics bound to the patient, and family members only subscribe to monitoring-related topics, realizing hierarchical data push and real-time synchronization of information across multiple terminals.

8. The intelligent annotation system according to claim 1, characterized in that, The pen-end wireless communication module uses a Bluetooth 5.3 BLE low-power module to establish an AES-128 encrypted communication link with the patient-end smart terminal; the patient-end smart terminal and the cloud-based intelligent analysis and annotation engine adopt a dual-protocol architecture of HTTPS protocol + WebSocket protocol, where the HTTPS protocol is used for non-real-time large data transmission and the WebSocket protocol is used for real-time bidirectional communication.

9. The intelligent annotation system according to claim 4, characterized in that, The cloud-based intelligent analysis and annotation engine also includes a personalized injection site recommendation module, which calculates the recommendation priority of injection sites using the following formula: Symbol definition: The higher the value, the higher the recommendation level for the injection site. The site's anatomical structure was scored (0-100 points). The score for the site of abnormal tissue is 0-100. Historical injection scores for the site (0-100 points). The site-specific blood glucose response score (0-100 points); These are the weighting coefficients.

10. An application control method for an intelligent annotation system, applied to the intelligent annotation system according to any one of claims 2-9, characterized in that, Includes the following steps: System initialization and user pairing: Complete the full hardware self-test of the insulin pen, and the user completes account registration, real-name authentication and device pairing and binding through the patient-side smart annotation App, establishes a personal electronic health record and stores it in the cloud; Pre-injection preparation and detection activation: After the user completes the installation of the insulin pen cartridge, needle, and dosage adjustment, and aligns the insulin pen tip with the intended injection site on the skin, when the infrared distance sensor detects that the distance between the pen tip and the skin is within the effective detection range of 0.5mm-5mm, an operation reminder is issued. After the user triggers the detection command, the high-frequency non-contact ultrasound subcutaneous detection module is activated to complete the acquisition of subcutaneous tissue echo signals. Echo signal preprocessing and data transmission: The FPGA at the pen end performs parallel delay summation beamforming, Hilbert transform envelope detection, logarithmic compression, and digital filtering preprocessing on the acquired echo signal to generate standardized subcutaneous tissue ultrasound images. After compression and encryption, the images are transmitted to the patient's app via Bluetooth link and then uploaded to the cloud-based intelligent analysis and annotation engine. Cloud-based intelligent analysis and injection site suitability assessment: The cloud-based system completes pixel-level segmentation of subcutaneous tissue using an improved U-Net segmentation network. Based on the segmentation results, five core assessment features are extracted. The system then uses the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method to complete a three-level graded assessment of injection site suitability, generating assessment results and corresponding safety interlock control commands. Intelligent annotation generation and multi-terminal synchronization: Based on the evaluation results, feature values ​​and segmentation results, the cloud generates standardized text annotations and visual annotation images. At the same time, it generates injection site recommendation schemes through personalized recommendation algorithms and synchronizes relevant information to the patient, medical staff and family members through WebSocket and MQTT protocols. Safety interlock control and injection permission management: The patient-side App encrypts and transmits the safety interlock control commands sent from the cloud to the insulin pen MCU. The MCU drives the safety interlock module to perform graded unlocking actions according to the commands: the appropriate injection level unlocks the injection permission directly, the cautious injection level requires the user to manually confirm the risk before unlocking, and the inappropriate injection level remains permanently locked and the manual unlocking channel is blocked. Injection execution and data recording and archiving: After the user unlocks injection privileges, the MCU completes the insulin injection. The MCU uses sensors to detect the injection dosage and execution status in real time, automatically records all data of this injection, synchronizes it to the App and the cloud, and archives it to the user's personal health record. Long-term data management and personalized health management: The cloud performs statistical analysis on the user's full detection, injection, and blood glucose data according to a preset cycle, generates intelligent annotated health reports and personalized health management plans, and synchronizes them to various terminals to realize full-cycle health management for home injections for diabetes.

Images