Puncture difficulty grading system and method fusing ultrasonic image and 3CG signal

Abstract

The invention discloses a puncture difficulty grading system and method fusing an ultrasonic image and a 3CG signal, and relates to the technical field of puncture difficulty grading machine learning. The puncture difficulty grading system fusing the ultrasonic image and the 3CG signal comprises a data quality evaluation and compensation triggering module; a multi-modal feature extraction and fusion judgment module; and a puncture difficulty grading execution module. According to the method, whether ultrasonic image adaptive compensation and signal adaptive compensation are triggered or not is determined by executing puncture reference data quality evaluation, then multi-modal feature extraction and multi-modal feature fusion are executed, and finally puncture difficulty grading is executed, so that the effect of improving the puncture difficulty grading accuracy is achieved; the problem of low puncture difficulty grading accuracy in the prior art is solved.

Description

A grading system and method for puncture difficulty that integrates ultrasound images and 3CG signals. Technical Field

[0001] This invention relates to the field of machine learning technology for puncture difficulty grading, and in particular to a puncture difficulty grading system and method that integrates ultrasound images and 3CG signals. Background Technology

[0002] The precise placement of the PICC (Peripherally Inserted Central Catheter) tip is crucial for the safety and efficacy of clinical treatment and has become a focus of research both domestically and internationally. Currently, the domestically accepted "gold standard" of chest X-ray placement suffers from diagnostic lag and radiation hazards. While intracardiac electrocardiography (ICG) placement is listed as a clinical standard internationally, it requires operators with strong ECG interpretation skills, hindering its widespread adoption in China. 3CG (3-Channel Intracardiac) placement... Electrogram (three-lead intracardiac electrogram) – an emerging real-time positioning technology in China, can significantly improve the success rate of tip placement on the first attempt, while shortening infusion delays and reducing radiation exposure and complication risks. However, the difficulty grading assessment of clinical puncture procedures such as PICC placement remains a key challenge: traditional puncture difficulty grading relies heavily on single ultrasound image information, which can only reflect static features such as lesion anatomy and morphology, making it difficult to capture dynamic information closely related to puncture difficulty, such as tissue physiological activity and hemodynamics. This can easily lead to grading errors, thereby increasing the risk of puncture operation (such as accidental injury to blood vessels, puncture failure, etc.). 3CG signals, however, can improve the accuracy of grading. The dynamic physiological features of the target area are effectively complemented by the anatomical features of the ultrasound image. Therefore, by fusing information from both modalities to construct a puncture difficulty grading scheme, the limitations of a single modality can be overcome, comprehensively improving the accuracy and reliability of the grading. In the construction and application of the puncture difficulty grading model, firstly, ultrasound image data and corresponding 3CG physiological signals of the target area are simultaneously acquired. The ultrasound image undergoes image preprocessing (such as denoising, enhancement, and region of interest segmentation), while the 3CG signal undergoes signal preprocessing (such as filtering and baseline correction) to remove interference information. Then, based on CNN (Convolutional Neural Networks), the puncture difficulty grading scheme is constructed. The convolutional neural network (CNN) model extracts visual features (such as lesion morphology, boundary clarity, and surrounding blood vessel distribution) from the preprocessed ultrasound image. Based on the LSTM (Long Short-Term Memory) network model (including a signal segment quality assessment module), physiological parameters (such as peak intensity, waveform rhythm, and time-domain and frequency-domain features) are extracted from the 3CG signal. Subsequently, a feature fusion algorithm is used to effectively integrate the two types of features to construct a puncture difficulty representation vector containing multi-dimensional information. This feature vector is then input into a pre-trained puncture difficulty grading model (such as random forest or deep learning model). The model infers and outputs the puncture success probability, ultimately achieving accurate puncture difficulty grading by fusing ultrasound images and 3CG signals.

[0003] The aforementioned technology has been found to have at least the following technical problems: Current ultrasound equipment for vascular puncture assessment primarily relies on manual judgment of key vascular parameters (such as arterial / venous type and vessel diameter), and subjective selection of the optimal puncture vessel based on these parameters. This subjective judgment-based approach lacks quantitative fusion analysis of multi-dimensional parameters (such as vascular tortuosity and interference from surrounding tissues) when dealing with complex cases (such as thin or deep vessels). This increases the risk of puncture failure and complications, and the efficiency of puncture difficulty grading is low. Furthermore, interference from body surface parameters, such as a thicker subcutaneous fat layer, may occur in some obese patients (body mass index greater than 30). Ultrasound signal attenuation, blurred vascular boundaries, and weakened 3CG signal features due to increased tissue resistance (such as weakened electromagnetic positioning signals and reduced P-wave feature recognition) exacerbate the difficulty of extracting key features from images and signals, potentially leading to distortion in key feature extraction. Consequently, when performing multimodal data fusion based on distorted key features of images and signals, the fused puncture difficulty representation vector may become disconnected from actual vascular conditions. This results in the puncture difficulty level being mismatched with the actual operational difficulty when classifying puncture difficulty levels based on a pre-set puncture difficulty grading model (such as random forest or deep learning models), ultimately leading to low accuracy in puncture difficulty grading. Summary of the Invention

[0004] To address the problem of low accuracy in puncture difficulty grading in existing technologies, this invention provides a puncture difficulty grading system and method that integrates ultrasound images and 3CG signals. The technical solution is as follows:

[0005] On one hand, a puncture difficulty grading system integrating ultrasound images and 3CG signals is provided, including: a data quality assessment and compensation triggering module, a multimodal feature extraction and fusion judgment module, and a puncture difficulty grading execution module. The data quality assessment and compensation triggering module performs a puncture baseline data quality assessment and, based on the acquired puncture baseline data quality assessment results, determines whether to trigger adaptive compensation for ultrasound images to correct image blurring and adaptive compensation for 3CG signal weakening. The multimodal feature extraction and fusion judgment module, after the puncture baseline data quality assessment is qualified, performs multimodal feature extraction to capture ultrasound image features and 3CG signal electrophysiological features, and, based on the acquired multimodal feature extraction results, determines whether to perform multimodal feature fusion. The puncture difficulty grading execution module, after the multimodal feature fusion is completed, performs a puncture difficulty grading to quantify the complexity of the puncture operation.

[0006] On the other hand, a method for grading puncture difficulty by fusing ultrasound images and 3CG signals is provided. This method is applied to a puncture difficulty grading system that fuses ultrasound images and 3CG signals. The method includes: S1, performing a puncture baseline data quality assessment, and determining whether to trigger adaptive compensation of ultrasound images to correct image blurring and adaptive compensation of 3CG signals to correct 3CG signal weakening based on the obtained puncture baseline data quality assessment results; S2, after the puncture baseline data quality assessment is qualified, performing multimodal feature extraction of ultrasound image features and 3CG signal electrophysiological features to capture factors affecting puncture difficulty, and determining whether to perform multimodal feature fusion based on the obtained multimodal feature extraction results; S3, after the multimodal feature fusion is completed, performing puncture difficulty grading to quantify the complexity of the puncture operation.

[0007] The beneficial effects of the technical solution provided by the embodiments of the present invention include at least the following: 1. By performing a quality assessment of the puncture baseline data and determining whether to trigger adaptive compensation of ultrasound images and 3CG signals based on the obtained quality assessment results, it helps to filter out low-quality ultrasound image data and interference from 3CG signal data at the source, providing qualified data for subsequent feature extraction and fusion, reducing data distortion caused by the lack of data processing for obese individuals, and thus reducing the input bias of the grading model, thereby ensuring the reliability of the entire solution's decision-making from the bottom up; after the puncture baseline data quality assessment is qualified, multimodal feature extraction is performed, and the determination of whether to proceed is based on the obtained feature confidence results. Multimodal feature fusion helps to overcome the limitations of existing technologies where a single ultrasound image can only reflect static information and a single 3CG signal can only reflect dynamic physiological information. By integrating the complementary advantages of the two types of features, a more comprehensive feature representation of puncture difficulty can be constructed. Compared with the single-modal assessment or simple feature splicing of existing technologies, it can significantly improve the adaptability of features to complex puncture scenarios. After the multimodal feature fusion is completed, puncture difficulty grading is performed to quantify the complexity of the puncture operation. This helps to replace the traditional subjective judgment that relies on experience, reduce the assessment bias caused by individual experience differences, and thus guide the operator to select the appropriate operation plan according to the grading results, thereby improving the puncture success rate and reducing the risk of complications.

[0008] 2. Existing technologies lack a systematic quality assessment scheme for ultrasound images and 3CG signals, relying mostly on single quality indicators or subjective judgments. This fails to comprehensively cover core quality dimensions such as the degree of separation between vascular signals and noise, the ability to distinguish vessel wall boundaries, and the recognizability of P-wave features. Furthermore, ultrasound image artifacts and 3CG signal noise can easily lead to misjudgments of data quality, resulting in low-quality data being directly incorporated into subsequent feature extraction and grading processes, causing deviations in the puncture difficulty grading. This proposed solution, by specifically selecting ultrasound image signal-to-noise ratio and vessel wall edge clarity, as well as 3CG signal-to-noise ratio and P-wave recognizability, comprehensively captures the core quality characteristics of both types of data, helping to address the one-sidedness of existing technologies that rely on single-indicator assessments. By weighting and coupling ultrasound image quality parameters and harmonic averaging 3CG signal quality parameters, this method comprehensively considers the impact weight of each quality dimension on the overall data quality, avoiding bias caused by single parameter evaluation. This makes the data quality quantification results more objective in reflecting the true reliability of the data. Based on the data quality quantification value, an adaptive compensation mechanism is triggered, which can accurately identify and optimize substandard conditions such as ultrasound image blurring and 3CG signal distortion. Compared with the current situation where existing technologies lack data preprocessing verification, this method removes low-quality data interference from the source, providing a standardized and high-quality data foundation for subsequent multimodal feature extraction and fusion, and reducing the puncture difficulty classification error caused by data quality issues.

[0009] 3. In special scenarios involving obese patients (body mass index greater than 30) with vascular anatomical variations or a history of multiple chemotherapy treatments (this scenario simultaneously presents the complexities of body fat interference and deteriorating vascular conditions, making conventional feature assessments inadequate), targeted feature accuracy assessments are required. However, existing technologies for feature accuracy assessment in puncture difficulty grading generally suffer from a single assessment dimension and a lack of weighted differentiation, resulting in an inability to comprehensively cover measurement biases in ultrasound image parameters (such as vessel diameter, depth, and tortuosity) and capture biases in unfused 3CG signals (such as P-wave amplitude). This leads to one-sided and distorted feature accuracy assessment results, failing to reflect the true quality of feature measurements and introducing potential errors into subsequent puncture difficulty grading. This proposed solution addresses these issues. By selectively acquiring quantitative parameters of accuracy deviation for four types of core features, and then weighting and coupling these parameters with parameters influencing them, we can use the results as a feature deviation evaluation index. This approach achieves comprehensive coverage of key multimodal feature deviations in ultrasound images and 3CG signals. Furthermore, by assigning weights, we highlight the actual differences in the impact of different feature deviations. This effectively compensates for the shortcomings of existing technologies, such as incomplete evaluation dimensions and lack of targeted weight design. As a result, the feature deviation evaluation results are more aligned with the actual needs of clinical puncture, accurately quantifying the true degree of deviation in multimodal feature measurements. This provides a high-quality and reliable feature foundation for subsequent multimodal feature fusion and puncture difficulty grading, reducing the risk of grading deviations caused by inaccurate feature accuracy assessments. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 is a schematic diagram of the puncture difficulty grading system fused with ultrasound images and 3CG signals provided in an embodiment of the present invention; Figure 2 is a flowchart of the general overview of the puncture difficulty grading method fused with ultrasound images and 3CG signals provided in an embodiment of the present invention; Figure 3 is a flowchart of the puncture difficulty grading method fused with ultrasound images and 3CG signals provided in an embodiment of the present invention; Figure 4 is a random forest ensemble architecture diagram of the puncture difficulty grading model of the puncture difficulty grading method fused with ultrasound images and 3CG signals provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0013] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0014] Example 1 provides a puncture difficulty grading system that integrates ultrasound images and 3CG signals. As shown in Figure 1, the system includes a data quality assessment and compensation triggering module, a multimodal feature extraction and fusion judgment module, and a puncture difficulty grading execution module. The data quality assessment and compensation triggering module performs a puncture baseline data quality assessment in a specified puncture difficulty grading scenario and determines whether to trigger adaptive ultrasound image compensation for correcting image blurring caused by obesity and tissue interference, and adaptive 3CG signal compensation for correcting 3CG signal weakening, based on the acquired baseline data quality assessment results. By monitoring the baseline data quality assessment results, it helps to reduce feature extraction bias caused by puncture baseline data distortion from the source, providing a high-quality ultrasound image and 3CG signal foundation for subsequent multimodal analysis.

[0015] The multimodal feature extraction and fusion judgment module is used to perform multimodal feature extraction of ultrasound image features and 3CG signal electrophysiological features to capture factors affecting puncture difficulty after the puncture baseline data quality assessment is qualified. Based on the obtained multimodal feature extraction results, it decides whether to perform multimodal feature fusion. By performing multimodal feature extraction and multimodal feature fusion, it helps to overcome the limitations of existing technologies that rely on single modal information or simple feature splicing. By integrating the complementarity of ultrasound image spatial features and 3CG physiological function features, a more accurate puncture difficulty feature system can be constructed.

[0016] The puncture difficulty grading module is used to perform puncture difficulty grading to quantify the complexity of puncture procedures after multimodal feature fusion. By performing puncture difficulty grading, it helps to provide an objective and quantitative difficulty reference standard for clinical procedures, reduce assessment bias caused by individual experience differences, and improve puncture safety and efficiency.

[0017] It should be understood that the puncture difficulty grading system integrating ultrasound images and 3CG signals provided in this solution has a dedicated database built during the R&D phase to store various core data. The data sources of this database include basic preset parameters configured by technicians, such as initial signal confidence reference values, initial image confidence reference values, and initial first quality reference values, as well as organized historical clinical case data (such as ultrasound images, 3CG signal records, and corresponding puncture result data of different patient groups) and model development experimental data (such as parameter adaptation results in multi-round feature extraction and fusion algorithm verification), providing practical support for the scientific validity and clinical adaptability of the preset parameters. The database adopts an encrypted distributed storage architecture, storing structured parameters (such as the specific values ​​of quality assessment thresholds and the range of feature fusion weights) through a relational database, while using a non-relational database to store unstructured data (such as original ultrasound image frames and 3CG signal waveform data). Clinical technicians can dynamically optimize and calibrate the preset values ​​in the database based on newly added clinical puncture data and quality feedback, so that the stored data can continuously adapt to the puncture scenario needs of different patients (such as obese patients and patients with vascular variations).

[0018] In this embodiment, the data quality assessment and compensation triggering module, the multimodal feature extraction and fusion judgment module, and the puncture difficulty grading execution module help to systematically solve the core pain points of traditional puncture difficulty assessment, which relies on subjective experience and is limited by single-modal data. This provides accurate and objective puncture difficulty grading criteria for clinical practice. Specifically, the data quality assessment and compensation triggering module, as a core pre-processing step, outputs qualified and compensated ultrasound images and 3CG signals, which directly determine the feature capture accuracy of the multimodal feature extraction and fusion judgment module. If the quality of the preceding data is substandard and not compensated, it will lead to distortion in subsequent feature extraction. The multimodal feature extraction and fusion judgment module provides the core input for the puncture difficulty grading execution module. The high-confidence fusion features it extracts are the key to the reliability of the grading results. The completeness and accuracy of feature fusion directly affect the accuracy of the grading quantification results, forming a closed-loop adaptation mechanism from data preprocessing to feature processing and then to puncture difficulty grading application.

[0019] Figure 2 shows a flowchart outlining the puncture difficulty grading method that fuses ultrasound images and 3CG signals. This flowchart illustrates the overall process of grading puncture difficulty using multimodal data from fused ultrasound images and 3CG signals as input, through data quality assessment and adaptive compensation, multimodal feature extraction and fusion, and finally completing the puncture difficulty grading. As shown in Figure 2, in a specified puncture difficulty grading scenario, a puncture baseline data quality assessment is performed, and a first quality quantization value and a second quality quantization value are obtained. It is determined whether the first quality quantization value is less than the initially set first quality reference value. If so, adaptive compensation of the ultrasound image is triggered; otherwise, multimodal feature extraction is performed. It is then determined whether the second quality quantization value is less than the initially set second quality reference value. If so, adaptive compensation of the 3CG signal is triggered; otherwise, multimodal feature extraction is performed. Adaptive compensation representation of the ultrasound image is performed, followed by frequency domain analysis and noise spectrum estimation to obtain the main signal frequency band and residual frequency band of the image. Adaptive adjustment of the main signal gain is performed for the main signal frequency band, and residual noise suppression is performed for the residual frequency band. The ultrasound image adaptive compensation is then performed. After compensation is completed, the first quality quantization value is checked for compliance. If it is, multimodal feature extraction is performed; otherwise, an image quality adjustment failure warning is sent. Adaptive compensation of the 3CG signal is performed, and the signal matching quantization value is obtained. It is then checked whether the signal matching quantization value is greater than the initial matching degree reference value. If it is, adaptive Wiener filtering is performed, followed by waveform repair. After waveform repair, adaptive Wiener filtering is performed again. After adaptive compensation of the 3CG signal is completed, the second quality quantization value is checked for compliance. If it is, multimodal feature extraction is performed; otherwise, a 3CG signal adaptive compensation failure warning is sent. Multimodal feature extraction is performed, and image feature confidence and signal feature confidence are obtained. It is then checked whether the image feature confidence is greater than the initial image confidence reference value. If it is, multimodal feature fusion is performed; otherwise, an image feature anomaly warning is sent. Finally, it is checked whether the signal feature confidence is greater than the initial signal confidence reference value. If not, a signal feature anomaly warning is sent; otherwise, multimodal feature fusion is performed. After multimodal feature fusion is completed, the puncture difficulty is graded.

[0020] Preferably, the specific process for puncture baseline data quality assessment is as follows: S101, Obtain puncture baseline data quality assessment parameters, specifically as follows: For ultrasound image data, define corresponding first puncture baseline data quality assessment parameters, including ultrasound image signal-to-noise ratio (SNR) to reflect the degree of separation between effective vascular signals and background noise in ultrasound images, and vascular wall edge sharpness to reflect the degree of distinction between the vascular wall and the surrounding tissue boundary. The specific definition process is as follows: The average of the ratios of the standard deviations of pixel values ​​in the vascular region to the standard deviations of pixel values ​​in the background region in each frame of the acquired images within the puncture baseline data quality assessment time period is expressed as ultrasound image SNR; ultrasound image data represents the data collected by ultrasound diagnostic equipment covering the target puncture vessel ( Color ultrasound images of peripheral veins, central veins, and surrounding tissues (such as fat, muscle, and fascia); the vascular region represents the internal region of the vascular cross-section identified based on image segmentation networks (such as U-Net); the vascular wall edge sharpness is represented by quantifying the ratio of the vascular wall edge sharpness index to the initial vascular wall edge sharpness reference value within the puncture baseline data quality assessment period. The vascular wall edge sharpness index is represented by the average gradient magnitude of the vascular wall boundary region in each acquired frame of image; the average gradient magnitude represents the result of obtaining the image gradient at the vascular wall boundary based on the Sobel operator and performing an arithmetic mean operation on the gradient magnitude of all boundary pixels; for 3CG signal data The corresponding second puncture baseline data quality assessment parameters are defined, including the 3CG signal-to-noise ratio (SNR), which quantifies the proportion of effective physiological signals in the 3CG signal, and the P-wave discernibility, which quantifies the discernibility of P-wave features (such as peak value and morphology) in the 3CG signal. The specific definition process is as follows: The 3CG signal-to-noise ratio is expressed as the average of the ratio of the P-wave signal power monitored by the 3CG intracardiac electrocardiogram (ICG) monitoring instrument to the baseline noise signal power for each cardiac cycle acquired during the puncture baseline data quality assessment period. The 3CG signal data represents the continuous physiological electrical signal data reflecting cardiac electrical activity monitored by the 3CG ICG monitoring instrument, with core components including catheter tip positioning and vascular function status. The characteristic wave segments such as the P wave and QRS complex are considered; the baseline noise segment signal power is represented by the average of the amplitude variance of the signal obtained during the stable TP (i.e., end-diastolic phase) period of the historical time period; the cardiac cycle is the time interval from the start point of one QRS complex in the intracardiac electrocardiogram to the start point of the next adjacent QRS complex. This interval covers a complete atrial and ventricular systole and diastole physiological process, including the P wave, QRS complex, T wave and the intervals of each wave segment, and is the basic periodic unit that recurs in the electrocardiogram signal; the P wave discernibility is represented by the average of the ratio of the P wave amplitude to the amplitude of the adjacent QRS complex in each cardiac cycle obtained during the puncture baseline data quality assessment period.Adjacent QRS complexes represent the QRS complexes within the cardiac cycle to which the current P wave belongs. The amplitude of adjacent QRS complexes is represented by the vertical distance between the R wave apex and the PR segment baseline. The puncture baseline data quality assessment time period includes multiple frames of images and a preset number of cardiac cycles. S102: After weighted coupling processing of the first puncture baseline data quality assessment parameters and the corresponding first quality weight parameters, a first quality quantification value for evaluating the overall quality level of ultrasound image data is obtained. The weighted coupling processing means multiplying each of the first puncture baseline data quality assessment parameters (ultrasound image signal-to-noise ratio, vessel wall edge sharpness) with the corresponding first quality weight parameters (image signal-to-noise ratio weight coefficient, vessel wall sharpness weight coefficient), and then summing all the product results to integrate the influence of different quality dimensions on the overall quality of ultrasound images and form a comprehensive quantification index. After harmonic averaging of the second puncture baseline data quality assessment parameters, a value for evaluating 3CG signal is obtained. According to the second quality quantification value of the overall quality level; the first quality weight parameter includes the image signal-to-noise ratio weight coefficient, which reflects the influence of the ultrasound image signal-to-noise ratio on the first quality quantification value, and the vessel wall clarity weight coefficient, which reflects the influence of the vessel wall edge clarity on the first quality quantification value; S103, the adaptive compensation mechanism is triggered, and the specific process is as follows: if the first quality quantification value is less than the initially set first quality reference value, the ultrasound image quality is determined to be substandard, the ultrasound image is marked as an ultrasound image to be compensated, and the ultrasound image adaptive compensation is triggered; otherwise, multimodal feature extraction is performed, wherein the initially set first quality reference value is represented by the average value of the first quality quantification value over a historical time period; if the second quality quantification value is less than the initially set second quality reference value, the 3CG signal quality is determined to be substandard, the 3CG signal is marked as a 3CG signal to be compensated, and the 3CG signal adaptive compensation is triggered; otherwise, multimodal feature extraction is performed, wherein the initially set second quality reference value is represented by the average value of the second quality quantification value over a historical time period.

[0021] It is important to note that both the baseline data quality assessment and the feature accuracy deviation assessment have a clear weighting and quantification system. The first quality weighting parameter includes the image signal-to-noise ratio weighting coefficient and the vessel wall clarity weighting coefficient. The feature accuracy deviation impact parameters cover the vessel diameter deviation impact factor, vessel depth deviation impact factor, vessel tortuosity deviation impact factor, and P wave amplitude deviation impact factor. These weighting parameters are pre-calibrated by technical personnel and stored in a dedicated database, providing the core basis for the calculation of the first quality quantification value, the comprehensive assessment of feature accuracy deviation, and the accurate matching of weight values. Furthermore, all weighting parameters use a numerical range of 0 to 1 to quantify the impact proportion, ensuring the operability of the assessment process and the reliability of the results.

[0022] Specifically, the process begins by extensively collecting historical clinical puncture data, encompassing ultrasound image and 3CG signal characteristics from different patient groups (including special cases such as obesity and vascular variations), along with the actual values ​​of various deviation parameters, initial weighting factors, and feature accuracy verification standards. Simultaneously, the complete matching combination of weighting parameters and evaluation results under each historical puncture scenario is recorded. Then, targeted weight quantification values ​​are assigned based on the actual impact of each parameter on the puncture difficulty assessment results. For example, when vessel diameter deviation directly affects the puncture needle positioning accuracy and significantly interferes with feature accuracy, a higher weight is assigned to the vessel diameter deviation influence factor. High weighting is used to enhance the proportion of weight parameters in feature accuracy assessment, and effective application records of weight parameters in various clinical scenarios are retained simultaneously. Subsequently, correlation analysis (such as Pearson correlation coefficient) is used to eliminate abnormal correlation data caused by temporary calibration deviations of monitoring equipment (such as drift of ultrasound diagnostic instrument parameters, instantaneous error of 3CG signal monitoring instrument) and temporary physiological fluctuations of patients (such as sudden changes in heart rate, local tissue congestion), and to screen out the correspondence of weight parameters with statistical stability. Finally, all effective clinical data are integrated to form a weight quantification system adapted to different puncture scenarios, providing data support for the scientificity and reliability of the assessment process.

[0023] In this embodiment, the assessment of puncture baseline data quality helps to identify data defects such as blurred ultrasound images and weakened 3CG signals from the source, and specifically reduces the negative impact of factors such as obesity and tissue interference on bimodal data. This reduces the risk of multimodal feature extraction deviation and inaccurate puncture difficulty grading results caused by data distortion, improves the reliability of puncture baseline data and the accuracy of subsequent feature processing, and realizes the quantitative evaluation and standardized screening of ultrasound image and 3CG signal quality. This lays a solid core data foundation for the scientific nature and clinical suitability of multimodal feature fusion and final puncture difficulty grading.

[0024] Preferably, the specific compensation process for adaptive compensation of ultrasound images is as follows: S201, performing frequency domain analysis and noise spectrum estimation, the specific process is as follows: based on two-dimensional fast Fourier transform, the ultrasound image to be compensated is transformed from the spatial domain to the frequency domain to obtain the corresponding image frequency domain spectrum; based on peak detection and background fitting algorithms, the image frequency domain spectrum is separated into the image main signal frequency band representing effective tissue structure and the image residual frequency band representing noise, the specific separation process is as follows: the peak frequency cluster with significantly higher energy than the surrounding area in the frequency domain spectrum is located by the peak detection algorithm (corresponding to the effective tissue structure signal), and at the same time, based on the background fitting algorithm, the frequency range covered by the peak cluster is defined as the image main signal frequency band, and the remaining frequency range is defined as the image residual frequency band; S202, performing corresponding adaptive adjustment of the image main signal gain for the image main signal frequency band; the adaptive adjustment of the main signal gain means taking the first quality quantization value and the average energy value of the image main signal frequency band as input parameters, querying the preset signal gain mapping table, and obtaining the corresponding value for enhancing the effective image. The main signal gain ratio coefficient is used as the adjustment amount. The amplitude gain of the main signal band of the image is gradually increased in the direction of gain increase (after each adjustment of the amplitude gain of the main signal band of the image, the first quality quantization value is re-acquired. If the first quality quantization value is less than the initial first quality reference value, the amplitude gain after this adjustment is used as the initial value for the next adjustment and the amplitude gain of the main signal band of the image is adjusted step by step in the direction of gain increase). This helps to reduce the oversaturation of the effective signal or loss of detail caused by a single large gain, and ensures that the adjustment amplitude is accurately controlled while enhancing the effective tissue structure signal. The average energy value of the main signal band of the image is represented by summing the amplitude values ​​of all frequency points monitored by the ultrasound diagnostic monitoring instrument in the main signal band of the image, and then quantifying the ratio with the total number of frequency points in the band monitored by the monitoring instrument. The average energy value of the main signal band of the image is used to quantify the overall intensity of the effective tissue structure signal of the ultrasound image. The ratio quantization represents the ratio calculation.

[0025] S203. Perform corresponding image residual noise suppression for the image residual frequency band; image residual noise suppression means taking the first quality quantization value and the average energy value of the image residual frequency band as input parameters, querying the preset noise suppression mapping table, and obtaining the corresponding noise suppression coefficient for suppressing background noise; using the noise suppression coefficient as the adjustment amount, gradually reduce the amplitude gain of the residual noise frequency band in the direction of gain decrease (after each adjustment of the amplitude gain of the residual noise frequency band, the first quality quantization value is re-acquired; if the first quality quantization value is less than the initial first quality reference value, the amplitude gain after this adjustment is used as the initial value for the next adjustment and the amplitude gain of the residual noise frequency band is adjusted step by step in the direction of gain decrease), which helps to reduce the false elimination of effective edge signals by single strong suppression, accurately weaken background noise while maximizing the preservation of effective structural information, and steadily improve the image signal-to-noise ratio and structural clarity; the average energy value of the image residual frequency band is the result of summing the amplitude values ​​of all frequency points in the image residual frequency band and then quantizing it by the proportion of the total number of frequency points in the band. The average energy value of the residual frequency band in the image is used to quantify the overall intensity of background noise. S204: Continuously monitor the first quality quantization value. When the first quality quantization value is not less than the initially set first quality reference value, image edge optimization is performed to obtain a qualified ultrasound image. Multimodal feature extraction is then performed based on the qualified ultrasound image. Otherwise, adaptive adjustment of the main signal gain and residual noise suppression continue. When the initially set maximum number of iterations for adaptive compensation of the ultrasound image is reached, if the first quality quantization value is still less than the initially set first quality reference value, an image quality adjustment failure prompt is sent. The initially set maximum number of iterations is preset by a designated person. The specific process of image edge optimization is as follows: Based on two-dimensional inverse Fourier transform, the main signal frequency band and the residual frequency band of the image are synthesized in the frequency domain to obtain a preliminary enhanced image where residual noise is selectively suppressed and the main signal is preserved to the maximum extent. Based on a multi-scale morphological gradient algorithm, the preliminary enhanced image undergoes edge sharpening and structure preservation processing to further enhance the edge contrast of the blood vessel wall, thereby generating a qualified ultrasound image.

[0026] It is important to note that the preset signal gain mapping table and preset noise suppression mapping table used in this solution for adaptive compensation of ultrasound images are both calibrated in advance by technicians using clinical data and stored in a dedicated database. This provides core support for accurately obtaining the main signal gain ratio coefficient and noise suppression coefficient. When the system performs ultrasound image enhancement and noise suppression operations, it can quickly match the adaptation coefficients from the corresponding mapping tables to ensure that the amplitude of main signal gain adjustment and noise suppression precisely matches the optimization needs of ultrasound images of different quality levels. This avoids the problem of insufficient gain leading to unclear effective vascular features or excessive suppression causing loss of vascular structural details.

[0027] Specifically, the construction of these two types of mapping tables relies on a large amount of effective ultrasound image processing data in clinical scenarios, covering different patient body types (such as obesity level, subcutaneous fat thickness), tissue interference intensity, and parameter combinations under the original image quality level (including the combination of the first quality quantization value and the average energy value of the main signal band of the image and the corresponding main signal gain ratio coefficient, the combination of the first quality quantization value and the average energy value of the residual band of the image and the corresponding noise suppression coefficient). Quantitative weights are assigned according to the degree of influence of the parameter combination on the image optimization effect—for example, when the first quality quantization value is significantly lower than the preset standard and the main signal energy is weak, a larger main signal gain ratio coefficient is matched to enhance the vascular structure signal. At the same time, the actual effective values ​​of the two types of coefficients in each clinical scenario are recorded, and abnormal correlation data caused by instantaneous calibration deviation of ultrasound equipment, temporary local tissue congestion of patients, etc., are eliminated through correlation analysis (such as Pearson correlation coefficient). Finally, the parameter correspondence with statistical stability is retained to ensure the adaptability and reliability of the mapping table.

[0028] In this embodiment, adaptive compensation of ultrasound images helps to address the core technical problems of blurred ultrasound images and blurred vascular structure boundaries caused by thick subcutaneous fat layers and tissue scattering interference in obese patients. By synergistically adjusting the main signal gain enhancement and residual noise suppression, effective vascular features are highlighted and background interference is weakened. This reduces the risk of vascular structure measurement errors and subsequent feature extraction distortion caused by substandard image quality. It improves the clarity and recognition accuracy of key anatomical information such as vascular diameter and wall edge regularity in ultrasound images of obese patients. It realizes the precise optimization of low-quality ultrasound images of obese patients into clinically usable images, and provides reliable image data support for multimodal feature fusion and puncture difficulty classification.

[0029] Preferably, the specific process of adaptive compensation for 3CG signals is as follows: S301, continuously collect 3CG signal data within a historical time period, and identify the complete cardiac cycle based on the R-wave peak detection algorithm, obtain the median amplitude value of each corresponding sampling point in the cardiac cycle waveform, and use the synthesized standard waveform as the initial 3CG reference signal; perform signal segmentation on the 3CG signal to be compensated based on the cardiac cycle to obtain segmented 3CG signals; perform alignment operation on each segmented 3CG signal and the initial 3CG reference signal based on the dynamic time planning algorithm, and obtain the minimum cumulative distance of the alignment path; sum the obtained minimum cumulative distance of the alignment path with the preset alignment constant, and then perform the reciprocal operation to represent the signal matching quantification value used to reflect the qualification of the segmented 3CG signal to be compensated, wherein the minimum cumulative distance of the alignment path represents the result obtained by finding the segmented segment in the dynamic time planning algorithm. The optimal alignment path between the 3CG signal and the initial 3CG reference signal (i.e., the best matching trajectory of the two signal waveforms on the time axis) is calculated, and the minimum total deviation value is obtained by accumulating the differences (such as amplitude deviation) of the signal values ​​at all corresponding times along this path. The preset alignment constant is set in advance by preset personnel to avoid the meaninglessness of the reciprocal calculation when the minimum cumulative distance of the alignment path is 0. S302, Signal difference compensation is performed based on the signal matching quantization value. The specific process is as follows: For segmented 3CG signals with a signal matching quantization value greater than the initial matching degree reference value, adaptive Wiener filtering is performed, where the initial matching degree reference value is represented by the average value of the signal matching quantization value over a historical time period; For segmented 3CG signals with a signal matching quantization value not greater than the initial matching degree reference value, waveform repair is performed. After the waveform repair is completed, adaptive Wiener filtering is performed again based on the obtained repaired 3CG signal.

[0030] Specifically, the waveform restoration process is as follows: Identify the maximum restoration interval between the segmented 3CG signal and the initial 3CG reference signal. The identification process is as follows: Calculate the amplitude difference between the signal segment and the initial 3CG reference signal within the P-wave time window by sampling point by sampling point. Determine the interval of consecutive sampling points with the largest cumulative difference as the maximum restoration interval. The amplitude difference is represented by the result of the absolute difference calculation between the amplitude values ​​of the sampling points in the signal segment and the amplitude values ​​of the corresponding sampling points in the initial 3CG reference signal. The cumulative difference is represented by the result of the summation of the amplitude differences of a preset number of consecutive sampling points within the P-wave time window. The preset number of sampling points is set in advance by a preset team. Based on the maximum restoration interval, weighted fusion is performed on each sampling point value within the maximum restoration interval of the original signal and the corresponding sampling point value in the initial 3CG reference signal to restore the 3CG signal waveform of the maximum restoration interval. The restored maximum restoration interval is then re-embedded into the original signal segment. The corresponding position of the segment is obtained by smoothing and filtering the signal connection point using a linear interpolation algorithm. The embedded means that the original data of the corresponding interval in the original signal segment is replaced by the data of the maximum repair interval after repair, so as to maintain the integrity of the time axis of the signal segment and the continuous arrangement of the data of the non-repair interval. The signal connection point means the connection position between the starting point of the maximum repair interval after repair and the ending point of the unrepaired data before the interval in the original signal segment, and the ending point of the repair interval and the starting point of the unrepaired data after the interval in the original signal segment. After the adaptive compensation of the 3CG signal is completed in S303, it is determined whether the re-acquired second quality quantization value is less than the initial second quality reference value. If so, a 3CG signal adaptive compensation failure prompt is sent. Otherwise, the 3CG signal after adaptive compensation is marked as a qualified 3CG signal, and multimodal feature extraction is performed based on the qualified 3CG signal.

[0031] In this embodiment, adaptive compensation of 3CG signals helps to address the core technical problems of weakened 3CG signals and blurred P-wave features caused by subcutaneous tissue attenuation in obese patients. By synergistically enhancing effective electrophysiological signals and suppressing background interference through signal waveform repair and matching degree calibration, the risk of deviations in the extraction of key features such as P-wave amplitude peaks and insufficient cross-modal fusion accuracy caused by signal distortion is reduced. This improves the stability of 3CG signals and the accuracy of core feature recognition, and achieves precise optimization of low-quality 3CG signals into clinically usable signals. It provides reliable electrophysiological data support for multimodal feature fusion and puncture difficulty grading.

[0032] Preferably, multimodal feature extraction is performed, and the specific process is as follows: S401, Perform ultrasound image feature extraction, and the specific process is as follows: Input qualified ultrasound images into a pre-trained multi-task convolutional neural network. The convolutional neural network includes a feature extraction backbone network and a confidence evaluation branch. The specific training process of the multi-task convolutional neural network is as follows: Divide the historical ultrasound image dataset into an ultrasound image training set and an ultrasound image test set. Based on the ultrasound image training set, perform end-to-end training by jointly optimizing the feature extraction loss (such as the error between features and labeled values) and the confidence evaluation loss (such as the deviation between predicted confidence and true quality), so that the network can simultaneously master feature extraction and confidence evaluation capabilities. Use the ultrasound image test set to perform multi-task convolutional neural network training. Network performance verification; Multi-level image feature vectors are extracted through the feature extraction backbone network, including shallow texture features and deep semantic features. Shallow texture features represent low-level visual information such as pixel grayscale distribution, local edge gradient, and texture roughness of the target blood vessel and surrounding tissue in the ultrasound image, which can intuitively reflect the basic structural details of the image and provide a foundation for subsequent quantification of blood vessel structure features. Deep semantic features represent high-level semantic information such as the complete morphological outline of the blood vessel, the regularity of the blood vessel wall, and the spatial relationship between the blood vessel and surrounding tissue, which are abstracted after hierarchical learning by the network. These directly relate to the core influencing factors of puncture difficulty and support accurate feature confidence assessment. The corresponding image feature confidence is obtained through the confidence assessment branch; S402 The 3CG signal feature extraction process is as follows: Qualified 3CG signals are segmented according to the complete cardiac cycle and then input into a pre-trained multi-task long short-term memory (LSTM) network. The specific training process of the LSTM network is as follows: The historical 3CG signal segmentation dataset is divided into a 3CG signal test set and a 3CG signal training set. Based on the 3CG signal test set, end-to-end training is performed by jointly optimizing the temporal feature extraction loss (e.g., the error between features and labeled values) and the rhythm evaluation loss (e.g., the deviation between predicted rhythm and true rhythm). This enables the network to simultaneously master signal temporal feature extraction and rhythm analysis capabilities. The 3CG signal training set is used to verify the effectiveness of the LSTM network. The LSTM network extracts signal features... The signal time-series feature vector includes short-term waveform features and long-term rhythm features. Short-term waveform features represent instantaneous characteristics within a single cardiac cycle, such as the peak amplitude, rise / fall slope, waveform distortion, and time interval between the QRS complex and the P wave, directly reflecting the local characteristics of the electrophysiological signal in a single heartbeat. Long-term rhythm features represent periodic patterns such as the stability of P wave intervals, amplitude fluctuation trends, and correlation with heart rate changes across multiple consecutive cardiac cycles, reflecting the overall rhythmic characteristics of the electrophysiological signal. The corresponding signal feature confidence level is obtained based on the signal segment quality assessment module, which is a sub-module in the multi-task long short-term memory network specifically used to evaluate the reliability of 3CG signal feature extraction.S403. Verify the feature accuracy evaluation of image feature and signal feature extraction quality, specifically including: if the image feature confidence level is greater than the initial image confidence level reference value, then mark the image feature vector as a qualified image feature vector, and perform multimodal feature fusion based on the qualified image feature vector; otherwise, send an image feature anomaly prompt. The initial image confidence level reference value is represented by the average image feature confidence level over a historical time period. If the signal feature confidence level is greater than the initial signal confidence level reference value, then mark the signal time series feature vector as a qualified signal time series feature vector. The vector is used to perform multimodal feature fusion based on the qualified signal time-series feature vector; otherwise, an abnormal signal feature prompt is sent. The initial signal confidence reference value is represented by the average signal feature confidence value over a historical time period. S404. The specific process of multimodal feature fusion is as follows: The qualified image feature vector and the qualified signal time-series feature vector are input into a preset multimodal fusion model (such as a neural network based on an attention mechanism), and the output is a multimodal feature set containing subsets of image features and signal features. The puncture difficulty is then graded based on the multimodal feature set.

[0033] In this embodiment, by using multimodal feature extraction and fusion, the complementarity of ultrasound images and 3CG signals can be fully utilized. Ultrasound images provide spatial features of the target vascular anatomy (such as diameter, depth, and wall regularity), while 3CG signals provide temporal features of cardiac electrical activity (such as P-wave characteristics and rhythm stability). Together, they cover the dual dimensions of anatomical morphology and physiological function required for puncture difficulty assessment. This solves the problem of key information loss caused by relying on a single modality (such as ultrasound images alone) or simple feature splicing in existing technologies. Compared with common feature extraction methods, this solution specifically selects multi-task convolutional neural networks and multi-task long short-term memory networks: multi-task convolutional neural networks, with their local receptive fields and hierarchical feature learning capabilities, can accurately capture the shallow texture details of ultrasound images. (e.g., edge gradients) and deep semantic information (e.g., the spatial relationship between blood vessels and tissues) are adapted to the spatial structural characteristics of image-type data; the multi-task long short-term memory network effectively handles the temporal dependence of 3CG signals through a gating mechanism, accurately extracting short-term waveform features (e.g., P-wave peak) and long-term rhythm features (e.g., periodic stability), adapting to the dynamic temporal characteristics of physiological electrical signals. The targeted selection of the two models ensures the accuracy of feature extraction for different modalities. Multimodal feature fusion is not a simple splicing, but a dynamic judgment of the necessity of fusion based on feature confidence. This avoids the interference of low-quality features on the fusion results and constructs a more comprehensive puncture difficulty feature system through the complementary fusion of high-quality features. Ultimately, it provides more reliable feature support for subsequent grading, improving the tightness of the correlation between features and puncture difficulty and the accuracy of assessment.

[0034] Example 2 provides an alternative scheme for feature accuracy assessment. In special scenarios involving obese patients (body mass index greater than 30) with vascular anatomy variations or a history of multiple chemotherapy treatments (this scenario simultaneously involves the complex factors of body fat interference and deteriorating vascular conditions, making conventional feature assessment difficult to adapt), targeted feature accuracy assessment is required. In this case, an alternative scheme for feature accuracy assessment is needed. The specific process of feature accuracy assessment is as follows: Obtaining feature accuracy deviation quantification parameters, including a vascular diameter deviation index (used to quantify the relative deviation between ultrasound-measured vascular diameter and the actual vascular diameter), a vascular depth deviation index (used to quantify the relative deviation between ultrasound-measured vascular diameter and the actual vascular diameter), a vascular tortuosity deviation index (used to quantify the relative deviation between ultrasound-assessed vascular tortuosity and the actual tortuosity), and a P-wave amplitude deviation index (used to quantify the relative deviation between 3CG-measured P-wave amplitude and the reference P-wave amplitude); and obtaining feature accuracy deviation quantification parameters reflecting the impact of vascular diameter deviation on the overall feature. The system includes a vascular diameter deviation influence factor (reflecting the weight of vascular depth deviation on the overall accuracy of the feature), a vascular tortuosity deviation influence factor (reflecting the weight of vascular tortuosity deviation on the overall accuracy of the feature), and a P-wave amplitude deviation influence factor (reflecting the weight of P-wave amplitude deviation on the overall accuracy of the feature). Based on the quantification parameter and the influence parameter, a feature deviation evaluation index is defined, specifically as follows: the result of weighted coupling processing of the quantification parameter and the influence parameter is used as the feature deviation evaluation index to quantify the measurement deviation between ultrasound image features and 3CG signal features. The system determines whether the feature deviation evaluation index is less than the initial feature deviation reference value. If so, multimodal feature fusion is performed; otherwise, a feature anomaly warning is sent. The initial feature deviation reference value is represented by the average value of the feature deviation evaluation index over a historical time period.

[0035] Specifically, the formula for the blood vessel diameter deviation index is as follows:

[0036] Among them, E D D represents the deviation index of blood vessel diameter. m D0 represents the target blood vessel diameter, and D0 represents the actual blood vessel diameter. Specifically, the target blood vessel diameter is automatically measured by the Canny operator and is used to reflect the accuracy of ultrasound measurement of blood vessel thickness under body fat interference. The actual blood vessel diameter is the actual blood vessel diameter confirmed by CTA / X-ray angiography and serves as a benchmark without attenuation interference to eliminate the influence of body fat on the judgment of blood vessel diameter.

[0037] Specifically, the formula for the vessel depth deviation index is as follows:

[0038] Among them, E H H represents the vessel depth deviation index. m H0 represents the actual vascular depth. Specifically, the vascular depth is automatically read from the ultrasound depth scale and is used to reflect the accuracy of ultrasound in locating deep blood vessels in obese patients. The actual vascular depth is obtained from CTA three-dimensional reconstruction and serves as the benchmark for vascular depth measurement, used to quantify the interference of body fat layer on the judgment of vascular depth.

[0039] Specifically, the formula for the vascular tortuosity deviation index is as follows:

[0040] Among them, E C C represents the deviation index of vascular tortuosity. m The tortuosity of the blood vessel is represented by C0, which represents the actual tortuosity of the blood vessel. Specifically, the tortuosity of the blood vessel is the ratio of the length of the tortuosity path obtained by the blood vessel's central axis (such as a B-spline curve) in the ultrasound image based on a fitting algorithm (such as a Bezier curve or a least squares fitting curve) to the straight-line distance between the two ends of the blood vessel. It is used to reflect the accuracy of ultrasound in assessing the morphology of fibrotic blood vessels after chemotherapy. The actual tortuosity of the blood vessel is obtained by CTA three-dimensional reconstruction and serves as a benchmark without ultrasound artifact interference to eliminate the influence of image blurring on the judgment of blood vessel morphology.

[0041] Specifically, the formula for the P-wave amplitude deviation index is as follows:

[0042] Among them, E A A represents the P-wave amplitude deviation index. m A0 represents the initial P-wave amplitude reference value. Specifically, the P-wave amplitude peak value is represented by the maximum amplitude value of the P-wave amplitude peak value monitored by the 3CG intracardiac electrocardiogram monitor within the P-wave time window. It is used to reflect the accuracy of the 3CG signal in capturing the core features of the P-wave in a high body fat scenario. The initial P-wave amplitude reference value is represented by the average value of the P-wave amplitude over a historical time period.

[0043] Specifically, the formula for the characteristic deviation evaluation index is as follows:

[0044] Among them, T P The following are the characteristic deviation assessment indicators: W1 represents the vascular diameter deviation influencing factor, W2 represents the vascular depth deviation influencing factor, W3 represents the vascular tortuosity deviation influencing factor, and W4 represents the P wave amplitude deviation influencing factor.

[0045] In this embodiment, feature accuracy assessment helps to systematically quantify and fuse the deviations between ultrasound image features and 3CG signal features in complex scenarios where body fat interference and vascular condition deterioration are superimposed. This reduces the risk of feature misjudgment caused by the limitations of single-modal measurement or interference from special physiological conditions, and improves the reliability and adaptability of feature accuracy assessment under the condition of obesity combined with vascular variation.

[0046] Specifically, the process of grading puncture difficulty is as follows: A multimodal feature set is input into a pre-defined puncture difficulty grading model (such as a random forest or deep learning model), which outputs the puncture success probability. Based on the success probability, the puncture difficulty level is classified. The specific training process of the puncture difficulty grading model is as follows: First, a training dataset is constructed. A high-quality multimodal feature set (containing a fused vector of multi-level ultrasound image features and 3CG signal temporal features) obtained after multimodal feature extraction and fusion from actual clinical puncture cases is used as the input features. The actual puncture results of the corresponding cases (such as success / failure labels) and the puncture difficulty level (low, medium, high) labeled by clinicians are used as supervision labels. The dataset is divided into a training set, a validation set, and a training set according to a pre-defined ratio. The test set is used for end-to-end iterative training based on the training set. An early stopping mechanism is introduced during training. If the core evaluation metrics on the validation set (accuracy for classification tasks and R² score for regression tasks) do not improve after a preset number of iterations, training is stopped to avoid overfitting. Finally, the test set is used to verify the model's generalization ability. If the test set metrics (such as classification accuracy ≥90% and regression MAE ≤0.05) meet the preset standards, the optimal model parameters are saved. If they do not meet the standards, the model returns to the hyperparameter optimization stage to adjust the range and retrain until the model performance meets the standards. Finally, a puncture difficulty classification model that can accurately map the correlation between multimodal features and puncture difficulty is obtained. The preset ratio, preset number of iterations, and preset standards are all set in advance by preset personnel.

[0047] The specific process for classifying the difficulty levels of punctures is as follows: If the success rate of the puncture meets the conditions for Level 1 puncture difficulty, the current puncture difficulty is marked as low difficulty; if the success rate of the puncture meets the conditions for Level 2 puncture difficulty, the current puncture difficulty is marked as medium difficulty; if the success rate of the puncture meets the conditions for Level 3 puncture difficulty, the current puncture difficulty is marked as high difficulty. Level 1 puncture difficulty means that the success rate of the puncture is greater than the initially set Level 1 success rate reference value; Level 2 puncture difficulty means that the success rate of the puncture is not greater than the initially set Level 1 success rate reference value, but is greater than the initially set Level 2 success rate reference value; Level 3 puncture difficulty means that the success rate of the puncture is not greater than the initially set Level 2 success rate reference value. The initially set Level 1 success rate reference value is set in advance by preset personnel, and the initially set Level 2 success rate reference value is represented by the average success rate of punctures over a historical period.

[0048] In this embodiment, by assessing feature accuracy and grading puncture difficulty, the interference of low-confidence features on the assessment results can be eliminated through quality verification of multimodal features of ultrasound images and 3CG signals. This ensures that the features input into the puncture grading model have a real correlation with the puncture difficulty, which helps to solve the grading bias problem caused by inconsistent quality when traditional features are used directly. Furthermore, the validated high-quality multimodal features can be transformed into quantified puncture success probabilities and corresponding difficulty levels, breaking the limitation of relying on the subjective experience of clinicians to judge the difficulty. This makes the assessment results more objective and consistent, improves the accuracy of puncture difficulty assessment, effectively reduces the risk of puncture failure, and ultimately achieves a dual improvement in the safety and efficiency of puncture operations.

[0049] Figure 3 shows the flowchart of the puncture difficulty grading method that fuses ultrasound images and 3CG signals. As shown in Figure 3, S1, data quality assessment and compensation triggering: In a specified puncture difficulty grading scenario, a baseline puncture data quality assessment is performed. Based on the obtained baseline puncture data quality assessment results, it is determined whether to trigger adaptive compensation for ultrasound images to correct image blurring caused by obesity and tissue interference, and adaptive compensation for 3CG signals to correct 3CG signal weakening. Through data quality assessment and compensation triggering, the clinical usability of ultrasound images and 3CG signals is guaranteed from the data source, providing a high-quality data foundation for subsequent feature processing; S2, multimodal feature extraction and fusion. After the baseline puncture data quality assessment is deemed satisfactory, multimodal feature extraction of ultrasound image features and 3CG electrophysiological features to capture factors influencing puncture difficulty is performed. Based on the obtained multimodal feature extraction results, a decision is made on whether to perform multimodal feature fusion. Through multimodal feature extraction and fusion judgment, it is helpful to deeply explore the complementary information of ultrasound image features and 3CG electrophysiological features, and strengthen the feature's ability to represent puncture difficulty. S3, Puncture Difficulty Grading: After the multimodal feature fusion is completed, puncture difficulty grading is performed to quantify the complexity of the puncture operation. This helps to transform multidimensional features into intuitive difficulty levels, providing quantitative decision support for clinical puncture strategy formulation.

[0050] In this embodiment, the three modules of data quality assessment and compensation triggering, multimodal feature extraction and fusion judgment, and puncture difficulty classification execution are sequentially connected to form a complete technical link from raw data processing to final result output. This systematically solves the core problems of unreliable data, single features, and subjective results in traditional puncture difficulty assessment. Data quality assessment and compensation triggering provide a qualified data foundation for feature extraction, and feature extraction and fusion judgment provide accurate feature support for classification execution, forming a dynamic iterative mechanism that adapts to actual clinical needs.

[0051] Figure 4 shows the random forest ensemble architecture of the puncture difficulty grading model for the fusion of ultrasound images and 3CG signals. As shown in Figure 4, the multi-dimensional standardized features of the fused ultrasound images and 3CG signals are used as inputs to the random forest model (such as ultrasound image modal features including vessel diameter and skin-to-vessel depth, and 3CG electrophysiological signal features such as P-wave amplitude peak and 3CG signal-to-noise ratio). Feature learning is performed through three functionally differentiated decision trees: Tree-1 focuses on ultrasound image modal features, capturing factors influencing puncture difficulty from the perspective of vascular structure visualization; Tree-2 focuses on 3CG electrophysiological signal features, analyzing operational difficulty constraints from the perspective of signal stability; and Tree-3 integrates cross-modal features to achieve fusion inference. Then, the Averaging (result fusion) module averages the predicted puncture success probability values ​​output by the three decision trees, and finally outputs a quantitative success probability result for puncture difficulty grading (e.g., an output success probability of 89%), providing an accurate basis for clinical puncture difficulty assessment.

[0052] In summary, by performing a baseline puncture data quality assessment and using the assessment results to determine whether to trigger adaptive compensation for ultrasound images and 3CG signals, we can filter out low-quality ultrasound image data and interference from 3CG signal data at the source. This provides qualified data for subsequent feature extraction and fusion, reduces data distortion caused by the lack of data processing for obese individuals, and consequently reduces input bias in the grading model, thus ensuring the reliability of the entire decision-making process from the ground up. After the baseline puncture data quality assessment is satisfactory, multimodal feature extraction is performed, and the confidence level of the acquired features is used to determine whether to perform multimodal feature fusion, which helps... This invention overcomes the limitations of existing technologies where a single ultrasound image can only reflect static information and a single 3CG signal can only reflect dynamic physiological information. By integrating the complementary advantages of these two types of features, a more comprehensive feature representation of puncture difficulty is constructed. Compared with the single-modal assessment or simple feature splicing of existing technologies, this invention significantly improves the adaptability of features to complex puncture scenarios. After multimodal feature fusion, a puncture difficulty grading is performed to quantify the complexity of the puncture operation. This helps to replace the traditional subjective judgment that relies on experience, reduce assessment bias caused by individual experience differences, and thus guide the operator to select an appropriate operation plan based on the grading results, thereby improving the puncture success rate and reducing the risk of complications.

[0053] 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 puncture difficulty grading system integrating ultrasound images and 3CG signals, characterized in that, include: The system comprises a data quality assessment and compensation triggering module, a multimodal feature extraction and fusion judgment module, and a puncture difficulty grading execution module. The data quality assessment and compensation triggering module performs a puncture baseline data quality assessment and determines whether to trigger adaptive ultrasound image compensation for correcting image blurring and adaptive 3CG signal compensation for correcting 3CG signal weakening based on the acquired baseline data quality assessment results. The multimodal feature extraction and fusion judgment module, after the baseline data quality assessment is deemed satisfactory, performs multimodal feature extraction to capture ultrasound image features and 3CG signal electrophysiological features, and determines whether to perform multimodal feature fusion based on the acquired multimodal feature extraction results. The puncture difficulty grading execution module, after multimodal feature fusion is completed, performs a puncture difficulty grading to quantify the complexity of the puncture operation.

2. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 1, characterized in that, The specific process of the puncture baseline data quality assessment is as follows: S101, Obtain puncture baseline data quality assessment parameters. The specific process is as follows: For ultrasound image data, define the corresponding first puncture baseline data quality assessment parameters, including ultrasound image signal-to-noise ratio and vessel wall edge clarity. The specific definition process is as follows: The average of the ratio of the standard deviation of pixel values ​​in the vessel area to the standard deviation of pixel values ​​in the background area in each frame of the obtained images during the puncture baseline data quality assessment period is expressed as the ultrasound image signal-to-noise ratio; The result of quantifying the ratio of the vessel wall edge clarity index to the initially set vessel wall edge clarity reference value during the puncture baseline data quality assessment period is expressed as the vessel wall edge clarity; For 3CG signal data, define the corresponding second puncture baseline data quality assessment parameters, including 3CG signal signal-to-noise ratio and P wave discernibility. The specific definition process is as follows: The result of averaging the ratio of the P wave signal power to the baseline noise signal power in each cardiac cycle obtained during the puncture baseline data quality assessment period is expressed as the 3CG signal signal-to-noise ratio; The result of quantifying the ratio of the standard deviation of pixel values ​​in the vessel area to the standard deviation of pixel values ​​in the background area in each frame of the obtained images during the puncture baseline data quality assessment period is expressed as the 3CG signal signal-to-noise ratio; The result of quantifying the ratio of the vessel wall edge clarity index to the initially set vessel wall edge clarity reference value during the puncture baseline data quality assessment period is expressed as the vessel wall edge clarity; For 3CG signal data, define the corresponding second puncture baseline data quality assessment parameters, including 3CG signal signal-to-noise ratio and P wave discernibility. The specific definition process is as follows: The result of averaging the ratio of the P wave signal power to the baseline noise signal power in each cardiac cycle obtained during the puncture baseline data quality assessment period is expressed as the 3CG signal signal-to-noise ratio; The result of quantifying the ratio of the standard deviation of pixel values ​​in the vessel area to the initial set vessel wall The average of the ratios of the P wave amplitude to the amplitudes of adjacent QRS complexes in each cardiac cycle is expressed as the P wave intelligibility; S102, after weighted coupling processing of the first puncture baseline data quality assessment parameters and the corresponding first quality weight parameters, a first quality quantification value for evaluating the overall quality level of ultrasound image data is obtained; after harmonic averaging processing of the second puncture baseline data quality assessment parameters, a second quality quantification value for evaluating the overall quality level of 3CG signal data is obtained; the first quality weight parameters include the image signal-to-noise ratio weight coefficient and blood... Tube wall clarity weighting coefficient; S103, triggering the adaptive compensation mechanism, the specific process is as follows: if the first quality quantification value is less than the initially set first quality reference value, it is determined that the quality of the ultrasound image is substandard, the ultrasound image is marked as an ultrasound image to be compensated, and adaptive compensation of the ultrasound image is triggered, otherwise, multimodal feature extraction is performed; if the second quality quantification value is less than the initially set second quality reference value, it is determined that the quality of the 3CG signal is substandard, the 3CG signal is marked as a 3CG signal to be compensated, and adaptive compensation of the 3CG signal is triggered, otherwise, multimodal feature extraction is performed.

3. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 2, characterized in that, The specific compensation process of the adaptive compensation of the ultrasound image is as follows: S201, perform frequency domain analysis and noise spectrum estimation, the specific process is as follows: convert the ultrasound image to be compensated from the spatial domain to the frequency domain to obtain the corresponding image frequency domain spectrum; separate the image frequency domain spectrum into the image main signal frequency band and the image residual frequency band; S202, perform adaptive adjustment of the image main signal gain, take the first quality quantization value and the average energy value of the image main signal frequency band as input parameters, query the preset signal gain mapping table, and obtain the corresponding main signal gain ratio coefficient; Using the main signal gain ratio coefficient as the adjustment amount, the amplitude gain of the main signal frequency band of the image is gradually increased in the direction of gain increase; S203, image residual noise suppression is performed, the first quality quantization value and the average energy value of the image residual frequency band are used as input parameters, and the corresponding noise suppression coefficient is obtained by querying the preset noise suppression mapping table. Using the noise suppression coefficient as the adjustment amount, the amplitude gain of the residual noise frequency band is gradually reduced in the direction of gain decrease; the average energy value of the residual frequency band is represented by summing the amplitudes of all frequency points in the residual frequency band and then quantifying the ratio of this summation to the total number of frequency points in the band. The average energy value of the residual frequency band is used to quantify the overall intensity of the background noise; S204, continuously monitor the first quality quantization value. When the first quality quantization value is not less than the initial first quality reference value, image edge optimization is performed to obtain a qualified ultrasound image, and multimodal feature extraction is performed based on the qualified ultrasound image. Otherwise, the main signal gain adaptive adjustment and residual noise suppression are continued. When the initial maximum number of iterations for adaptive compensation of the ultrasound image is reached, if the first quality quantization value is still less than the initial first quality reference value, an image quality adjustment failure prompt is sent.

4. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 3, characterized in that, The specific process for image edge optimization is as follows: the main signal band and residual signal band of the image are subjected to frequency domain spectrum synthesis processing to obtain a preliminary enhanced image; the preliminary enhanced image is subjected to edge sharpening and structure preservation processing to generate a qualified ultrasound image.

5. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 4, characterized in that, The specific process of adaptive compensation for the 3CG signal is as follows: S301. Collect continuous 3CG signal data and identify the complete cardiac cycle. Obtain the median amplitude value of each corresponding sampling point in the cardiac cycle waveform and use the synthesized standard waveform as the initial 3CG reference signal. Based on the cardiac cycle, perform signal segmentation on the 3CG signal to be compensated to obtain segmented 3CG signals. Perform alignment operation on each segmented 3CG signal and the initial 3CG reference signal, and obtain the minimum cumulative distance of the alignment path. After summing the obtained minimum cumulative distance of the alignment path with the preset alignment constant, perform the reciprocal operation to represent the signal matching quantification value used to reflect the qualification degree of the segmented 3CG signal to be compensated. S302. Perform signal differentiation compensation based on the signal matching quantification value. The specific process is as follows: For segmented 3CG signals with a signal matching quantification value greater than the initial matching degree reference value, perform adaptive Wiener filtering. For segmented 3CG signals with a signal matching quantification value not greater than the initial matching degree reference value, perform waveform repair. After the waveform repair is completed, perform adaptive Wiener filtering again based on the obtained repaired 3CG signal.

6. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 5, characterized in that, The specific process of waveform restoration is as follows: Identify the maximum restoration interval in the segmented 3CG signal that corresponds to the initial 3CG reference signal. The specific identification process is as follows: Based on the maximum restoration interval, perform weighted fusion on the value of each sampling point in the maximum restoration interval of the original signal and the corresponding sampling point value in the initial 3CG reference signal; re-embed the restored maximum restoration interval into the corresponding position of the original signal segment, and perform smoothing filtering on the signal junction to obtain the restored 3CG signal; after S303 and 3CG signal adaptive compensation, determine whether the re-acquired second quality quantization value is less than the initial second quality reference value. If so, send a 3CG signal adaptive compensation failure prompt; otherwise, mark the 3CG signal after 3CG signal adaptive compensation as a qualified 3CG signal, and perform multimodal feature extraction based on the qualified 3CG signal.

7. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 6, characterized in that, The multimodal feature extraction process is as follows: S401. Perform ultrasound image feature extraction, specifically as follows: Input qualified ultrasound images into a pre-trained multi-task convolutional neural network, which includes a feature extraction backbone network and a confidence evaluation branch; extract multi-level image feature vectors through the feature extraction backbone network, including shallow texture features and deep semantic features; obtain the corresponding image feature confidence through the confidence evaluation branch; S402. Perform 3CG signal feature extraction, specifically as follows: Segment qualified 3CG signals according to the complete cardiac cycle and input them into a pre-trained multi-task long short-term memory network; extract the signal temporal feature vectors of the signal through the multi-task long short-term memory network; obtain the corresponding signal feature confidence based on the signal segment quality evaluation module; S403. Verify the image features and signal features respectively. The feature accuracy assessment of the extracted quality specifically includes: if the image feature confidence is greater than the initial image confidence reference value, the image feature vector is marked as a qualified image feature vector, and multimodal feature fusion is performed based on the qualified image feature vector; otherwise, an image feature anomaly prompt is sent. If the signal feature confidence is greater than the initial signal confidence reference value, the signal time series feature vector is marked as a qualified signal time series feature vector, and multimodal feature fusion is performed based on the qualified signal time series feature vector; otherwise, a signal feature anomaly prompt is sent. S404, the specific process of multimodal feature fusion is as follows: the qualified image feature vector and the qualified signal time series feature vector are input into the preset multimodal fusion model, the multimodal feature set containing the image feature subset and the signal feature subset is output, and the puncture difficulty is graded based on the multimodal feature set.

8. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 7, characterized in that, The specific process for feature accuracy assessment is as follows: Quantitative parameters of feature accuracy deviation, including vessel diameter deviation index, vessel depth deviation index, and vessel tortuosity deviation index, are obtained; influence parameters of feature accuracy deviation, including vessel diameter deviation influence factor, vessel depth deviation influence factor, vessel tortuosity deviation influence factor, and P-wave amplitude deviation influence factor, are obtained; based on the quantitative parameters of feature accuracy deviation and the influence parameters of feature accuracy deviation, a feature deviation assessment index is defined, specifically as follows: the result of weighted coupling processing of the quantitative parameters of feature accuracy deviation and the influence parameters of feature accuracy deviation is used as the feature deviation assessment index for quantifying the measurement deviation between ultrasound image features and 3CG signal features; it is determined whether the feature deviation assessment index is less than the initially set feature deviation reference value. If so, multimodal feature fusion is performed; otherwise, a feature anomaly warning is sent.

9. The puncture difficulty grading system fusing ultrasound images and 3CG signals according to claim 7, characterized in that, The specific process of grading the puncture difficulty is as follows: A multimodal feature set is input into a preset puncture difficulty grading model, and the puncture success probability is output. The puncture difficulty level is then classified based on the success probability. The specific process of classifying the puncture difficulty level is as follows: If the success probability meets the first-level puncture difficulty condition, the current puncture difficulty is marked as low difficulty; if the success probability meets the second-level puncture difficulty condition, the current puncture difficulty is marked as medium difficulty; if the success probability meets the third-level puncture difficulty condition, the current puncture difficulty is marked as high difficulty. The first-level puncture difficulty condition indicates that the puncture success rate is greater than the initially set first-level success rate reference value; the second-level puncture difficulty condition indicates that the puncture success rate is not greater than the initially set first-level success rate reference value, but greater than the initially set second-level success rate reference value; the third-level puncture difficulty condition indicates that the puncture success rate is not greater than the initially set second-level success rate reference value.

10. A method for grading puncture difficulty by fusing ultrasound images and 3CG signals, applied to the puncture difficulty grading system by fusing ultrasound images and 3CG signals as described in any one of claims 1-9, characterized in that, include: S1. Perform a baseline puncture data quality assessment and, based on the assessment results, determine whether to trigger adaptive ultrasound image compensation for correcting image blurring and adaptive 3CG signal compensation for correcting 3CG signal weakening. S2. After the baseline puncture data quality assessment is passed, perform multimodal feature extraction of ultrasound image features and 3CG signal electrophysiological features to capture factors affecting puncture difficulty, and, based on the multimodal feature extraction results, determine whether to perform multimodal feature fusion. S3. After multimodal feature fusion, perform puncture difficulty grading to quantify the complexity of the puncture procedure.

Images