Ultrasonic image diagnosis system and method

Abstract

The invention relates to the technical field of sound wave measurement, in particular to an ultrasonic image diagnosis system and method.According to the ultrasonic image diagnosis system and method, matching between reflection characteristics and tissue characteristics is made to have the self-adaptive learning ability through a deep neural network, the targeted recognition effect is improved in the aspect of signal classification, and based on the judgment result of the reflection characteristics, the diagnosis accuracy is improved. When boundary partitioning is carried out on a tissue area, the edge structure is judged by using the combination of three parameters of gray range, gradient direction and texture continuity, contour fuzziness caused by single index judgment is avoided, and the boundary partitioning accuracy is improved by extracting a gray distribution center, an amplitude change track and an edge area continuous change sequence. And the region positioning result is input into a support vector machine, accurate recognition of lesion properties is realized according to boundary classification comparison of morphological structure quantitative features and historical benign and malignant feature data, secondary verification of the structural form is performed after the image is formed, and the diagnosis integrity and the judgment confidence are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of acoustic wave measurement, and particularly to an ultrasonic image diagnosis system and method. Background Art

[0002] The technical field of acoustic wave measurement aims to utilize the propagation characteristics of acoustic waves in different media to perform non-contact, quantitative, and visual detection and analysis on the structure, state, and change process of physical objects. Based on physical mechanisms such as the propagation speed, reflection, scattering, attenuation, and Doppler effect of acoustic waves, measurement models are established and related devices and algorithms are developed to achieve high-precision acquisition and characterization of information such as displacement, thickness, material properties, defects, tissue structure, and physiological parameters.

[0003] The purpose of an ultrasonic image diagnosis system is to construct two-dimensional or three-dimensional image information of the target area by analyzing the reflection and scattering responses of tissue structures to ultrasonic waves, achieve high resolution, fast imaging, clear recognition of tissue boundaries, and auxiliary analysis of lesion areas, improve the diagnosis efficiency, and meet the clinical requirements for precise, quantitative, and intelligent image analysis.

[0004] The existing technology usually relies on the results of ultrasonic image imaging during the processing to perform tissue boundary recognition and lesion judgment, and fails to fully utilize key physical reflection features such as frequency and phase in the echo signal stage, resulting in a decline in the diagnostic ability in the initial stage of lesion formation or areas with blurred tissue boundaries. Moreover, in image enhancement processing, it mainly relies on fixed threshold contrast adjustment and cannot perform differential detail extraction according to different tissue reflection patterns, resulting in blurred boundary contours and small lesions being easily masked. There is a lack of dynamic change modeling and evolution trend tracking, resulting in the inability to identify the expansion or offset trajectory of lesions between different frames, affecting the accurate judgment of the development stage of lesions and limiting the reliability of benign and malignant results. Summary of the Invention

[0005] The purpose of the present invention is to solve the drawbacks existing in the prior art and propose an ultrasonic image diagnosis system and method.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions: An ultrasonic image diagnosis system includes: A signal recognition module: acquires ultrasonic echo signals through a sensor, extracts signal reflection features, obtains frequency, amplitude, and phase data through time-frequency analysis, calculates the time-frequency distribution of the reflection signals, and uses a deep neural network to compare with known tissue features to determine whether the signals belong to normal tissues and lesion areas, and generates signal classification results; Tissue analysis module: Based on the signal classification result, by comparing the frequency, amplitude, and phase changes, filter out the reflection patterns that conform to the physiological tissue characteristics, perform regional matching judgment, extract the tissue regions that conform to the reflection patterns, and generate an optimized tissue classification result; Image processing module: Based on the optimized tissue classification result, use a multi-scale convolutional neural network to adjust the contrast of the target region, enhance the tissue boundary contrast through adaptive contrast enhancement, and enhance the details of the lesion region to improve the regional display effect, and generate a processed image; Abnormality detection module: Based on the processed image, analyze the time series of the signal, extract the signal change characteristics, identify abnormal signals through the isolation forest method, perform abnormal region identification, and generate an abnormality detection result; Diagnostic decision-making module: Based on the abnormality detection result and the processed image, use a support vector machine to judge whether the image contains benign and malignant lesions through regional classification comparison, analyze the morphological characteristics of the lesion region, and combine historical data comparison to perform diagnostic result confirmation and generate a lesion diagnosis result.

[0007] As a further solution of the present invention, the signal recognition module includes: Signal extraction sub-module: Obtain the ultrasonic echo signal through a sensor, perform time-frequency analysis on the echo signal, perform Fourier transform on the received ultrasonic echo signal, extract the frequency distribution characteristics of the signal at different time points and the energy distribution of each frequency component on the time axis, and extract the frequency, amplitude, and phase information, calculate the time-frequency distribution of the signal, perform segmented analysis on the signal according to the time window, set the sliding time window length, divide the continuous echo signal sequence into multiple adjacent time segments, and independently calculate the frequency, amplitude, and phase parameters for each time segment to form corresponding local feature vectors, and obtain the frequency, amplitude, and phase characteristics of each segment to generate reflection feature data; Feature matching sub-module: Based on the reflection feature data, compare the frequency, amplitude, and phase characteristics of each segment of the signal with the known physiological tissue characteristics, select the matching region according to the feature similarity, perform feature comparison and screening, obtain the reflection pattern that conforms to the physiological tissue characteristics, and generate a tissue reflection feature matching result; Classification judgment sub-module: Based on the tissue reflection feature matching result, use a deep neural network to judge whether the matching region of the signal belongs to normal tissue and lesion regions, perform category division according to the threshold, and generate a signal classification result.

[0008] As a further solution of the present invention, the deep neural network, according to the formula:

[0009] Where: Represents the classification probability, indicating the likelihood that the signal region is a lesion region. Represents the sigmoid activation function. Represents the first weight coefficient. Represents the second weight coefficient. Represents the third weight coefficient. Represents the fourth weight coefficient. Represents the first reflection feature. Represents the second reflection feature. Represents the third reflection feature. Represents the fourth reflection feature. Represents the bias term.

[0010] As a further solution of the present invention, the tissue analysis module includes: Reflection feature extraction sub-module: Based on the signal classification result, extract frequency, amplitude, and phase changes, compare and screen out reflection patterns that conform to physiological tissue characteristics, perform similarity calculation and screening, and generate a feature comparison result. Region screening sub-module: Based on the feature comparison result, screen the matching regions, determine whether the regions conform to the reflection pattern, perform screening and elimination operations, and generate a region screening result. Classification optimization sub-module: Based on the region screening result, optimize tissue classification, reclassify the regions according to the clustering situation of the features within the regions, and generate an optimized tissue classification result.

[0011] As a further solution of the present invention, the image processing module includes: Contrast adjustment sub-module: Based on the optimized tissue classification result, adjust the brightness and contrast of the target region. By calculating the brightness difference between the target region and the surrounding background, adjust the brightness value of each pixel point to make the tissue boundary more prominent against the background, and generate a contrast adjustment result. Detail enhancement sub-module: Based on the contrast adjustment result, enhance the edge details of the lesion region, extract the texture features of the lesion region, enhance small-scale details, and generate a detail enhancement result through local contrast enhancement and detail strengthening operations. Image generation sub-module: Based on the detail enhancement result, perform image synthesis on the entire target region. By performing color balance and brightness adjustment, fuse the detail enhancement region with other parts, optimize the display effect of the region boundary, and generate a processed image.

[0012] As a further solution of the present invention, the multi-scale convolutional neural network follows the formula:

[0013] Where: Indicates the detailed enhancement result diagram, Indicates the total number of convolution scales, Indicates the fusion weight at the th scale, Indicates the input image of the lesion area, Indicates the edge mask diagram, Indicates the per-pixel multiplication operation, Indicates the texture compensation coefficient, Indicates the local texture feature map extracted at the th scale, Indicates the corresponding convolution kernel at the th scale, Indicates the two-dimensional convolution operation,

[0014] As a further solution of the present invention, the anomaly detection module includes: Time series analysis sub-module: Based on the processed image, extract the time series features of the signal, analyze the signal changes at each time point, extract the change rate and pattern features by calculating the amplitude difference, frequency drift and phase mutation between consecutive time points, detect signal fluctuations, and obtain the trend of the signal changing with time, generating time series feature data; Anomaly recognition sub-module: Based on the time series feature data, compare and classify the signal changes, mark the abnormal changes in the signal that are different from the historical data, perform the screening and recognition of abnormal signals, generating an anomaly recognition result; Anomaly region calibration sub-module: Based on the anomaly recognition result, locate and calibrate the detected anomaly region, determine the region corresponding to the abnormal signal by comparing the difference between the abnormal signal and the normal signal, generating an anomaly region recognition result.

[0015] As a further solution of the present invention, the diagnosis and decision-making module includes: Region classification sub-module: Based on the anomaly detection result and the processed image, use a support vector machine to extract features for each region in the image, analyze the color, texture, and edge information of the region, compare the contrast and brightness differences between different regions, perform classification judgment, identify the lesion region and the normal region, generating a region classification result; Morphological analysis sub-module: Based on the region classification result, perform boundary analysis on the lesion region, extract the geometric morphological features of the lesion by measuring the morphological parameters of the lesion region, and calculate its morphological similarity, generating a morphological feature analysis result; Result confirmation sub-module: Based on the morphological feature analysis result, combined with historical data comparison, determine the benign and malignant nature of the lesion area. By comparing the morphological features of the current lesion area with the lesion features in the historical data, judge its possible lesion type and generate a lesion diagnosis result.

[0016] As a further solution of the present invention, the support vector machine is calculated according to the formula:

[0017] Where: represents the classification result, indicating whether the ultrasonic image area is a lesion area, represents the Lagrange multiplier, represents the label of the sample point, represents the kernel function, represents the weight coefficient of the color feature, represents the average brightness value of the area, represents the weight coefficient of the texture feature, represents the texture complexity index, represents the bias term.

[0018] An ultrasonic image diagnosis method, which is executed based on the above ultrasonic image diagnosis system, includes the following steps: S1: Based on the ultrasonic echo signal obtained by the sensor, parallelly extract the reflection intensity sequence in the time dimension, the energy concentration distribution in the frequency dimension and the jump positions in the phase sequence of the signal. By corresponding the amplitude change gradient interval with the phase mutation index set under the same time axis, calculate the frequency reconstruction trend trajectory and the energy dense peak arrangement structure of the signal segment in the continuous time window. Input the above sequences into the deep neural network, assign labels to the feature channel values output by the network, obtain the category identification vector corresponding to the signal segment, and establish the reflection feature segmentation result; S2: Based on the reflection feature segmentation result, extract the corresponding frequency main component value, the total number of phase jumps, the ratio of the amplitude extreme value to the amplitude mean value from each signal segment, construct a four-dimensional vector at the same time, and uniformly normalize it to the unit feature space. Load the known tissue type sample feature vector group, judge the attribution of the vector positions of each signal segment, generate the tissue type number corresponding to each signal segment, and summarize to obtain the tissue type classification result; S3: Based on the tissue type classification result, use a multi-scale convolutional neural network to retrieve the image pixel region identified by the corresponding number on the ultrasound image frame, sequentially extract the average pixel gray value, the density of the gray extreme position, the change value of the regional edge direction, and the gradient superposition intensity distribution under the region, and construct a regional feature map in the same space. Compare the gradient drops of the four indicators of the average pixel gray value, the density of the gray extreme position, the change value of the regional edge direction, and the gradient superposition intensity distribution between adjacent pixel blocks, and calculate the maximum value of the internal edge transition intensity of each region. Perform contrast adjustment and enhancement operations on the image boundary region to obtain the tissue boundary contrast result; S4: Based on the tissue boundary contrast result, select a continuous sequence of tissue regions from adjacent frames. For each region sequence, extract its central point sequence of pixel gray distribution, the main axis of the maximum gray gradient direction, the edge area value, and the reflection amplitude trajectory value. Perform differential calculation on the above values of the same region in different frames to obtain the change rate sequence in the time dimension. Combine the sequence matching value between the decoded output and the predefined offset pattern set to determine whether an anomaly is formed, identify the regions that continuously deviate beyond the set standard in multiple-frame images, and establish the lesion temporal change result; S5: Based on the lesion temporal change result, filter out all the region indices with cross-frame offset anomalies. Sequentially obtain the region area value, the boundary closure degree, the standard deviation of the pixel gray distribution, and the contour gradient two-way symmetry ratio of the corresponding region in each frame of the image, and construct a multi-dimensional vector to input into the support vector machine. Use the well-trained benign and malignant lesion boundary function of the model to compare the attribution positions of the current region vector on both sides of the function, assign a morphological type to the current region, and obtain the morphological characteristics of the lesion tissue.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, through the deep neural network, the matching between the reflection characteristics and the tissue characteristics has an adaptive learning ability, which improves the targeted recognition effect in signal classification. Based on the determination result of the reflection characteristics, when partitioning the tissue region boundary, three parameters of gray level range, gradient direction, and texture continuity are used to judge the edge structure, avoiding contour blur caused by a single index judgment; In the present invention, a multi-scale convolutional neural network is used to perform contrast adjustment and detail enhancement on the target region in the ultrasound image, which can extract the deep features of the tissue boundary and the lesion region at different scales, significantly improving the image detail retention ability and boundary clarity, enhancing the visualization effect of the lesion region, and improving the accuracy of image diagnosis and the clinical interpretation efficiency; In the present invention, by extracting the continuous change sequences of the gray distribution center, the amplitude change trajectory, and the edge area, and performing inter-frame offset detection, the abnormal expansion behavior can be recognized during the evolution process; In the present invention, by inputting the regional positioning result into a support vector machine, and classifying and comparing the morphological structure quantization features with the historical benign and malignant feature data at the boundary, the accurate identification of the nature of the lesion is realized. After the image is formed, the structural morphology is verified for the second time, effectively improving the diagnostic integrity and determination confidence level. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the system flow chart of the present invention; Figure 2 is the schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0022] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0023] Please refer to Figure 1 , the present invention provides a technical solution: an ultrasonic image diagnosis system includes: Signal recognition module: Obtain the ultrasonic echo signal through a sensor, extract the signal reflection characteristics, obtain frequency, amplitude, and phase data through time-frequency analysis, calculate the time-frequency distribution of the reflection signal, and use a deep neural network to compare with known tissue characteristics to determine whether the signal belongs to normal tissue and the lesion area, and generate a signal classification result; Tissue analysis module: Based on the signal classification result, by comparing the frequency, amplitude, and phase changes, screen out the reflection patterns that conform to the physiological tissue characteristics, perform regional matching judgment, extract the tissue areas that conform to the reflection patterns, and generate an optimized tissue classification result; Image processing module: Based on the optimized tissue classification result, use a multi-scale convolutional neural network to adjust the contrast of the target area, enhance the tissue boundary contrast through adaptive contrast enhancement, and enhance the details of the lesion area to improve the area display effect, and generate a processed image; Anomaly detection module: Based on the processed image, analyze the time series of the signal, extract the signal change characteristics, identify the abnormal signal through the Isolation Forest method, perform anomaly region identification, and generate the anomaly detection result; Diagnostic decision-making module: Based on the anomaly detection result and the processed image, use the Support Vector Machine. Through region classification and comparison, determine whether the image contains benign and malignant lesions, analyze the morphological characteristics of the lesion area, combine with historical data comparison, perform diagnostic result confirmation, and generate the lesion diagnosis result.

[0024] The signal recognition module includes: Signal extraction sub-module: Obtain the ultrasonic echo signal through the sensor, perform time-frequency analysis on the echo signal, perform Fourier transform on the received ultrasonic echo signal, extract the frequency distribution characteristics of the signal at different time points and the energy distribution of each frequency component on the time axis, and extract frequency, amplitude, and phase information, calculate the time-frequency distribution of the signal, perform segmented analysis on the signal according to the time window, set the sliding time window length, divide the continuous echo signal sequence into multiple adjacent time segments, and calculate the frequency, amplitude, and phase parameters independently for each time segment to form the corresponding local feature vector, and obtain the frequency, amplitude, and phase characteristics of each segment, and generate the reflection feature data; Feature matching sub-module: Based on the reflection feature data, compare the frequency, amplitude, and phase characteristics of each segment of the signal with the known physiological tissue characteristics, select the matching region according to the feature similarity, perform feature comparison and screening, obtain the reflection pattern that conforms to the physiological tissue characteristics, and generate the tissue reflection feature matching result; Classification and judgment sub-module: Based on the tissue reflection feature matching result, use a deep neural network to determine whether the matching region of the signal belongs to normal tissue and lesion regions, perform category division according to the threshold, and generate the signal classification result; Signal extraction sub-module: Based on the ultrasonic echo signal obtained by the sensor, perform time-frequency analysis on the signal using the Short-Time Fourier Transform method. This method sets the window function parameter as the Hamming window, the window length is set to 256, and the overlap rate is 50%. Segment the signal every 256 sampling points, apply the FFT function to each segment of the signal, intercept the first 128 frequency component values, and record the real and imaginary parts of the complex number in its amplitude spectrum and phase spectrum to form a feature vector. Perform a window movement operation on the corresponding index position of the obtained feature vector in the time series and generate a complete feature matrix. Further segment the matrix according to the time dimension using a time window with a length of 10 milliseconds, and extract three sets of values: the maximum frequency, the average amplitude, and the number of phase mutation points in each segment to generate the reflection feature data; Feature matching sub-module: Based on the reflection feature data, the Euclidean distance calculation method is used to perform pairing operations on the three-dimensional feature vectors of frequency, amplitude, and phase of each segment of the signal and the known tissue sample vectors stored in the tissue feature database. For each pair of vectors, the sum of the squares of the coordinate dimension differences is calculated, and then the square root is taken to obtain the distance value. The similarity judgment threshold is set to 0.2. For the samples with a distance less than or equal to the threshold in all samples, comparison records are performed. All the matching results are sorted from high to low according to the confidence level, and the tissue number corresponding to the feature sample with the highest matching frequency is selected, and the tissue category number to which the current paragraph signal belongs is marked to generate the tissue reflection feature matching result; Classification judgment sub-module: Based on the tissue reflection feature matching result, a deep neural network is used to judge the category to which the signal matching area belongs. The neural network structure is a convolutional neural network model. The network contains 3 convolutional layers, and the convolutional kernel sizes are set to 3×3, 5×5, and 3×3 in sequence, the stride is set to 1, and the activation function is ReLU for all. After each convolutional layer, a pooling layer is connected. The pooling method is max pooling, and the window size is 2×2. Then two fully connected layers are connected, and the number of hidden units is 128 and 64 in sequence. The final output is the Softmax function for binary classification. The input is a matrix composed of a three-dimensional feature vector sequence, and the output is the binary classification result of normal tissue and lesion area. According to the comparison between the lesion probability component value in the output vector and the threshold of 0.5, the category division result takes the one with the larger probability as the judgment output to generate the signal classification result.

[0025] The deep neural network, according to the formula:

[0026] Where: Represents the classification probability, representing the possibility that the signal area is a lesion area, Represents the sigmoid activation function, Represents the first weight coefficient, Represents the second weight coefficient, Represents the third weight coefficient, Represents the fourth weight coefficient, Represents the first reflection feature, Represents the second reflection feature, Represents the third reflection feature, Represents the fourth reflection feature, Represents the bias term; Execution process: First, multiple reflection features are extracted from the ultrasonic image, including amplitude features , frequency features , phase information , and time delay features , which reflect the reflection characteristics of different regions in the image. Then, through the weight coefficients Each eigenvalue is weighted and adjusted, and the weight coefficient reflects the importance of each feature in the classification judgment. Then, the weighted eigenvalue and the bias term are linearly combined together to obtain an intermediate value, and the intermediate value passes through the sigmoid activation function for mapping and is converted into a classification probability between 0 and 1 . Finally, according to the set threshold , if is greater than or equal to , then the region is classified as a diseased region. If is less than , then it is classified as a normal tissue region.

[0027] The category is divided according to the threshold. The threshold division mechanism adopts a dynamic setting method based on the distribution of training data. Specifically, by statistically calculating indicators such as the frequency mean, amplitude variance, and phase jump density of the reflection characteristics corresponding to various tissues in the labeled samples, a multi-dimensional distribution map is constructed in the feature space, and the minimum classification error points are extracted at the intersection areas of various distribution boundaries as the division threshold. The threshold is not a fixed constant and dynamically adjusts with the overall statistical characteristics of the input samples. The setting of the threshold is based on the principle of maximum between-class difference and minimum within-class difference, and is iteratively corrected in combination with the loss feedback in the training stage of the deep neural network and automatically updated according to the newly added training samples.

[0028] The tissue analysis module includes: Reflection feature extraction sub-module: Based on the signal classification result, extract the frequency, amplitude, and phase changes, compare and screen out the reflection patterns that conform to the physiological tissue characteristics, perform similarity calculation and screening, and generate a feature comparison result; Region screening sub-module: Based on the feature comparison result, screen the matching regions, judge whether the regions conform to the reflection pattern, perform screening and elimination operations, and generate a region screening result; Classification optimization sub-module: Based on the region screening result, optimize the tissue classification, re-classify the regions according to the clustering situation of the features within the regions, and generate an optimized tissue classification result; Reflection feature extraction sub-module: Based on the signal classification result, use the Pearson correlation coefficient calculation method to perform pairwise operations on the frequency, amplitude, and phase numerical sequences in each region with the corresponding feature templates in the standard physiological tissue sample library dimension by dimension. For each pair of results, perform covariance evaluation and standard deviation normalization processing respectively, and use the mean of the correlation coefficients of the three feature dimensions as the final similarity value. Set the similarity screening threshold to 0.85, mark the paragraphs with a correlation coefficient higher than this threshold in all signal segments, archive and save their original feature sequences in a triple format, and then perform the maximum difference statistics on each dimension in the triple to verify the feature stability and generate a feature comparison result; Region screening sub-module: Based on the feature comparison results, use the Boolean logic judgment method to independently judge three conditions for each matching region, namely whether the frequency peak position, the amplitude maximum index, and the number of phase mutation points are all within the range given by the standard template. The frequency peak tolerance is set to ±5%, the amplitude maximum tolerance is set to ±8%, and the phase mutation point tolerance is set to ±2 points. The three conditions form Boolean expressions respectively and perform a logical AND operation. When all three are true, the region is retained; if any is false, the region is excluded. The excluded regions are marked as 0 in the record, and the retained regions are marked as 1. Combine and sort all regions marked as 1 according to the index to generate the region screening result; Classification optimization sub-module: Based on the region screening result, use the K-means clustering method to construct a three-dimensional feature vector from the three parameter combinations of the frequency mean, amplitude standard deviation, and phase mean square error of all retained regions, and set the initial number of clustering centers K to 2. After performing the initial assignment of the clustering centers, iteratively re-divide the region belonging by the Euclidean distance method. The maximum number of clustering iterations is set to 100, and the convergence condition is that the movement of the clustering center is less than 0.001. Reassign classification labels to each region after clustering is completed to generate the optimized organization classification result.

[0029] The image processing module includes: Contrast adjustment sub-module: Based on the optimized organization classification result, adjust the brightness and contrast of the target region. By calculating the brightness difference between the target region and the surrounding background, adjust the brightness value of each pixel point to make the tissue boundary more prominent against the background, and generate the contrast adjustment result; Detail enhancement sub-module: Based on the contrast adjustment result, use a multi-scale convolutional neural network to enhance the edge details of the lesion region, extract the texture features of the lesion region, and enhance small-scale details. Through local contrast enhancement and detail strengthening operations, generate the detail enhancement result; Image generation sub-module: Based on the detail enhancement result, perform image synthesis on the entire target region. By adjusting color balance and brightness, fuse the detail enhancement region with other parts to optimize the display effect of the region boundary and generate the processed image; Contrast adjustment sub-module: Based on the optimized tissue classification results, the linear contrast stretching method is used to adjust the brightness and contrast of the target area. The specific implementation steps are as follows: First, calculate the minimum and maximum gray values of all pixels in the target area to obtain the current gray distribution range. Then, according to the average gray difference between the target area and its adjacent background area, establish a numerical mapping table of the gain adjustment factor and the offset factor, update the gray value of each pixel according to the mapping result, and based on the row and column of the current pixel, perform gradient extension processing in eight directions around it in turn. After all pixel points are updated, the adjusted area is output in the format of a gray matrix, and the gray distribution density map of the entire area is resampled to generate the contrast adjustment result; Detail enhancement sub-module: Based on the contrast adjustment result, the multi-scale convolutional neural network and the Laplace high-pass filtering method are used to enhance the edge details of the lesion area. First, a three-row and three-column sliding analysis window is constructed within the lesion area, and the gray differences between the central pixel and the pixels in four directions are extracted for each window to mark the local gradient change information. The edge change response intensity is recorded according to the degree of gradient change. Subsequently, according to the index order in the response intensity table, the gray values of the corresponding pixels in the original image are readjusted and numerically enhanced. At the same time, the local gray variance value within each window is extracted, and the linear stretching process is performed on the low-variance area within the set range to make the image details in the area of small gray changes more obvious, generating the detail enhancement result; Image generation sub-module: Based on the detail enhancement result, the weighted image fusion method is used to synthesize the image of the entire target area. First, the enhanced area and the non-enhanced area are respectively assigned unique numbers, and the channel means of the gray values of the red, green, and blue channels of all pixels in each numbered area are statistically calculated. The difference level between the enhanced area and the surrounding area is calculated according to the channel means, and the fusion weight is assigned according to the level; Subsequently, all numbered areas are merged according to the weight to generate a fused channel gray matrix at each pixel point. After the fusion is completed, the overall brightness of the image is corrected, and the channel values of all pixels are scaled proportionally using a fixed brightness adjustment coefficient, and then the full-image brightness histogram redistribution operation is performed on the scaled channel values to generate the processed image.

[0030] The multi-scale convolutional neural network is calculated according to the formula:

[0031] Where: represents the detail enhancement result map, represents the total number of convolutional scales, represents the th scale of the fusion weight, represents the image contrast enhancement factor, represents the input of the lesion area image, Indicates the edge mask image, Indicates the per-pixel multiplication operation, Indicates the texture compensation coefficient, Indicates the local texture feature map extracted at the th scale, Indicates the corresponding convolution kernel at the th scale, Indicates the weighted fusion operation of feature maps at multiple scales; Execution process: First, the lesion area image is used as the input, combined with the edge mask to perform per-pixel multiplication, emphasizing the structural boundaries and suppressing the invalid areas. Subsequently, the brightness of the edge area is amplified by the contrast enhancement factor to improve the visual clarity. At the same time, the texture feature map at the th scale is obtained from the auxiliary texture extraction network, and it is added to the image enhancement expression according to the texture compensation coefficient . After that, the composite image is convolved with the convolution kernel at the th scale to extract the detailed information. Finally, the convolution responses at all scales are weighted and fused through the corresponding fusion weights to synthesize the detail enhancement result , thereby effectively strengthening the edges, textures, and small-scale structures of the lesion area, and improving the resolution and interpretability of the diagnostic image.

[0032] The anomaly detection module includes: The time series analysis sub-module: Based on the processed image, extract the time series features of the signal, analyze the signal changes at each time point, extract the change rate and pattern features by calculating the amplitude difference, frequency drift, and phase mutation between consecutive time points, detect the signal fluctuations, and obtain the trend of the signal changing over time, generating the time series feature data; The anomaly recognition sub-module: Based on the time series feature data, compare and classify the signal changes, mark the abnormal changes in the signal that are different from the historical data, perform the screening and recognition of the abnormal signals, generating the anomaly recognition result; The anomaly area calibration sub-module: Based on the anomaly recognition result, locate and calibrate the detected anomaly area, determine the area corresponding to the abnormal signal by comparing the differences between the abnormal signal and the normal signal, generating the anomaly area recognition result; Time series analysis sub-module: Based on the processed image, the long short-term memory network method is used to extract the time series features of the signal. A time series input matrix is constructed for the reflection intensity values corresponding to each pixel position in consecutive frames. The input sequence length is set to 50 frames, the stride is 1, the number of hidden units is set to 128, the activation function is tanh, and the input gate, forget gate, and output gate parameter matrices in the gating mechanism are enabled between the input layer and the hidden layer. The weight initialization adopts a uniform distribution method. When training the time series input, forward propagation is performed with a batch size of 64, the output state at each moment is recorded, and the end state is extracted as the encoding result in the time dimension. After smoothing all the output states, the maximum change rate, average change amount, and fluctuation frequency indexes of each pixel position at consecutive time points are extracted to generate time series feature data; Anomaly recognition sub-module: Based on the time series feature data, the isolation forest method is used to compare and classify the changes in the signal. The maximum change rate, average change amount, and fluctuation frequency combination corresponding to each pixel point are constructed into a three-dimensional input vector. The number of trees in the forest is set to 100, the sample sampling ratio is 0.2, and the node splitting condition is the maximum information gain. The decision path is constructed and the average path length of each input vector is calculated. The vectors with a path length shorter than the average path length are marked as anomaly points. All the anomaly points are marked and encoded, and integrated into an anomaly pixel set according to the image coordinate index to generate the anomaly recognition result; Anomaly region calibration sub-module: Based on the anomaly recognition result, the connected region extraction method is used to locate and calibrate the detected anomaly regions. First, all the anomaly pixels are marked with 8-neighborhood expansion according to the image coordinates, the adjacency status of each pixel in the row and column directions is recorded, a connected graph index table is constructed, the area of all the connected pixel sets is statistically analyzed, the connected regions with an area less than 25 pixels are excluded, the boundary coordinates of the circumscribed rectangles of the remaining regions are calculated respectively and stored as the region recognition numbers. The internal signal sequences of each numbered region are extracted and compared point by point with the average mode of the normal signal in terms of mean and standard deviation. The region numbers with difference values exceeding the set deviation threshold are marked with a highlighted code to generate the anomaly region recognition result.

[0033] The diagnostic decision module includes: Region classification sub-module: Based on the anomaly detection result and the processed image, the support vector machine is used to extract the features of each region in the image. By analyzing the color, texture, and edge information of the region, comparing the contrast and brightness differences of different regions, classification judgment is carried out to identify the diseased regions and normal regions, and generate the region classification result; Morphological analysis sub-module: Based on the region classification result, the boundary of the diseased region is analyzed. By measuring the morphological parameters of the diseased region, the geometric morphological features of the lesion are extracted, and its morphological similarity is calculated to generate the morphological feature analysis result; Result confirmation sub-module: Based on the morphological feature analysis results and combined with historical data comparison, determine the benign and malignant nature of the lesion area. By comparing the morphological features of the current lesion area with the lesion features in the historical data, judge its possible lesion type and generate a lesion diagnosis result; Region classification sub-module: Based on the anomaly detection results and the processed image, use the support vector machine method to extract features and classify each region in the image. First, divide the image into non-overlapping region blocks, each with a size of 32×32 pixels. Extract six indicators including the average value of the color channels, the contrast index in the gray-level co-occurrence matrix, the number of edge gradient direction changes, the standard deviation of boundary sharpness, the variance of brightness distribution, and the contrast range from each region to construct a six-dimensional feature vector. Use the radial basis kernel function as the kernel mapping function, set the kernel function γ to 0.01, and the penalty factor C to 100. Use all region feature vectors as the input training set, and the label set comes from the existing manually annotated image region classification data. After the model training is completed, classify each new image region one by one, output the attribution category label of each region and perform coordinate index mapping to generate a region classification result; Morphological analysis sub-module: Based on the region classification results, use contour extraction and geometric calculation methods to analyze the boundary of the lesion area and extract morphological parameters. First, extract the outermost contour path of each lesion area to obtain the pixel coordinates of all boundary points. Calculate six morphological parameters including the perimeter, area, major axis and minor axis lengths, boundary concavity rate, shape compactness, and the number of edge fold lines of the region based on the coordinate set. After combining the above parameters to construct a feature vector, standardize the vector. Subsequently, divide the lesion area into multiple fan-shaped sectors, reconstruct and statistically analyze the morphological parameters of each sector. By calculating the Euclidean distance and structural similarity value between each sector, perform encoding analysis on the overall morphological structure to generate morphological feature analysis results; Result confirmation sub-module: Based on the morphological feature analysis results, use the feature vector comparison method combined with historical data comparison to determine the benign and malignant nature of the lesion area. First, extract the morphological feature vector sample sets of the lesion areas that have been confirmed as benign and malignant from the database. Each sample contains the six standardized morphological index values of the corresponding region. Calculate the cosine similarity of the feature vector of the current lesion area with each sample vector in the sample set in the feature space in turn. The output value range of the similarity calculation is limited between 0 and 1. Set the classification judgment threshold to 0.75. If the similarity with any malignant sample exceeds this threshold, it is marked as a malignant area, otherwise it is marked as a benign area. Number and map the coordinates of all judgment results and output them to generate a lesion diagnosis result.

[0034] Support vector machine, according to the formula:

[0035] Where: Indicates the classification result, representing whether the ultrasonic image region is a lesion region. Indicates the Lagrange multiplier. Indicates the label of the sample point. Indicates the kernel function. Indicates the weight coefficient of the color feature. Indicates the average brightness value of the region. Indicates the weight coefficient of the texture feature. Indicates the texture complexity index. Indicates the bias term.

[0036] Execution process: First, each region in the ultrasonic image is extracted as a feature vector , which contains the color, texture, and edge information of the region, and the Lagrange multiplier is used to weight the contribution of each training sample to the classification decision boundary. The sample label determines whether the region is a lesion region or a normal region. The kernel function is used to measure the similarity between the input features and the support vectors . To enhance the classification accuracy, both color and texture features are introduced simultaneously. The color feature is extracted by calculating the average brightness of the region, and the weight coefficient determines the influence of this feature on classification. The texture feature is extracted by calculating the texture complexity of the region, and the weight coefficient determines the influence of the texture feature on classification. Finally, the bias term is used to adjust the decision boundary to distinguish between lesion regions and normal regions.

[0037] An ultrasonic image diagnosis method. The ultrasonic image diagnosis method is performed based on the above ultrasonic image diagnosis system and includes the following steps: S1: Based on the ultrasonic echo signal obtained by the sensor, parallel extraction is performed on the reflection intensity sequence in the time dimension, the energy concentration distribution in the frequency dimension, and the jump positions in the phase sequence of the signal. By corresponding mapping the amplitude change gradient interval and the phase mutation index set under the same time axis, the frequency reconstruction trend trajectory and the energy dense peak arrangement structure of the signal segment in the continuous time window are calculated. The above sequences are input into the deep neural network, label assignment is performed on the values of each feature channel output by the network, the category identification vector corresponding to the signal segment is obtained, and the reflection feature segmentation result is established. S2: Based on the segmentation results of reflection features, extract the corresponding frequency main component values, total number of phase jumps, ratio of the number of amplitude extrema to the amplitude mean value from each signal segment. At the same time, construct a four-dimensional vector, uniformly normalize it to the unit feature space, load the feature vector group of known tissue type samples, determine the attribution of the vector positions of each signal segment, generate the tissue type number corresponding to each segment of the signal, and summarize to obtain the tissue type classification result; S3: Based on the tissue type classification result, use a multi-scale convolutional neural network to retrieve the image pixel area identified by the corresponding number on the ultrasound image frame, sequentially extract the average pixel gray value, density of gray value extreme positions, regional edge direction change value, and gradient superposition intensity distribution under the region, and construct a regional feature map in the same space. Compare the gradient drops of the four indicators of the average pixel gray value, density of gray value extreme positions, regional edge direction change value, and gradient superposition intensity distribution between adjacent pixel blocks, and calculate the maximum value of the internal edge transition intensity of each region. Perform contrast adjustment and enhancement operations on the image boundary region to obtain the tissue boundary contrast result; S4: Based on the tissue boundary contrast result, select a continuous sequence of tissue regions from adjacent frames. Extract the sequence of central points of pixel gray value distribution, the main axis of the maximum gray gradient direction, the edge area value, and the reflection amplitude trajectory value of each region sequence. Perform differential calculation on the above values of the same region in different frames to obtain the change rate sequence in the time dimension. Combine the sequence matching value between the decoded output and the predefined offset pattern set to determine whether it constitutes an abnormality, identify the regions that continuously deviate beyond the set standard in multiple frames of images, and establish the lesion time series change result; S5: Based on the lesion time series change result, screen out all the region indexes with cross-frame offset abnormalities, sequentially obtain the region area value, boundary closure degree, standard deviation of pixel gray value distribution, and contour gradient two-way symmetry ratio of the corresponding regions under each frame of the image, and construct them into a multi-dimensional vector and input it into the support vector machine. Use the well-trained benign and malignant lesion boundary function of the model to compare the attribution positions of the current region vector on both sides of the function, assign a morphological type to the current region, and obtain the morphological characteristics of the lesion tissue.

[0038] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical content of the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still belong to the protection scope of the technical solution of the present invention.

Claims

1. An ultrasonic image diagnosis system, characterized in that, The system includes: A signal recognition module: It acquires ultrasonic echo signals through sensors, extracts signal reflection features, obtains frequency, amplitude, and phase data through time-frequency analysis, calculates the time-frequency distribution of the reflection signals, and uses a deep neural network to compare with known tissue features to determine whether the signals belong to normal tissues and lesion regions, and generates signal classification results. A tissue analysis module: Based on the signal classification results, it filters out reflection patterns that conform to physiological tissue features by comparing frequency, amplitude, and phase changes, performs regional matching judgments, extracts tissue regions that conform to the reflection patterns, and generates optimized tissue classification results. An image processing module: Based on the optimized tissue classification results, it uses a multi-scale convolutional neural network to adjust the contrast of the target region, enhances the tissue boundary contrast through adaptive contrast enhancement, and enhances the details of the lesion region to improve the regional display effect, and generates a processed image. An anomaly detection module: Based on the processed image, it analyzes the time series of the signals, extracts signal change features, identifies abnormal signals through the Isolation Forest method, performs abnormal region identification, and generates anomaly detection results. A diagnostic decision-making module: Based on the anomaly detection results and the processed image, it uses a support vector machine to determine whether the image contains benign and malignant lesions through regional classification comparison, analyzes the morphological features of the lesion regions, combines historical data comparison, and performs diagnostic result confirmation to generate lesion diagnostic results.

2. The ultrasonic image diagnostic system according to claim 1, wherein The signal recognition module includes: A signal extraction sub-module: It acquires ultrasonic echo signals through sensors, performs time-frequency analysis on the echo signals, performs Fourier transform on the received ultrasonic echo signals, extracts the frequency distribution features of the signals at different time points and the energy distribution of each frequency component on the time axis, extracts frequency, amplitude, and phase information, calculates the time-frequency distribution of the signals, analyzes the signals in segments according to time windows, sets the sliding time window length, divides the continuous echo signal sequence into multiple adjacent time segments, and independently calculates the frequency, amplitude, and phase parameters for each time segment to form corresponding local feature vectors, and obtains the frequency, amplitude, and phase features of each segment, generating reflection feature data. A feature matching sub-module: Based on the reflection feature data, it compares the frequency, amplitude, and phase features of each segment of the signals with known physiological tissue features, selects matching regions according to feature similarity, performs feature comparison and screening, obtains reflection patterns that conform to physiological tissue features, and generates tissue reflection feature matching results. A classification judgment sub-module: Based on the tissue reflection feature matching results, it uses a deep neural network to determine whether the matching regions of the signals belong to normal tissues and lesion regions, and performs category division according to thresholds, generating signal classification results.

3. The ultrasonic image diagnosis system according to claim 2, wherein, The deep neural network, according to the formula: Wherein: represents the classification probability, indicating the possibility that the signal region is a lesion region, represents the sigmoid activation function, represents the first weight coefficient, represents the second weight coefficient, represents the third weight coefficient, represents the fourth weight coefficient, represents the first reflection feature, represents the second reflection feature, represents the third reflection feature, represents the fourth reflection feature, represents the bias term.

4. The ultrasonic image diagnosis system according to claim 1, characterized in that, The tissue analysis module includes: A reflection feature extraction sub-module: Based on the signal classification results, it extracts frequency, amplitude, and phase changes, compares and filters out reflection patterns that conform to physiological tissue features, performs similarity calculation and screening, and generates feature comparison results. Region screening sub-module: Based on the feature comparison result, screen the matching regions, determine whether the regions meet the reflection mode, perform screening and elimination operations, and generate a region screening result; Classification optimization sub-module: Based on the region screening result, optimize the tissue classification, re-classify the regions according to the clustering of features within the regions, and generate an optimized tissue classification result.

5. The ultrasonic image diagnostic system according to claim 1, wherein The image processing module includes: Contrast adjustment sub-module: Based on the optimized tissue classification result, adjust the brightness and contrast of the target region. By calculating the brightness difference between the target region and the surrounding background, adjust the brightness value of each pixel point to make the tissue boundary and the background prominent, and generate a contrast adjustment result; Detail enhancement sub-module: Based on the contrast adjustment result, use a multi-scale convolutional neural network to enhance the edge details of the lesion region, extract the texture features of the lesion region, enhance small-scale details, and generate a detail enhancement result through local contrast enhancement and detail strengthening operations; Image generation sub-module: Based on the detail enhancement result, perform image synthesis on the entire target region, fuse the detail enhancement region with other parts through color balance and brightness adjustment, optimize the display effect of the region boundary, and generate a processed image.

6. The ultrasonic image diagnosis system according to claim 5, characterized in that, The multi-scale convolutional neural network is in accordance with the formula: Among them: represents the detailed enhancement result image, represents the total number of convolution scales, represents the fusion weight at the image contrast enhancement factor, represents the input image of the lesion area, represents the edge mask image, represents the per-pixel multiplication operation, represents the texture compensation coefficient, represents the local texture feature map extracted at the represents the corresponding convolution kernel at the represents the two-dimensional convolution operation, represents the weighted fusion operation of feature maps at multiple scales.

7. The ultrasonic image diagnostic system according to claim 1, wherein The anomaly detection module includes: Time series analysis sub-module: Based on the processed image, extract the time series features of the signal, analyze the signal changes at each time point, extract the change rate and pattern features by calculating the amplitude difference, frequency drift, and phase mutation between consecutive time points, detect signal fluctuations, and obtain the trend of the signal changing over time, and generate time series feature data; Anomaly recognition sub-module: Based on the time series feature data, compare and classify the signal changes, mark the abnormal changes in the signal that are different from the historical data, perform screening and recognition of abnormal signals, and generate an anomaly recognition result; Anomaly region calibration sub-module: Based on the anomaly recognition result, locate and calibrate the detected anomaly regions. By comparing the differences between the abnormal signals and the normal signals, determine the regions corresponding to the abnormal signals, and generate an anomaly region recognition result.

8. The ultrasonic image diagnostic system according to claim 1, wherein The diagnosis decision module includes: Region classification sub-module: Based on the anomaly detection result and the processed image, use a support vector machine to extract the features of each region in the image. By analyzing the color, texture, and edge information of the regions, comparing the contrast and brightness differences between different regions, perform classification judgment, identify the lesion regions and normal regions, and generate a region classification result; Morphological analysis sub-module: Based on the region classification result, perform boundary analysis on the lesion regions. By measuring the morphological parameters of the lesion regions, extract the geometric morphological features of the lesions, and calculate their morphological similarity, and generate a morphological feature analysis result; Result confirmation sub-module: Based on the morphological feature analysis result, combined with historical data comparison, determine the benign and malignant nature of the lesion regions. By comparing the morphological features of the current lesion regions with the lesion features in the historical data, judge the possible lesion types, and generate a lesion diagnosis result.

9. The ultrasonic image diagnostic system according to claim 1, wherein The support vector machine is in accordance with the formula: Wherein: represents the classification result, indicating whether the ultrasonic image area is a lesion area, represents the Lagrange multiplier, represents the label of the sample point, represents the kernel function, represents the weight coefficient of the color feature, represents the average brightness value of the area, represents the weight coefficient of the texture feature, represents the texture complexity index, represents the bias term.

10. An ultrasonic image diagnosis method, characterized in that, Execute according to the ultrasonic image diagnosis system described in any one of claims 1-9, including the following steps: S1: Based on the ultrasonic echo signals obtained by the sensor, parallel extraction is performed on the reflection intensity sequence in the time dimension, the energy concentration distribution in the frequency dimension, and the jump positions in the phase sequence of the signals. By corresponding mapping the amplitude change gradient interval and the phase mutation index set under the same time axis, calculate the frequency reconstruction trend trajectory and the energy dense peak arrangement structure of the signal segment within a continuous time window. Input the above sequences into a deep neural network, assign labels to the values of each feature channel output by the network, obtain the category identification vector corresponding to the signal segment, and establish the reflection feature segmentation result; S2: Based on the reflection feature segmentation result, extract the corresponding frequency main component value, total number of phase jumps, ratio of amplitude extreme value number to amplitude mean value from each signal segment, construct a four-dimensional vector at the same time, and uniformly normalize it to the unit feature space. Load the known tissue type sample feature vector group, judge the attribution of the vector positions of each signal segment, generate the tissue type number corresponding to each signal segment, and summarize to obtain the tissue type classification result; S3: Based on the tissue type classification result, use a multi-scale convolutional neural network to retrieve the image pixel area identified by the corresponding number on the ultrasonic image frame, sequentially extract the average pixel gray value, gray extreme position density, regional edge direction change value, and gradient superposition intensity distribution under the area, and construct a regional feature map in the same space. Compare the gradient drops of the four indicators of average pixel gray value, gray extreme position density, regional edge direction change value, and gradient superposition intensity distribution between adjacent pixel blocks, and calculate the maximum value of the internal edge transition intensity of each area. Perform contrast adjustment and enhancement operations on the image boundary area to obtain the tissue boundary contrast result; S4: Based on the tissue boundary contrast result, select a continuous sequence of tissue areas from adjacent frames, extract the central point sequence of pixel gray distribution, the main axis of the maximum gray gradient direction, the edge area value, and the reflection amplitude trajectory value of each area sequence. Perform differential calculation on the above values of the same area in different frames to obtain the change rate sequence in the time dimension. Combine the sequence matching value between the decoded output and the predefined offset mode set to judge whether it constitutes an abnormality, and identify the areas with continuous offsets exceeding the set standard in multiple-frame images to establish the lesion time series change result; S5: Based on the lesion time series change result, screen out all the area indexes with cross-frame offset abnormalities, sequentially obtain the area value, boundary closure degree, standard deviation of pixel gray distribution, and contour gradient bidirectional symmetry ratio of the corresponding areas under each frame of image, and construct a multi-dimensional vector and input it into a support vector machine. Use the well-trained benign and malignant lesion boundary function of the model to compare the attribution positions of the current area vector on both sides of the function, assign a morphological type to the current area, and obtain the morphological characteristics of the lesion tissue.

Images