Peripheral lung cancer pleural invasion evaluation system based on ultrasonic targeted radiography

Abstract

The invention relates to the technical field of medical imaging, and discloses a peripheral lung cancer pleural invasion evaluation system based on ultrasonic targeted radiography, comprising: a double-frequency ultrasonic probe configured to independently work at different frequencies to realize depth imaging and high-resolution imaging of lung tissues and suitable for different types of lung lesions; the signal processing module can receive, analyze and process echo signals transmitted by the ultrasonic probe, generate images in real time and extract feature information; and the dynamic motion compensation module is used for automatically adjusting the emission time sequence and parameter setting of the ultrasonic signal based on the real-time breathing motion state of the patient. By adopting the design of the double-frequency ultrasonic probe and the core-shell structure, high-resolution imaging and depth detection of lung tissues are realized. By dynamically adjusting the thickness of the probe shell to be matched with the frequency, the system can adapt to complex tissue environments under different physiological conditions, and image definition and stability are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical imaging technology, in particular to a peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast. BACKGROUND

[0002] The accurate evaluation of peripheral lung cancer pleural invasion has always been a difficulty in clinical diagnosis. Traditional ultrasound imaging technology is limited by probe design and cannot balance penetration depth and resolution. Single frequency probes often have signal attenuation in complex lung tissue, resulting in acceptable superficial imaging and unclear deep tissue. Doctors have to adjust the equipment parameters repeatedly, which is time-consuming and cannot guarantee stable image quality. This technical bottleneck directly affects the detection rate of early micro lesions.

[0003] Existing systems lack effective means to deal with dynamic interference of respiratory motion. The relative displacement of the ultrasound probe and the tissue caused by the patient's breathing produces motion artifacts. Most devices use fixed parameters for acquisition and cannot correct the signal timing in real time, resulting in trailing or fault in the final image. This problem is particularly prominent in long-term monitoring, and doctors often have to rely on experience to guess the true lesion morphology, which greatly reduces the reliability of diagnosis.

[0004] On the data analysis level, traditional methods are mostly limited to single-dimensional feature extraction. Independent analysis in time or frequency domain easily loses the spatiotemporal correlation information of the lesion tissue, like missing key pieces of a puzzle. Manual interpretation is not only inefficient but also affected by subjective factors, and different doctors may draw contradictory conclusions from the same image. This extensive analysis mode cannot meet the needs of quantitative and standardized diagnosis in the era of precision medicine. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application provides a peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast, which solves the problems of difficulty in balancing penetration depth and resolution in traditional ultrasound imaging, interference of respiratory motion leading to artifacts, and incomplete single-dimensional feature analysis.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: a peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast, comprising:

[0007] A dual-frequency ultrasound probe configured to work independently at different frequencies to achieve depth imaging and high-resolution imaging of lung tissue, suitable for different types of lung lesions;

[0008] A signal processing module capable of receiving, analyzing and processing echo signals emitted by the ultrasound probe, generating images in real time and extracting feature information;

[0009] A dynamic motion compensation module automatically adjusts the emission timing and parameter settings of the ultrasound signals based on the real-time respiratory motion state of the patient.

[0010] A spatio-temporal feature fusion algorithm module extracts multi-dimensional features from the data obtained by the signal processing module, including pathological features related to pleural invasion of peripheral lung cancer.

[0011] A regional clustering analysis module uses clustering algorithms to analyze different tissue regions in the ultrasound images and groups adjacent tissues based on similarity.

[0012] A result display and feedback module can display the evaluation results to the operator in a visual manner and provide corresponding adjustment suggestions based on the evaluation results.

[0013] Further, by integrating various technical elements, the detection accuracy of pleural invasion of peripheral lung cancer is improved. The functions of each module are closely integrated to ensure that the system not only effectively images but also analyzes and feeds back in real time based on the imaging results, promoting accurate diagnosis by clinicians. The design of the entire system reflects a deep understanding of modern medical needs and actual operations, focusing on both the advancement of technology and user experience.

[0014] Preferably, the dual-frequency ultrasound probe adopts a core-shell structure design, ensuring clear images under different physiological conditions through the penetration ability and imaging clarity of sound waves.

[0015] Further, the innovation of this design lies in its enhanced adaptability of the ultrasound probe to different tissue interfaces. The application of the core-shell structure not only improves the conduction efficiency of sound waves but also reduces the attenuation of sound waves during transmission, ensuring clear ultrasound imaging even under complex physiological conditions. Such a design concept adapts to the variability of clinical environments and has wide practical application value.

[0016] Preferably, the thickness of the probe shell varies according to the emission frequency, optimizing the acoustic transmission characteristics of different levels of tissue.

[0017] Further, optimizing the thickness of the probe shell allows it to better adapt to the acoustic characteristics required by different frequency emissions, effectively improving the imaging quality of different depth tissues. With the change of emission frequency, the physical properties of the probe are adjusted to ensure that the signal is transmitted to the target tissue at the best angle and amplitude, thereby maximizing the resolution of the ultrasound image and providing strong support for accurate identification of lesions.

[0018] Preferably, the signal processing module includes signal separation algorithms based on time and frequency domain analysis, which can process complex signals and prevent background noise interference.

[0019] Further, by introducing a comprehensive analysis of time and frequency domains, the in-depth exploration of ultrasonic signals is ensured. The design of the algorithm allows the system to identify effective signals in a noisy environment, reflecting more accurate ultrasonic images. The effectiveness of signal separation lies in its effective suppression of background noise, thereby improving the reliability of subsequent image analysis. The implementation of this technology will significantly improve image quality and provide a solid foundation for the diagnosis process.

[0020] Preferably, the time and frequency domain analysis method combines signal separation, descriptive model construction and image reconstruction technology.

[0021] Further, the comprehensiveness and progressiveness in signal processing. By comprehensively using multiple analysis methods, the system not only can restore the lost image information, but also can reconstruct clear ultrasonic images, thereby providing high-quality support for doctors' clinical judgment. This innovative processing capability is particularly important in the diagnosis of complex cases, demonstrating the potential of the invention in clinical applications.

[0022] Preferably, the dynamic motion compensation module uses a biomechanical model to obtain the patient's breathing pattern in real time and adjust the transmission timing and phase of the ultrasonic signal in a timely manner.

[0023] Further, the use of a biomechanical model provides the system with unique dynamic adjustment capabilities, enabling it to overcome the image blurring problem caused by respiratory motion. By tracking the patient's breathing behavior in real time, this module can adaptively optimize the signal transmission timing, ensuring the stability and continuity of the image. This innovative design improves the application flexibility of the system, especially in clinical scenarios that require continuous monitoring, ensuring the high consistency of image quality.

[0024] Preferably, the biomechanical model uses real-time acquired respiratory frequency and amplitude information to dynamically adjust the transmission parameters.

[0025] Further, by combining biomechanical methods, the detection system is injected with intelligent elements. The real-time feedback capability of the biomechanical model ensures that the transmission parameters can accurately adapt to the patient's physiological state, thereby improving the accuracy and practicality of ultrasonic imaging, enabling doctors to make faster and more effective diagnoses in complex situations.

[0026] Preferably, the spatiotemporal feature fusion algorithm module uses tensor decomposition technology to extract ultrasonic features in space and time dimensions, and comprehensively analyzes different physical parameters. The fusion algorithm formula is as follows:

[0027]

[0028] Wherein:

[0029] X ijk represents the extracted feature value;

[0030] W abc denotes an element of a core tensor;

[0031] U ia denotes an element of a spatial feature matrix;

[0032] V jb denotes an element of a temporal feature matrix;

[0033] S kc denotes a weight related to a physical parameter;

[0034] A denotes the dimension of spatial features;

[0035] B denotes the dimension of temporal features;

[0036] C denotes the dimension of physical parameters.

[0037] Further, the spatio-temporal features in the ultrasound image data are deeply mined. Through tensor decomposition technology, the system can integrate multi-dimensional information, thereby enhancing the integrity and accuracy of feature extraction. This method effectively reduces the probability of information loss, making the diagnosis of peripheral pulmonary carcinoma pleural invasion more accurate, providing a more solid guarantee for clinical judgment.

[0038] Preferably, the tensor decomposition technology includes selecting a suitable core tensor structure, optimizing parameter settings and feature fusion strategies.

[0039] Further, through reasonable core tensor structure, the processing speed and accuracy of the algorithm can be significantly improved. This optimization enables the system to maintain high efficiency when facing complex data, ensuring the reliability of the extraction results. The implementation of this technology will greatly improve the intelligent level of ultrasound image analysis and better support medical decision-making.

[0040] Preferably, the regional clustering analysis module combines deep learning methods to automatically identify and label potential lesion areas using spatial features of ultrasound images, as follows:

[0041]

[0042] wherein:

[0043] R denotes the clustering result;

[0044] X ij denotes the pixel features in the ultrasound image;

[0045] C r denotes the center point of the rth class;

[0046] C k denotes the center point of other classes;

[0047] N represents the number of rows of the ultrasound image;

[0048] M represents the number of columns of the ultrasound image;

[0049] represents finding the class r that best fits the current pixel features X ij in all classes.

[0050] Further, combined with the regional clustering analysis module of the deep learning algorithm, not only the identification efficiency of the potential lesion area is improved, but also the burden of doctors is reduced through automatic labeling. This technology can quickly filter out important pathological information from massive data, greatly improving the efficiency and accuracy of preliminary diagnosis, providing support for subsequent treatment plans, and showing the foresight and practicality of the application.

[0051] The application provides a peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast.

[0052] The application has the following beneficial effects:

[0053] 1. The application adopts a dual-frequency ultrasound probe and a core-shell structure design, realizing high-resolution imaging and deep detection of lung tissue. Compared with the limitations of single-frequency probes in the prior art, the penetration and imaging blurring problems caused by tissue differences are solved. By dynamically adjusting the probe shell thickness and frequency matching, the system can adapt to complex tissue environments under different physiological conditions, ensuring image clarity and stability.

[0054] 2. The application combines biomechanical models and dynamic motion compensation modules, and the system can track the patient's breathing state in real time and adjust signal parameters. The artifacts caused by respiratory motion in the prior art are effectively suppressed, and the continuity and accuracy of the image are significantly improved. This technical solution breaks through the limitations of traditional static imaging and provides reliable support for dynamic monitoring.

[0055] 3. The application integrates multi-dimensional data of space, time and physical parameters through tensor decomposition technology, extracts the pathological features of peripheral lung cancer pleural invasion, and solves the problems of incomplete feature extraction and low information utilization rate compared with traditional single-dimensional analysis methods. Combined with the regional clustering module driven by deep learning, the system can automatically identify the lesion area and reduce the error of artificial intervention.

[0056] 4. The application realizes automatic processing and intuitive presentation of ultrasound images through the collaborative design of the regional clustering analysis module and the result display feedback module. The inefficient mode of relying on manual interpretation in the prior art is broken, and the problems of complicated diagnosis process and strong subjectivity are solved. The system significantly shortens the analysis time through intelligent labeling and optimization suggestions, reduces the operation threshold, and provides an efficient and standardized diagnosis tool for clinicians. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A system framework diagram of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0059] Please refer to the accompanying Figure 1 The embodiments of the present application provide a peripheral lung cancer pleural invasion evaluation system based on ultrasonic targeted contrast, comprising:

[0060] A dual-frequency ultrasonic probe is configured to work independently at different frequencies to realize deep imaging and high-resolution imaging of lung tissue, and is suitable for different types of lung lesions;

[0061] A signal processing module can receive, analyze and process echo signals emitted by the ultrasonic probe, generate images in real time and extract feature information;

[0062] A dynamic motion compensation module automatically adjusts the emission timing and parameter settings of the ultrasonic signal based on the real-time respiratory motion state of the patient;

[0063] A spatiotemporal feature fusion algorithm module extracts pathological features related to peripheral lung cancer pleural invasion by performing multi-dimensional feature extraction on the data obtained by the signal processing module;

[0064] A regional clustering analysis module uses a clustering algorithm to analyze different tissue regions in the ultrasonic image and groups adjacent tissues according to similarity;

[0065] A result display and feedback module can display the evaluation results to the operator in a visual manner and provide corresponding adjustment suggestions based on the evaluation results.

[0066] Module 1: Dual-frequency ultrasonic probe

[0067] In this embodiment, the dual-frequency ultrasonic probe module adopts a core-shell layered structure design, and realizes multi-modal imaging through acoustic impedance matching and frequency response optimization. The probe shell is composed of acoustic impedance gradient materials, and its acoustic characteristics change continuously from the outer layer to the inner layer. The acoustic impedance of the outer layer material is close to the characteristics of biological soft tissue, and the acoustic impedance of the inner layer material is matched with the characteristics of piezoelectric crystals. This gradient design reduces the reflection loss of acoustic waves at the probe-tissue interface and improves the efficiency of acoustic energy transmission.

[0068] The probe is internally integrated with two sets of independently controlled annular transducer arrays. The first array is composed of high-frequency piezoelectric ceramic elements, working in the high-resolution imaging frequency band; the second array adopts low-frequency composite piezoelectric materials, working in the deep penetration imaging frequency band. The two sets of arrays are dynamically switched through electronic switches, and the switching logic is based on the real-time collected tissue depth data: when the depth of the target region is detected to exceed the preset threshold, the low-frequency array is automatically activated to enhance the penetration capability.

[0069] Preferably, the probe shell thickness is dynamically adjusted according to the working frequency. In the high-frequency mode, a quarter-wavelength matching layer design is adopted to suppress surface echo interference through interference cancellation principle; in the low-frequency mode, a half-wavelength matching layer design is switched to, which utilizes the acoustic superposition effect to enhance the penetration performance. The shell material is selected to be a biocompatible elastomer, the elastic modulus of which can be adaptively adjusted with the contact pressure to ensure flexible coupling with human tissues.

[0070] Further, a biocompatible matching layer is provided on the surface of the probe, which is composed of nanoparticles and a polymer matrix. By adjusting the particle filling rate, the acoustic impedance value is controlled to realize the impedance transition between the probe and the skin. The matching layer surface is designed with microstructure grooves to reduce sidelobe noise by guiding the scattering direction of acoustic waves, while enhancing the directivity of the acoustic field.

[0071] In this embodiment, the electronic control system of the probe includes a digital beamforming unit. The unit precisely controls the transmission timing of each array element through a programmable delay line to form a dynamic focused acoustic beam.

[0072] In the transmission phase, the time delay of the excitation signal of each array element is calculated by the following formula:

[0073]

[0074] Wherein:

[0075] τ n represents the signal delay time of the nth array element;

[0076] x n ,y n represents the coordinate position of the nth array element in the planar array;

[0077] x f ,y f ,z f represents the three-dimensional spatial coordinates of the target focal point;

[0078] c represents the acoustic velocity value in the current medium.

[0079] The reception stage adopts a dynamic aperture technology, which dynamically adjusts the number of effective receiving array elements according to the echo arrival time, thereby improving the signal-to-noise ratio while ensuring the spatial resolution.

[0080] Preferably, the probe is integrated with a temperature sensing and power regulation module. When the probe surface temperature is detected to exceed a safety threshold, a multi-stage protection mechanism is automatically triggered: first, the transmit voltage amplitude is reduced, and then the active cooling system is started. The module monitors the thermal distribution state through a distributed temperature sensor array, combines with the heat conduction model to predict the temperature change trend, and ensures the stability and safety of the device in long-time work.

[0081] Module 2: Signal processing module

[0082] In this embodiment, the signal processing module adopts a time-frequency joint analysis method to realize deep analysis and feature enhancement of the ultrasonic echo signal through a multi-stage processing chain. The module receives the original electric signal from the dual-frequency ultrasonic probe, sequentially performs time domain preprocessing, frequency domain decomposition and signal reconstruction, and finally outputs a data matrix containing complete acoustic features.

[0083] In the original signal processing flow, first, the baseline dynamic correction is performed on the collected time domain signal, and the mathematical expression is defined as:

[0084]

[0085] Wherein:

[0086] t: time axis variable, representing the current time point of signal processing;

[0087] s raw (t): original ultrasonic echo signal, representing the time domain voltage signal received by the probe;

[0088] s corrected (t): baseline corrected output signal, used for subsequent processing links;

[0089] T: time width parameter of sliding window, determines the time window length of baseline estimation;

[0090] w(·): window function (preferably Hanning window), satisfying the normalization condition;

[0091] τ: integral operation variable, representing the time sampling point within the window;

[0092] dτ: differential operator symbol, representing the Lebesgue integral of variable τ.

[0093] The signal baseline estimator is dynamically constructed by time-domain convolution operation, wherein the window function weighting strategy can effectively suppress the interference of sudden signal on baseline estimation. Specifically, the sliding window moves point by point along the time axis, and the weighted integral mean of the signal in the window is calculated at each position t, which reflects the low-frequency drift component caused by the change of probe contact pressure. By dynamically deducting this baseline component from the original signal, both low-frequency interference and the truncation effect of traditional high-pass filter on signal edge features are avoided.

[0094] Further, the module optimizes the design of the wavelet multi-scale analysis unit. The continuous wavelet transform is performed on the baseline-corrected signal, and the Morlet wavelet basis function with time-frequency localization characteristics is used to achieve full-band coverage from sub-millimeter fine structure to macroscopic organizational features through scale parameter adjustment. The instantaneous phase information of the signal is preserved during the transformation process, and a complex wavelet coefficient matrix is constructed to provide multi-dimensional feature input for subsequent organizational elasticity parameter inversion.

[0095] In the signal reconstruction stage, the module introduces over-complete dictionary sparse representation technology. The dictionary is constructed by Gabor atoms and curvelet atoms, wherein the Gabor atoms are used to capture the local time-frequency features of the signal, and the curvelet atoms enhance the representation ability of the anisotropic structure of the tissue. The sparse coefficients are iteratively solved by the orthogonal matching pursuit algorithm, which significantly reduces the data dimension and improves the processing efficiency on the premise of ensuring the reconstruction accuracy.

[0096] Preferably, the module integrates an independent component analysis unit to process multi-channel signals. The received signals of each channel are modeled as linear mixed observations of source signals, and the target tissue reflection signal and probe coupling noise are separated by the demixing matrix. The fixed-point algorithm based on negative entropy maximization is used to realize blind source separation using the non-Gaussian characteristics of the signal, effectively suppressing environmental electromagnetic interference and local device noise.

[0097] In terms of system implementation, the module adopts a three-stage pipeline architecture: the time-domain processing unit performs dynamic baseline correction and data buffering; the frequency domain analysis unit completes wavelet transform and feature extraction; and the reconstruction unit is responsible for sparse representation and data compression. The data ping-pong operation is realized through the double buffering storage mechanism between the units to ensure the continuity of real-time processing. The output data after processing includes time-domain corrected signals, complex wavelet coefficients and sparse feature vectors, forming a standardized data interface for downstream modules to call.

[0098] Module 3: Dynamic motion compensation module

[0099] In this embodiment, the dynamic motion compensation module realizes accurate compensation for respiratory motion artifacts through biomechanical modeling and multi-modal data fusion technology. The module integrates a respiratory motion monitoring unit, a signal timing control unit and a parameter adaptive adjustment unit to construct a closed-loop control system, ensuring the dynamic synchronization of ultrasound imaging sequences and respiratory phases.

[0100] In a specific implementation, the construction of the biomechanical model is based on the coupling motion mechanism of the thoracic cavity and the lung. The thoracic displacement, abdominal wall motion, and airflow rate data are collected in real time by a wearable respiratory sensor array to establish a three-dimensional respiratory motion vector field. After the sensor data is denoised by Kalman filtering, it is input into a long short-term memory network to predict the respiratory phase change trend in the next 200 ms. The prediction result is fused with the spatial positioning data of the ultrasound probe to generate a probe-tissue relative motion compensation amount.

[0101] The signal emission timing control unit adopts a double-loop adjustment mechanism. The inner loop dynamically adjusts the pulse repetition frequency based on the respiratory phase angle. When the inspiratory phase acceleration period is detected, the emission frequency is automatically increased to the preset upper limit to capture rapid tissue displacement; during the expiratory phase, the baseline frequency is restored. The outer loop applies different emission delay compensation in the axial and lateral directions according to the real-time calculated pleural displacement gradient, wherein the axial compensation amount has a nonlinear relationship with the respiratory depth, and millisecond-level delay fine-tuning is achieved through a lookup table.

[0102] The parameter adaptive adjustment unit integrates multiple groups of configurable strategy libraries, including compensation parameter templates in different states such as regular breathing, deep breathing, and coughing. When a sudden change in breathing pattern is detected, the system matches the optimal template through dynamic time warping algorithm to smooth the transition of compensation parameters. Preferably, the system sets an emergency compensation mode, which automatically switches to a prediction compensation mode based on tissue elasticity characteristics when apnea or abnormal wheezing is detected, to maintain image continuity.

[0103] At the system implementation level, the modules adopt a heterogeneous computing architecture. The front-end signal preprocessing unit is deployed on the FPGA to achieve hard synchronization of sensor data and ultrasound clock signals; the middle-end model operation unit uses a multi-core DSP array to perform respiratory phase prediction, motion vector calculation, and parameter optimization in parallel; the back-end control unit sends adjustment instructions to the probe driving module through a high-speed serial bus. A three-level buffer mechanism is set in the data pipeline to ensure that the motion compensation amount calculation and the timing of ultrasound pulse emission are strictly aligned.

[0104] The formula is as follows:

[0105]

[0106] Wherein:

[0107] ΔT: ultrasound pulse emission timing compensation amount;

[0108] η: respiratory-tissue motion coupling coefficient, obtained by biomechanical model calibration;

[0109] The respiratory phase prediction value P output by the long short-term memory network lstmThe partial derivative of time t represents the instantaneous change rate of the respiratory phase.

[0110] P lstm : Respiratory phase prediction value output by the long short-term memory network;

[0111] D thorax : Real-time thoracic displacement vector, including three-dimensional space displacement components;

[0112] θ: Angle between the probe axis and the main direction of respiratory motion;

[0113] α: Probe contact pressure adaptive adjustment factor;

[0114] δ probe : Probe spatial position offset;

[0115] β: Respiratory state mutation detection coefficient, calculated by airflow rate variance within the sliding window;

[0116] t: Current system timestamp;

[0117] t0: The latest respiratory cycle start timestamp;

[0118] e: Natural logarithm base, mathematical constant.

[0119] Preferably, the module integrates a multi-modal data fusion interface, supporting the import of CT / MRI prior anatomical models. Through a non-rigid registration algorithm, the real-time ultrasound image is aligned with the preoperative three-dimensional model, and the spatial distribution of the compensation vector is constrained by the anatomical landmark points to avoid tissue structure distortion caused by excessive compensation. The compensated data stream output is a time-domain aligned radio frequency signal sequence, accompanied by respiratory phase markers and motion compensation logs for subsequent feature analysis module calls.

[0120] Module 4: Spatiotemporal feature fusion algorithm module

[0121] In this embodiment, the spatiotemporal feature fusion algorithm module realizes efficient extraction of surrounding lung cancer pleural invasion related pathological features through multi-dimensional feature tensor construction and adaptive weight fusion technology. The module receives time-frequency domain feature data streams from the signal processing module, combines displacement field information output by the dynamic motion compensation module, and constructs a feature expression system with spatiotemporal continuity.

[0122] The formula is as follows:

[0123]

[0124] Wherein:

[0125] W ijk : Fusion weight coefficient of spatial position (i, j), time slice τ, and physical parameter k;

[0126] Y: global normalization factor, calculated from the total energy of the feature tensor;

[0127] Feature saliency score of spatial location (i,j) within time window T;

[0128] F k : feature vector of the k-th physical parameter, containing time-frequency domain statistics;

[0129] t: absolute timestamp of the current frame;

[0130] t0: central time reference point of the current analysis window;

[0131] s: time decay coefficient, controlling the rate of influence decay of historical features;

[0132] exp: natural exponential function, used to construct smooth transition of time decay;

[0133] tanh: hyperbolic tangent function, a nonlinear mapping that compresses input values to the interval [-1, 1];

[0134] a: physiological parameter coupling strength adjustment factor;

[0135] C physio : real-time physiological parameter vector magnitude from biosensors.

[0136] In implementation, the feature tensor construction process includes three orthogonal dimensions: spatial dimension maps the ultrasound image pixel coordinates, time dimension relates the time-varying characteristics of consecutive frame sequences, and physical parameter dimension integrates multi-modal features such as backscatter coefficient, elastic modulus, blood flow velocity, etc. Through high-order singular value decomposition technology, the core feature subspace is extracted from the original tensor, which retains the cross-dimensional correlation characteristics and effectively captures key information such as changes in pleural interface microstructure and acoustic property abnormalities caused by tumor infiltration.

[0137] The feature fusion process adopts a dynamic weight distribution mechanism, setting adaptive fusion coefficients for different physical parameters. The backscatter feature weight is dynamically adjusted according to tissue homogeneity, reducing the weight in homogeneous areas to reduce noise interference, and increasing the weight in heterogeneous areas to enhance lesion recognition; the elastic feature weight is negatively related to the probe contact pressure, automatically decaying when the pressure exceeds the threshold to avoid measurement bias introduced by tissue deformation; the blood flow feature weight is dynamically adjusted by the signal-to-noise ratio of the Doppler signal to ensure that weak signals in low flow rate areas are not overwhelmed.

[0138] The fusion algorithm formula is as follows:

[0139]

[0140] where:

[0141] X ijk denotes the extracted feature values;

[0142] W abc denotes an element of the core tensor;

[0143] U ia denotes an element of the spatial feature matrix;

[0144] V jb denotes an element of the temporal feature matrix;

[0145] S kc denotes a weight related to a physical parameter;

[0146] A denotes a dimension of the spatial feature;

[0147] B denotes a dimension of the temporal feature;

[0148] C denotes a dimension of the physical parameter.

[0149] Preferably, the module integrates a feature enhancement unit based on clinical prior knowledge. By importing a pleural anatomy template, an attention mask is applied in the spatial dimension to strengthen the extraction of subtle features in key areas such as costophrenic angle and mediastinal pleura. A sliding analysis window is set in the temporal dimension, and a long short-term memory network is used to capture the timing rules of feature evolution to identify dynamic pathological indicators such as pleural line discontinuity and visceral pleural thickening.

[0150] At the system implementation level, the module adopts a phased processing architecture. In the preprocessing stage, the input features are standardized and missing value compensation is performed, and the feature holes caused by motion compensation failure are filled by the K-nearest neighbor algorithm. In the core processing stage, a tensor decomposition engine is deployed, and the alternating least squares method is used to iteratively solve the core tensor and factor matrix, where the spatial factor matrix is initialized by a convolutional neural network to incorporate local texture characteristics. In the post-processing stage, feature dimensionality reduction and visualization are implemented, and high-dimensional features are projected into a two-dimensional space by the t-distributed stochastic neighbor embedding algorithm to generate feature distribution maps for clinical interpretation.

[0151] In terms of data interface, the module outputs a three-channel data stream: the main channel transmits the integrated feature matrix after fusion, the auxiliary channel provides feature contribution degree heat map and abnormal area marking information, and the log channel records the weight adjustment parameters and confidence indicators in the feature fusion process. All output data are attached with time synchronization identifiers to ensure strict alignment with the ultrasound image sequence.

[0152] The module preferably incorporates a feature interpretability enhancement mechanism. Backpropagation significance analysis visualizes the contribution of each physical parameter to the final fusion result, helping clinicians understand the basis for the algorithm's decision-making. An interactive parameter adjustment interface is also integrated, allowing for fine-tuning of the spatial attention region and temporal analysis window length for specific cases, balancing algorithm sensitivity and specificity.

[0153] Module 5: Regional Cluster Analysis Module

[0154] In this example, the regional clustering analysis module uses multimodal feature clustering and dynamic neighborhood optimization technology to accurately delineate the pleural invasion area of ​​peripheral lung cancer. This module receives the high-dimensional feature matrix output by the spatiotemporal feature fusion module and combines it with the spatial topology information of the ultrasound image to construct a pathology-specific tissue partitioning model.

[0155] In specific implementation, a hybrid distance metric is employed in the feature space construction process. For backscatter features, normalized mutual information is used to capture tissue heterogeneity; for elasticity features, dynamic time warping distance is used to compensate for timing measurement deviations caused by probe pressure changes; and for blood flow features, directional consistency cosine similarity is calculated to enhance the continuity of vascular orientation. The weight coefficients of each feature subspace are adaptively adjusted using an expectation-maximization algorithm to ensure the robustness of the metric system under different imaging conditions.

[0156] Dynamic Time Warping Distance:

[0157]

[0158] in:

[0159] DTW(E i ,E j ); elastic characteristic sequence E i With E j Dynamic time warping distance;

[0160] E i : elastic characteristic time series of the i-th sampling point;

[0161] E j : elastic characteristic time series of the jth sampling point;

[0162] π: optimal alignment path, consisting of coordinate pairs that satisfy monotonicity and continuity constraints;

[0163] p: sequence E i The time point index;

[0164] q: sequence E j The time point index;

[0165] Sequence Ei measurement at time point p;

[0166] sequence E j measurement at time point q.

[0167] direction consistency cosine similarity:

[0168]

[0169] wherein:

[0170] S(v p ,v q ): blood flow direction similarity of pixel points p and q;

[0171] v p : blood flow direction vector of point p, containing speed size and azimuth angle information;

[0172] v q : blood flow direction vector of point q;

[0173] ||v p ||: the length of vector v p , i.e. the scalar value of blood flow speed;

[0174] ||v q ||: the length of vector v q .

[0175] The clustering process adopts an improved density peak algorithm to automatically determine the clustering center through local density calculation and relative distance analysis. In view of the speckle noise interference specific to the ultrasound image, the algorithm introduces anisotropic diffusion filtering pretreatment to retain the sharpness of the tissue boundary while suppressing random noise. Preferably, the system sets a dynamic neighborhood radius adjustment mechanism to automatically contract or expand the search range according to the feature space distribution density, expands the neighborhood in the homogeneous tissue area to enhance the clustering stability, and reduces the neighborhood in the lesion boundary area to improve the boundary resolution.

[0176] The regional clustering analysis module combines a deep learning method to automatically identify and label potential lesion regions using the spatial features of the ultrasound image, and the formula is as follows:

[0177]

[0178] wherein:

[0179] R represents the clustering result;

[0180] X ij represents the pixel feature in the ultrasound image;

[0181] C r represents the center point of the rth class;

[0182] C k representing the center point of other classes;

[0183] N represents the number of rows of the ultrasound image;

[0184] M represents the number of columns of the ultrasound image;

[0185] representing the class r found in all classes that best fits the current pixel feature X ij .

[0186] The spatial constraint optimization unit integrates a Markov random field model, combining the feature space clustering result with the spatial continuity of the original image. By defining an energy function minimization criterion, isolated misjudgment areas caused by feature similarity are eliminated, while the topological integrity of anatomical structures such as pleural lines and blood vessel branches is maintained. The energy function includes a unary potential term and a binary smoothing term, where the unary term reflects the feature clustering confidence, and the binary term enforces the label consistency of spatially adjacent pixels.

[0187] Preferably, the module deploys a deep feature enhancement channel. Multi-scale texture features are extracted through a pre-trained convolutional neural network, complementing the aforementioned handcrafted features. The network uses a skip-connection structure to fuse shallow detail information and deep semantic features, and suppresses the interference of irrelevant background areas through an attention gate mechanism. The fusion weights of deep features and handcrafted features are dynamically adjusted through learnable parameters, optimized end-to-end in the training phase.

[0188] At the system implementation level, the module adopts a two-stage processing flow. In the offline training phase, a standard feature dictionary is constructed from historical case data, and an initial clustering template is generated using a spectral clustering algorithm; in the online analysis phase, template parameters are fine-tuned in combination with real-time data, and the feature dictionary is updated through an incremental learning algorithm. During processing, an anomaly detection mechanism is set, which automatically triggers a temporary clustering mode based on a graph cut algorithm when the matching degree of the input features with the dictionary is below a threshold, ensuring the basic analysis capability of the system under unknown lesion types.

[0189] In terms of data output, the module generates three levels of hierarchical results: the primary output is a pixel-level clustering label map, labeling different tissue type regions; the intermediate output provides region contour vector data, including morphological parameters such as area, perimeter, and compactness; the advanced output integrates multi-frame temporal analysis results to generate a dynamic evolution map of pleural invasion progression. All result data are accompanied by confidence scores, reflecting the reliability indicators of the clustering results.

[0190] The preferred module integrates interactive correction interface. It allows the clinician to manually adjust the region boundary based on the automatic clustering result through the pen touch interaction, and the system will feed back the correction result to the feature dictionary to achieve self-learning optimization. At the same time, it provides multi-scheme comparison view and displays the analysis results under different clustering number settings side by side to assist the doctor in selecting the most clinically meaningful division scheme.

[0191] Module 6: Result display and feedback module

[0192] In this embodiment, the dynamic decision feedback module realizes the closed-loop optimization of the diagnosis and decision of peripheral lung cancer pleural invasion through multi-source information fusion and adaptive learning mechanism. The module receives multi-dimensional data stream from the feature extraction and clustering analysis module, integrates the clinical prior knowledge base, and constructs an intelligent decision system with self-evolution ability.

[0193] Feature contribution degree calculation:

[0194]

[0195] Wherein:

[0196] C f : The contribution degree of feature f in the current decision, the value range is (0, 1);

[0197] f: Feature index variable;

[0198] exp: Exponential function with natural constant e as base;

[0199] β: Temperature coefficient, positive real number, control the smoothness of the contribution degree distribution;

[0200] w f : Dynamic weight parameter of feature f, updated through online learning mechanism;

[0201] w k : Dynamic weight parameter of the kth feature, which is the same as w f , which belongs to weight vector element;

[0202] k: Feature index variable in summation operation, traverse all N features;

[0203] N: Total number of features participating in the decision process, including image features, clinical parameters, etc.;

[0204] Denominator item Sum the exponential weighted values of all N features.

[0205] In a specific implementation, the decision engine adopts a probabilistic graphical model framework to map heterogeneous data such as image features, physiological parameters, and pathological markers into a unified probabilistic inference network. Feature nodes are modeled by Gaussian mixture models to capture the distribution characteristics of different pathological states, and decision nodes use softmax functions to achieve multi-class probability output, with each class corresponding to a specific pleural invasion grade. Model parameters are iteratively optimized by the expectation-maximization algorithm, and the prior distribution is automatically updated in each diagnosis process.

[0206] Preferably, the module deploys a real-time feedback learning channel. After the system outputs the diagnosis result, the system automatically collects the manual correction records of the operating doctor and the final pathological report, and generates an error signal through comparative analysis. This signal drives the backpropagation mechanism to adjust the feature weight coefficients, and enhances the decision contribution of related features for high-frequency misjudgment case types. The learning process uses the momentum gradient descent method to accelerate convergence while avoiding local optimal traps.

[0207] In terms of knowledge base construction, the system integrates a transfer learning mechanism. By pre-training a base model on a large medical image dataset, it extracts general feature representations; then through a domain adaptation network, it maps the general feature space to the specificity space of pleural invasion diagnosis. The adaptation network uses a residual connection structure to preserve the generalization ability of the base model while enhancing the sensitivity to subtle lesions in the pleural region.

[0208] The dynamic weight adjustment unit adopts a design that combines attention mechanisms and memory networks. The attention module calculates the similarity between the current case features and historical typical cases, generating a context-related feature enhancement vector; the memory network stores the corrected diagnosis case features and provides analog reasoning support through nearest neighbor retrieval. The outputs of the two are fused by a gating mechanism and jointly act on the final decision layer.

[0209] Preferably, the module sets an uncertainty quantification subsystem. It estimates the reliability of the diagnosis result through the Monte Carlo Dropout method, and when the confidence level is below a preset threshold, it automatically triggers a multi-model voting mechanism. The voting system integrates support vector machines, random forests, and deep networks into three heterogeneous classifiers, generates a final diagnosis suggestion through a weighted majority decision rule, and dynamically allocates weight coefficients based on the historical accuracy of each model.

[0210] Uncertainty quantification:

[0211]

[0212] Where:

[0213] U: Uncertainty measure of the diagnosis result;

[0214] c: Diagnosis category index variable;

[0215] p c: Predicted probability of diagnosis category c

[0216] log: Logarithm function with base 2 for information entropy calculation

[0217] M: Total number of optional diagnosis categories, including normal tissue, local invasion, extensive invasion, etc.

[0218] Summation term Accumulate all M diagnosis categories.

[0219] At the system implementation level, the microservice architecture is used to realize the modularization of the decision-making process. The feature preprocessing service is responsible for data standardization and missing value filling; the real-time inference service deploys a lightweight neural network to ensure response speed; the model update service runs asynchronously in the background to ensure that the learning process does not affect the front-end diagnosis efficiency. The message queue is used to realize the loose coupling communication between services, supporting horizontal expansion and fault isolation.

[0220] In terms of data interface, the module outputs structured diagnostic reports and decision-making basis graphs. The report content includes invasion probability scores, key sign lists, and key points for differential diagnosis, etc. The decision-making basis graph visually displays the contribution distribution of each feature to the diagnosis conclusion, and marks the image regions with differential significance. At the same time, API interface is provided for the hospital information system to realize the automatic archiving of diagnostic data and electronic medical records.

[0221] Preferably, the module integrates a sandbox test environment. It allows medical institutions to import local case data for model fine-tuning in an isolated area, update the center model parameters through federated learning mechanism, and realize multi-center collaborative optimization under the premise of protecting patient privacy. The fine-tuning process uses differential privacy technology to add Gaussian noise when updating the model gradient to prevent information leakage of training data.

[0222] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast, characterized by, Comprise: Dual-frequency ultrasound probe, configured to work independently at different frequencies to achieve deep imaging and high-resolution imaging of lung tissue, suitable for different types of lung lesions; Signal processing module, capable of receiving, analyzing and processing echo signals emitted by the ultrasound probe, generating images in real time and extracting feature information; Dynamic motion compensation module, based on the real-time respiratory motion state of the patient, automatically adjusts the emission timing and parameter settings of the ultrasound signal; Spatiotemporal feature fusion algorithm module, through multi-dimensional feature extraction of the data obtained by the signal processing module, extracts pathological features related to surrounding lung cancer pleural invasion; Regional clustering analysis module, using clustering algorithms to analyze different tissue regions in ultrasound images, grouping adjacent tissues by similarity; Results display and feedback module, which can display the evaluation results to the operator in a visual way, and provide corresponding adjustment suggestions according to the evaluation results.

2. The ultrasound-targeted contrast-enhancement-based system for evaluating pleural invasion by peripheral lung cancer according to claim 1, characterized by, The dual-frequency ultrasound probe adopts a core-shell structure design, which ensures clear images under different physiological conditions through the penetration ability and imaging clarity of sound waves. 3.The peripheral lung cancer pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, characterized in that, The thickness of the probe shell varies according to the transmission frequency, optimizing the acoustic transmission characteristics of different levels of tissue. 4.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, characterized in that, The signal processing module includes signal separation algorithms based on time and frequency domain analysis, which can handle complex signals and prevent background noise interference.

5. The ultrasound-targeted contrast imaging-based system for evaluating pleural invasion by peripheral lung cancer according to claim 4, characterized in that, The time and frequency domain analysis method combines signal separation, descriptive model construction and image reconstruction techniques. 6.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, wherein, The dynamic motion compensation module uses biomechanical models to obtain the patient's breathing pattern in real time and adjust the emission timing and phase of the ultrasound signal in a timely manner. 7.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, wherein, The biomechanical model uses real-time respiratory frequency and amplitude information to dynamically adjust the emission parameters. 8.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, wherein, The spatiotemporal feature fusion algorithm module uses tensor decomposition techniques to extract spatial and temporal ultrasound features and comprehensively analyzes different physical parameters. The fusion algorithm formula is as follows: Where: X ijk represents the extracted feature value; W abc denotes an element of the core tensor; U ia denotes an element of the spatial feature matrix; V jb denotes an element of the time feature matrix; S kc represents a weight related to the physical parameter; A represents the dimension of spatial features; B represents the dimension of temporal features; C represents the dimension of physical parameters. 9.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeted contrast according to claim 1, wherein, Tensor decomposition techniques include selecting appropriate core tensor structures and optimizing parameter settings and feature fusion strategies. 10.The peripheral pulmonary carcinoma pleural invasion evaluation system based on ultrasound targeting contrast of claim 1, wherein, The regional clustering analysis module combines deep learning methods to automatically identify and label potential lesion areas using spatial features of ultrasound images, with the formula as follows: Where: R represents the clustering result; X ij represents a pixel feature in an ultrasound image; C r denotes the center point of the rth class; C k representing the center point of other classes; N represents the number of rows of the ultrasound image; M represents the number of columns of the ultrasound image; represents finding the class r that best fits the current pixel characteristics X in all classes. ij represents finding the class r that best fits the current pixel characteristics X in all classes.

Images