Ultrasonic image diagnosis system and method

Abstract

The present invention relates to the field of sound wave measurement technology, and specifically to an ultrasonic image diagnosis system and method. In the present invention, a deep neural network is used to enable the matching between reflection features and tissue characteristics to have adaptive learning capabilities, thereby improving the targeted recognition effect in signal classification. Based on the judgment result of the reflection feature, when performing boundary partitioning of the tissue area, a combination of three parameters: grayscale extreme difference, gradient direction and texture continuity is used to judge the edge structure, thereby avoiding contour blurring caused by the judgment of a single indicator. By extracting the grayscale distribution center, amplitude change trajectory and edge area continuous change sequence, and inputting the regional positioning result into a support vector machine, accurate identification of the nature of the lesion is achieved based on the boundary classification comparison of the morphological structure quantitative features and historical benign and malignant feature data. After the image is formed, the structural morphology is secondary verified, thereby effectively improving the diagnostic integrity and judgment confidence.

Description

Technical Field

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

[0002] The field of acoustic wave measurement technology aims to utilize the propagation characteristics of acoustic waves in different media to conduct non-contact, quantitative, and visual detection and analysis of the structure, state, and change process of physical objects. Based on the physical mechanisms of acoustic wave propagation speed, reflection, scattering, attenuation, and Doppler effect, measurement models are established and related equipment 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 diagnostic system is to construct two-dimensional or three-dimensional image information of the target area by analyzing the reflection and scattering response of tissue structure to ultrasound, achieve high-resolution, rapid imaging, clear identification of tissue boundaries and auxiliary analysis of lesion areas, improve diagnostic efficiency, and meet clinical needs for precise, quantitative and intelligent image analysis.

[0004] During the processing process, existing technologies usually rely on ultrasound imaging results to identify tissue boundaries and judge lesions. They fail to fully utilize key physical reflection characteristics such as frequency and phase at the echo signal stage, resulting in a decrease in diagnostic capabilities in the early stages of lesion formation or in areas with blurred tissue boundaries. In addition, image enhancement processing mainly relies on fixed threshold contrast adjustment, and is unable to extract differentiated details based on different tissue reflection patterns, resulting in blurred boundary contours and small lesions being easily concealed. 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 lesion development stage and limiting the reliability of benign and malignant results. Summary of the Invention

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

[0006] In order to achieve the above object, the present invention adopts the following technical solution: an ultrasonic imaging diagnostic system comprising:

[0007] Signal recognition module: This module 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 reflected signals, and uses deep neural networks to compare them with known tissue features to determine whether the signals belong to normal tissue or diseased areas, generating signal classification results.

[0008] Tissue analysis module: Based on the signal classification results, by comparing the frequency, amplitude, and phase changes, it screens out the reflection patterns that meet the physiological tissue characteristics, performs regional matching judgment, extracts the tissue areas that meet the reflection patterns, and generates optimized tissue classification results;

[0009] Image processing module: Based on the optimized tissue classification results, a multi-scale convolutional neural network is used to adjust the contrast of the target area, improve the tissue boundary contrast through adaptive contrast enhancement, and enhance the details of the lesion area to improve the regional display effect and generate a processed image;

[0010] Anomaly 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, identify abnormal areas, and generate anomaly detection results;

[0011] Diagnostic decision module: Based on the abnormality detection results and processed images, a support vector machine is used to determine whether the image contains benign lesions and malignant lesions through regional classification and comparison, and the morphological characteristics of the lesion area are analyzed. Combined with historical data comparison, the diagnosis result is confirmed to generate the lesion diagnosis result.

[0012] As a further solution of the present invention, the signal recognition module includes:

[0013] Signal extraction submodule: Acquire ultrasonic echo signals through sensors, perform time-frequency analysis on the echo signals, perform Fourier transform on the received ultrasonic echo signals, extract the frequency distribution characteristics of the signals at different time points and the energy distribution of each frequency component on the time axis, extract frequency, amplitude, and phase information, calculate the time-frequency distribution of the signals, perform segmented analysis on the signals 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 the corresponding local feature vector, and obtain the frequency, amplitude, and phase characteristics of each segment to generate reflection feature data;

[0014] Feature matching submodule: Based on the reflection feature data, the frequency, amplitude, and phase characteristics of each signal segment are compared with known physiological tissue characteristics, matching regions are selected based on feature similarity, feature comparison and screening are performed, and reflection patterns that meet physiological tissue characteristics are obtained to generate tissue reflection feature matching results;

[0015] Classification judgment submodule: Based on the tissue reflection feature matching results, a deep neural network is used to determine whether the matching area of the signal belongs to normal tissue and diseased area, and the category is divided according to the threshold to generate the signal classification result.

[0016] As a further solution of the present invention, the deep neural network is according to the formula:

[0017]

[0018] in: Represents the classification probability, which represents the possibility that the signal area is the 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 characteristic, represents the fourth reflection characteristic, represents the bias term.

[0019] As a further solution of the present invention, the tissue analysis module includes:

[0020] Reflection feature extraction submodule: Based on the signal classification results, extract frequency, amplitude, and phase changes, compare and screen reflection patterns that meet physiological tissue characteristics, perform similarity calculation and screening, and generate feature comparison results;

[0021] Region screening submodule: Based on the feature comparison results, the matching area is screened, whether the area meets the reflection mode is determined, screening and elimination operations are performed, and the region screening result is generated;

[0022] Classification optimization submodule: Based on the regional screening results, optimize the tissue classification, reclassify the region according to the clustering of features within the region, and generate optimized tissue classification results.

[0023] As a further solution of the present invention, the image processing module includes:

[0024] Contrast adjustment submodule: Based on the optimized tissue classification results, the brightness and contrast of the target area are adjusted. By calculating the brightness difference between the target area and the surrounding background, the brightness value of each pixel is adjusted to make the tissue boundary and the background more prominent, and a contrast adjustment result is generated;

[0025] Detail enhancement submodule: Based on the contrast adjustment result, edge details of the lesion area are enhanced, texture features of the lesion area are extracted, small-scale details are enhanced, and detail enhancement results are generated through local contrast enhancement and detail enhancement operations;

[0026] Image generation submodule: Based on the detail enhancement results, the entire target area is synthesized, the detail enhancement area is integrated with other parts through color balance and brightness adjustment, the display effect of the area boundary is optimized, and the processed image is generated.

[0027] As a further solution of the present invention, the multi-scale convolutional neural network is based on the formula:

[0028]

[0029] in: Represents the detail enhancement result map, represents the total number of convolution scales, Indicates the The fusion weight at each scale is represents the image contrast enhancement factor, represents the lesion area image input, represents the edge mask image, represents a pixel-by-pixel multiplication operation, represents the texture compensation coefficient, Indicates the The local texture feature map extracted at different scales, Indicates the The corresponding convolution kernel at each scale is represents a two-dimensional convolution operation, Represents the weighted fusion operation of feature maps at multiple scales.

[0030] As a further solution of the present invention, the anomaly detection module includes:

[0031] Time series analysis submodule: Based on the processed image, it extracts the time series characteristics of the signal, analyzes the signal changes at each time point, calculates the amplitude difference, frequency drift and phase mutation of the signal between consecutive time points, extracts the change rate and pattern characteristics, detects signal fluctuations, and obtains the trend of signal changes over time to generate time series feature data;

[0032] Anomaly identification submodule: Based on the time series feature data, it compares and classifies the changes in the signal, marks the abnormal changes in the signal that are different from the historical data, performs screening and identification of abnormal signals, and generates anomaly identification results;

[0033] Abnormal area calibration submodule: Based on the abnormal identification result, the detected abnormal area is located and calibrated, and the area corresponding to the abnormal signal is determined by comparing the difference between the abnormal signal and the normal signal, thereby generating the abnormal area identification result.

[0034] As a further solution of the present invention, the diagnosis decision module includes:

[0035] Regional classification submodule: Based on the abnormality detection results and the processed image, a support vector machine is used to extract features from each region in the image. By analyzing the color, texture, and edge information of the region, and comparing the contrast and brightness differences between different regions, classification judgment is made, and the diseased area and normal area are identified to generate the regional classification result;

[0036] Morphological analysis submodule: Based on the regional classification results, the boundary analysis of the lesion area is performed, the geometric morphological features of the lesion are extracted by measuring the morphological parameters of the lesion area, and the morphological similarity is calculated to generate the morphological feature analysis results;

[0037] Result confirmation submodule: Based on the morphological feature analysis results and combined with historical data comparison, the benign or malignant nature of the lesion area is judged. By comparing the morphological features of the current lesion area with the lesion features in the historical data, the possible lesion type is determined and the lesion diagnosis result is generated.

[0038] As a further solution of the present invention, the support vector machine is according to the formula:

[0039]

[0040] in: Indicates the classification result, indicating whether the ultrasound 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 color feature, Represents the average brightness value of the area, represents the weight coefficient of texture features, represents the texture complexity index, represents the bias term.

[0041] An ultrasonic image diagnosis method is provided, which is performed based on the ultrasonic image diagnosis system and includes the following steps:

[0042] S1: Based on the ultrasonic echo signal acquired by the sensor, the reflection intensity sequence of the signal in the time dimension, the energy concentration distribution in the frequency dimension, and the jump position in the phase sequence are extracted in parallel. By mapping the amplitude change gradient interval and the phase mutation index set on the same time axis, the frequency reconstruction trend trajectory and energy-intensive peak arrangement structure of the signal segment in the continuous time window are calculated. The above sequence is input into the deep neural network, and the labels are assigned to the feature channel values output by the network. The category identification vector corresponding to the signal segment is obtained, and the reflection feature segmentation result is established;

[0043] S2: Based on the reflection feature segmentation results, the corresponding frequency main component value, total number of phase jumps, and ratio of amplitude extreme values to amplitude mean are extracted from each signal segment. A four-dimensional vector is constructed and normalized to the unit feature space. The known tissue type sample feature vector group is loaded, and the position of each signal segment vector is determined. The tissue type number corresponding to each signal segment is generated and summarized to obtain the tissue type classification result;

[0044] S3: Based on the tissue type classification result, a multi-scale convolutional neural network is used to retrieve the image pixel area identified by the corresponding number on the ultrasound image frame, and the pixel grayscale average value, grayscale extreme value position density, regional edge direction change value and gradient superposition intensity distribution under the region are extracted in sequence. A regional feature map is constructed in the same space, and the gradient difference between adjacent pixel blocks in terms of the four indicators of pixel grayscale average value, grayscale extreme value position density, regional edge direction change value and gradient superposition intensity distribution is compared. The maximum edge transition intensity within each region is calculated, and contrast adjustment and enhancement operations are performed on the image boundary area to obtain tissue boundary comparison results;

[0045] S4: Based on the tissue boundary comparison results, a continuous sequence of tissue regions is selected from adjacent frames. For each region sequence, a sequence of pixel grayscale distribution center points, a principal axis of the maximum grayscale gradient direction, an edge area value, and a reflection amplitude trajectory value are extracted. A differential calculation is performed on the above values of the same region in different frames to obtain a sequence of change rates in the time dimension. The sequence matching value between the decoded output and a predefined offset pattern set is combined to determine whether an abnormality is constituted. The region whose continuous offset exceeds the set standard in multiple frames of images is identified to establish a temporal change result of the lesion.

[0046] S5: Based on the temporal change results of the lesion, all regional indexes with cross-frame offset anomalies are screened out, and the regional area value, boundary closure, pixel grayscale distribution standard deviation, and contour gradient bidirectional symmetry ratio of the corresponding region in each frame image are obtained in turn, and a multidimensional vector is constructed and input into the support vector machine. Using the benign and malignant lesion boundary function trained by the model, the position of the current regional vector on both sides of the function is compared, the morphological type is assigned to the current region, and the morphological characteristics of the lesion tissue are obtained.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are:

[0048] In the present invention, a deep neural network is used to enable adaptive learning capabilities in matching reflection features with tissue characteristics, thereby improving the targeted recognition effect in signal classification. Based on the determination results of the reflection features, when dividing the tissue area into boundary zones, a combination of three parameters, namely grayscale extremes, gradient direction, and texture continuity, is used to determine the edge structure, thus avoiding the contour blurring caused by the judgment of a single indicator.

[0049] In this invention, a multi-scale convolutional neural network is used to adjust the contrast and enhance the details of the target area in the ultrasound image. It can extract the deep features of tissue boundaries and lesion areas at different scales, significantly improve the image detail retention ability and boundary clarity, enhance the visualization effect of the lesion area, and improve the accuracy of image diagnosis and the efficiency of clinical interpretation.

[0050] In the present invention, by extracting the grayscale distribution center, amplitude change trajectory and edge area continuous change sequence and performing inter-frame offset detection, abnormal expansion behavior can be identified during the evolution process;

[0051] In the present invention, by inputting the regional positioning results into a support vector machine, accurate identification of the nature of the lesion is achieved based on the boundary classification comparison of the morphological structure quantitative characteristics and historical benign and malignant feature data, and secondary verification of the structural morphology is performed after the image is formed, which effectively improves the diagnostic integrity and judgment confidence. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, 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 intended to limit the present invention.

[0054] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0055] See also Figure 1 The present invention provides a technical solution: an ultrasonic imaging diagnostic system comprising:

[0056] Signal recognition module: This module 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 reflected signals, and uses deep neural networks to compare them with known tissue features to determine whether the signals belong to normal tissue or diseased areas, generating signal classification results.

[0057] Tissue analysis module: Based on the signal classification results, by comparing the frequency, amplitude, and phase changes, it screens out the reflection patterns that meet the physiological tissue characteristics, performs regional matching judgment, extracts the tissue areas that meet the reflection patterns, and generates optimized tissue classification results;

[0058] Image processing module: Based on the optimized tissue classification results, a multi-scale convolutional neural network is used to adjust the contrast of the target area, improve the tissue boundary contrast through adaptive contrast enhancement, and enhance the details of the lesion area to improve the regional display effect and generate a processed image;

[0059] Anomaly detection module: Based on the processed image, it analyzes the time series of the signal, extracts the signal change characteristics, identifies abnormal signals through the isolation forest method, identifies abnormal areas, and generates anomaly detection results;

[0060] Diagnostic decision module: Based on the abnormality detection results and processed images, a support vector machine is used to determine whether the image contains benign and malignant lesions through regional classification and comparison. The morphological characteristics of the lesion area are analyzed, and combined with historical data comparison, the diagnosis result is confirmed to generate the lesion diagnosis result.

[0061] The signal recognition module includes:

[0062] Signal extraction submodule: Acquire ultrasonic echo signals through sensors, perform time-frequency analysis on the echo signals, perform Fourier transform on the received ultrasonic echo signals, extract the frequency distribution characteristics of the signals at different time points and the energy distribution of each frequency component on the time axis, extract frequency, amplitude, and phase information, calculate the time-frequency distribution of the signals, perform segmented analysis on the signals 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 the corresponding local feature vector, and obtain the frequency, amplitude, and phase characteristics of each segment to generate reflection feature data;

[0063] Feature matching submodule: Based on the reflection feature data, the frequency, amplitude, and phase characteristics of each signal segment are compared with the known physiological tissue characteristics. The matching area is selected according to the feature similarity. Feature comparison and screening are performed to obtain the reflection pattern that meets the physiological tissue characteristics and generate the tissue reflection feature matching result.

[0064] Classification judgment submodule: Based on the tissue reflection feature matching results, a deep neural network is used to determine whether the matching area of the signal belongs to normal tissue and diseased area, and the classification is performed according to the threshold to generate the signal classification result;

[0065] Signal extraction submodule: Based on the ultrasonic echo signal acquired by the sensor, the short-time Fourier transform method is used to perform time-frequency analysis on the signal. This method sets the window function parameter to a Hamming window, the window length to 256, and the overlap rate to 50%. The signal is segmented into 256 sampling points. The FFT function is applied to each signal segment, the first 128 frequency component values are intercepted, and the complex-valued real and imaginary parts in the amplitude spectrum and phase spectrum are recorded to form a feature vector. The resulting feature vector is subjected to a window shift operation at the corresponding index position in the time series to generate a complete feature matrix. The matrix is further segmented according to the time dimension using a time window of 10 milliseconds in length, and three sets of values, namely the maximum frequency, average amplitude, and number of phase mutation points, are extracted in each segment to generate reflection feature data.

[0066] Feature matching submodule: Based on the reflection feature data, the Euclidean distance calculation method is used to pair the frequency, amplitude, and phase three-dimensional feature vectors of each signal segment with the known tissue sample vectors stored in the tissue feature database. For each pair of vectors, the square difference in the coordinate dimension is summed, and the square root is taken to obtain the distance value. The similarity judgment threshold is set to 0.2, and all samples with a distance less than or equal to the threshold are compared and recorded. All matching results are sorted from high to low according to confidence. The tissue number corresponding to the feature sample with the highest matching frequency is screened out, and the tissue category number to which the current segment signal belongs is marked to generate the tissue reflection feature matching result;

[0067] Classification judgment submodule: Based on the tissue reflection feature matching results, a deep neural network is used to judge the category of the signal matching area. The neural network structure is a convolutional neural network model. The network contains 3 convolutional layers. The convolution kernel sizes are set to 3×3, 5×5, and 3×3 respectively, the step size is set to 1, and the activation function is ReLU. Each convolution layer is connected to a pooling layer. The pooling method is maximum pooling, the window size is 2×2, and then two fully connected layers are connected. The number of hidden units is 128 and 64 respectively. 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 0.5 threshold, the category division result takes the one with the larger probability as the judgment output to generate the signal classification result.

[0068] Deep neural network, according to the formula:

[0069]

[0070] in: Represents the classification probability, which represents the possibility that the signal area is the 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 characteristic, represents the fourth reflection characteristic, represents the bias term;

[0071] Execution process: First, multiple reflection features are extracted from the ultrasonic image, including amplitude features , frequency characteristics , phase information and time delay characteristics , which reflects the reflection characteristics of different areas in the image, and then the weight coefficient , each eigenvalue is weighted and adjusted, and the weight coefficient reflects the importance of each feature in the classification judgment, and then the weighted eigenvalue is combined with the bias term Perform linear combination together to get an intermediate value, which is activated by the sigmoid function Mapping is performed to a classification probability between 0 and 1 , and finally according to the set threshold ,like Greater than or equal to , then the area is classified as a lesion area. Less than , it is classified as normal tissue area.

[0072] Classification is performed based on the threshold, and the threshold division mechanism adopts a dynamic setting method based on the distribution of training data. Specifically, by statistically analyzing the frequency mean, amplitude variance, phase jump density and other indicators of the corresponding reflection characteristics of various types of tissues in the labeled samples, a multidimensional distribution map is constructed in the feature space, and the minimum classification error point is extracted in the intersection area of the boundaries of various distributions as the division threshold. The threshold is not a fixed constant, but is dynamically adjusted as the overall statistical characteristics of the input sample change. The threshold is set based on the principle of maximum inter-class difference and minimum intra-class difference, and is iteratively corrected in combination with the loss feedback in the deep neural network training stage, and automatically updated according to the new training samples.

[0073] The organizational analysis module includes:

[0074] Reflection feature extraction submodule: Based on the signal classification results, it extracts frequency, amplitude, and phase changes, compares and screens reflection patterns that meet the characteristics of physiological tissues, performs similarity calculation and screening, and generates feature comparison results;

[0075] Region screening submodule: Based on the feature comparison results, the matching area is screened to determine whether the area meets the reflection mode, and the screening and elimination operations are performed to generate the region screening results;

[0076] Classification optimization submodule: Based on the regional screening results, optimize the tissue classification, reclassify the region according to the clustering of features within the region, and generate optimized tissue classification results;

[0077] Reflection feature extraction submodule: Based on the signal classification results, the Pearson correlation coefficient calculation method is used to perform dimension-by-dimension pairing operations on the three types of numerical sequences of frequency, amplitude, and phase in each region with the corresponding feature templates in the standard physiological tissue sample library. Covariance evaluation and standard deviation normalization are performed on each pairing result. The mean correlation coefficient of the three feature dimensions is used as the final similarity value. The similarity screening threshold is set to 0.85. All signal segments with correlation coefficients above this threshold are marked, and their original feature sequences are archived and saved in triple format. The maximum difference statistics are then performed on each dimension in the triple to verify feature stability and generate feature comparison results.

[0078] Region screening submodule: Based on the feature comparison results, a Boolean logic judgment method is used to perform three independent conditional judgments on whether the frequency peak position, amplitude maximum index, and number of phase mutation points in each matching area are all within the range interval 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 and perform logical AND operations. When all three are true, the area is retained; if any is false, it is eliminated. The eliminated area is marked as 0 in the record, and the retained area is marked as 1. All areas marked as 1 are sorted by index to generate the region screening results.

[0079] Classification optimization submodule: Based on the regional screening results, the K-means clustering method is used to construct a three-dimensional feature vector for the three parameters of frequency mean, amplitude standard deviation and phase mean square error of all retained regions. The initial number of cluster centers K is set to 2. After the initial allocation of cluster centers, the regional affiliation is iteratively redivided through the Euclidean distance method. The maximum number of clustering iterations is set to 100, and the convergence condition is that the cluster center movement is less than 0.001. After clustering is completed, the classification labels are reassigned to each region to generate optimized tissue classification results.

[0080] The image processing module includes:

[0081] Contrast adjustment submodule: Based on the optimized tissue classification results, the brightness and contrast of the target area are adjusted. By calculating the brightness difference between the target area and the surrounding background, the brightness value of each pixel is adjusted to make the tissue boundary and the background more prominent, generating a contrast adjustment result.

[0082] Detail enhancement submodule: Based on the contrast adjustment results, a multi-scale convolutional neural network is used to enhance the edge details of the lesion area, extract the texture features of the lesion area, enhance small-scale details, and generate detail enhancement results through local contrast enhancement and detail enhancement operations;

[0083] Image generation submodule: Based on the detail enhancement results, the entire target area is synthesized, the detail enhancement area is integrated with other parts through color balance and brightness adjustment, the display effect of the area boundary is optimized, and the processed image is generated;

[0084] Contrast adjustment submodule: 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 execution steps are as follows: first, the minimum and maximum grayscale values of all pixels in the target area are calculated to obtain the current grayscale distribution range. Then, based on the average grayscale difference between the target area and its adjacent background area, a numerical mapping table of gain adjustment factors and offset factors is established. The grayscale value of each pixel is updated according to the mapping result. Based on the row and column where the current pixel is located, gradient extension processing is performed in the eight surrounding directions in sequence. After all pixels are updated, the adjusted area is output as a grayscale matrix format, and the grayscale distribution density map in the entire area is resampled to generate the contrast adjustment result.

[0085] Detail enhancement submodule: Based on the contrast adjustment results, a multi-scale convolutional neural network and Laplace high-pass filtering method are used to enhance the edge details of the lesion area. First, a sliding analysis window with three rows and three columns is constructed within the lesion area. For each window, the grayscale difference between the central pixel and the pixels in the four directions is extracted, and the local gradient change information is marked. The edge change response intensity is recorded according to the degree of gradient change. Then, according to the index order in the response intensity table, the grayscale values of the corresponding pixels in the original image are readjusted and numerically enhanced. At the same time, the local grayscale variance value within each window is extracted. Linear stretching is performed on low-variance areas within a set range to make the image details in areas with slight grayscale changes more obvious, generating a detail enhancement result.

[0086] Image generation submodule: Based on the detail enhancement results, the weighted image fusion method is used to synthesize the image of the entire target area. First, the enhanced area and the unenhanced area are assigned unique numbers respectively, and the channel mean statistics of the red, green and blue channel grayscale values of all pixels in each numbered area are performed. The difference level between the enhanced area and the surrounding area is calculated according to the channel mean, and the fusion weight is assigned according to the level; then the grayscale values of all numbered areas are merged according to the weights, and a fusion channel grayscale matrix is generated 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 geometrically scaled using a fixed brightness adjustment coefficient. Then, the full-image brightness histogram redistribution operation is performed on the scaled channel values to generate the processed image.

[0087] The multi-scale convolutional neural network is based on the formula:

[0088]

[0089] in: Represents the detail enhancement result map, represents the total number of convolution scales, Indicates the The fusion weight at each scale is represents the image contrast enhancement factor, represents the lesion area image input, represents the edge mask image, represents a pixel-by-pixel multiplication operation, represents the texture compensation coefficient, Indicates the The local texture feature map extracted at different scales, Indicates the The corresponding convolution kernel at each scale is represents a two-dimensional convolution operation, Represents the weighted fusion operation of feature maps at multiple scales;

[0090] Implementation process: First, the lesion area image As input, combined with the edge mask Perform pixel-by-pixel multiplication to emphasize structure boundaries and suppress invalid areas, and then pass the contrast enhancement factor Amplify the brightness of the edge area to improve visual clarity, and obtain the first Texture feature map at different scales , and according to the texture compensation coefficient Add it to the image enhancement expression, then the composite image is Convolution kernel at different scales Perform convolution operations to extract detail information. Finally, the convolution responses at all scales are combined through the corresponding fusion weights. Perform weighted fusion and synthesize the detail enhancement result , thereby effectively enhancing the edges, textures and small-scale structures of the lesion area, and improving the resolution and readability of diagnostic images.

[0091] The anomaly detection module includes:

[0092] Time series analysis submodule: Based on the processed image, it extracts the time series characteristics of the signal and analyzes the signal changes at each time point. By calculating the amplitude difference, frequency drift, and phase mutation of the signal between consecutive time points, it extracts the change rate and pattern characteristics, detects signal fluctuations, and obtains the trend of signal changes over time to generate time series feature data;

[0093] Anomaly Identification Submodule: Based on time series feature data, it compares and classifies signal changes, marks abnormal changes in signals that are different from historical data, performs screening and identification of abnormal signals, and generates anomaly identification results;

[0094] Abnormal area calibration submodule: Based on the abnormal recognition results, the detected abnormal area is located and calibrated. By comparing the difference between the abnormal signal and the normal signal, the area corresponding to the abnormal signal is determined and the abnormal area recognition result is generated;

[0095] Time series analysis submodule: Based on the processed image, the long short-term memory network method is used to extract the time series features of the signal. The time series input matrix is constructed for the reflection intensity values corresponding to each pixel position in the continuous 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 weights are initialized using 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 terminal state is extracted as the encoding result in the time dimension. After smoothing all output states, the maximum change rate, average change, and fluctuation frequency indicators of each pixel position at continuous time points are extracted to generate time series feature data;

[0096] Anomaly Identification Submodule: Based on time series feature data, the isolation forest method is used to compare and classify signal changes. The maximum change rate, average change, and fluctuation frequency corresponding to each pixel are combined to construct a three-dimensional input vector. The number of trees in the forest is set to 100, the sample sampling ratio is set to 0.2, and the node splitting condition is set to maximum information gain. A decision path is constructed and the average path length of each input vector is calculated. Vectors with path lengths shorter than the average path length are marked as outliers. All outliers are marked and encoded and integrated into an abnormal pixel set according to image coordinate index to generate anomaly identification results.

[0097] Abnormal area calibration submodule: Based on the abnormal recognition results, the connected area extraction method is used to locate and calibrate the detected abnormal areas. First, all abnormal pixels are marked with 8-neighborhood expansion according to the image coordinates, and the adjacency status of each pixel in the row and column directions is recorded. A connected graph index table is constructed, and the area statistics of all connected pixel sets are performed. The connected domains with an area of less than 25 pixels are excluded. The boundary coordinates of the circumscribed rectangles of the remaining areas are calculated and stored as area identification numbers. For each numbered area, its internal signal sequence is extracted and compared with the normal signal average pattern for point-by-point difference in mean and standard deviation. The area numbers of the areas whose difference values exceed the set deviation threshold are highlighted and coded to generate the abnormal area recognition results.

[0098] The diagnostic decision module includes:

[0099] Regional classification submodule: Based on the anomaly detection results and processed images, a support vector machine is used to extract features from each region in the image. By analyzing the color, texture, and edge information of the region and comparing the contrast and brightness differences between different regions, classification judgment is made, and the diseased and normal areas are identified to generate regional classification results.

[0100] Morphological analysis submodule: Based on the regional classification results, the boundary analysis of the lesion area is performed, the morphological parameters of the lesion area are measured, the geometric morphological features of the lesion are extracted, and the morphological similarity is calculated to generate the morphological feature analysis results;

[0101] Result confirmation submodule: Based on the morphological feature analysis results and combined with historical data comparison, the lesion area is judged to be benign or malignant. By comparing the morphological features of the current lesion area with the lesion features in the historical data, the possible lesion type is determined and the lesion diagnosis result is generated;

[0102] Region classification submodule: Based on the anomaly detection results and processed images, the support vector machine method is used to extract features and classify each region in the image. First, the image is divided into non-overlapping regional blocks, each of which is 32×32 pixels in size. Six indicators are extracted from each region: the average value of the color channel, the contrast index in the grayscale co-occurrence matrix, the number of edge gradient direction changes, the standard deviation of boundary clarity, the brightness distribution variance, and the contrast range. A six-dimensional feature vector is constructed, and the radial basis kernel function is used as the kernel mapping function. The kernel function γ is set to 0.01, and the penalty factor C is set to 100. All regional feature vectors are used as the input training set. The label set comes from the existing manually annotated image region classification data. After the model training is completed, the new image regions are classified one by one, and the category label of each region is output and the coordinate index mapping is performed to generate the regional classification result;

[0103] Morphological analysis submodule: Based on the regional classification results, contour extraction and geometric calculation methods are used to perform boundary analysis and morphological parameter extraction on the lesion area. First, the outermost contour path of each lesion area is extracted to obtain the pixel coordinates of all boundary points. Based on the coordinate set, six morphological parameters of the area are calculated: perimeter, area, major and minor axis lengths, boundary concavity, shape compactness, and number of edge polylines. After the above parameters are combined to form a feature vector, the vector is normalized. The lesion area is then divided into multiple sectors. The morphological parameters of each sector are reconstructed and statistically analyzed. By calculating the Euclidean distance and structural similarity between the sectors, the overall morphological structure is coded and analyzed to generate the morphological feature analysis results.

[0104] Result confirmation submodule: Based on the results of morphological feature analysis, the feature vector comparison method is used in combination with historical data comparison to determine the benign and malignant nature of the lesion area. First, a set of morphological feature vector samples of confirmed benign and malignant lesion areas is extracted from the database. Each sample contains six standardized morphological index values of the corresponding area. The feature vector of the current lesion area is calculated in turn with each sample vector in the sample set for cosine similarity in the feature space. The output value range of the similarity calculation is limited to between 0 and 1. The classification judgment threshold is set to 0.75. When the similarity with any malignant sample exceeds this threshold, it is marked as a malignant area, otherwise it is marked as a benign area. All judgment results are numbered and coordinate mapped for output to generate the lesion diagnosis result.

[0105] Support vector machine, according to the formula:

[0106]

[0107] in: Indicates the classification result, indicating whether the ultrasound 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 color feature, Represents the average brightness value of the area, represents the weight coefficient of texture features, represents the texture complexity index, represents the bias term.

[0108] Implementation process: First, each region in the ultrasound image is extracted as a feature vector , the feature vector contains the color, texture and edge information of the region, and is obtained through the Lagrange multiplier To weight the contribution of each training sample to the classification decision boundary, the sample label Determine whether the region is a lesion area or a normal area, kernel function Used to measure input features and support vectors In order to enhance the accuracy of classification, color and texture features are introduced at the same time. The color feature is calculated by the average brightness of the region. To extract the weight coefficient Determine the impact of this feature on classification. Texture features are calculated by the texture complexity of the region. To extract the weight coefficient Determines the impact of texture features on classification, and finally the bias term Used to adjust the decision boundary and distinguish between diseased areas and normal areas.

[0109] An ultrasonic image diagnosis method is provided, which is performed based on the ultrasonic image diagnosis system and includes the following steps:

[0110] S1: Based on the ultrasonic echo signal acquired by the sensor, the reflection intensity sequence of the signal in the time dimension, the energy concentration distribution in the frequency dimension, and the jump position in the phase sequence are extracted in parallel. By mapping the amplitude change gradient interval and the phase mutation index set on the same time axis, the frequency reconstruction trend trajectory and energy-intensive peak arrangement structure of the signal segment in the continuous time window are calculated. The above sequence is input into the deep neural network, and the labels are assigned to the feature channel values output by the network. The category identification vector corresponding to the signal segment is obtained, and the reflection feature segmentation result is established;

[0111] S2: Based on the reflection feature segmentation results, the corresponding frequency main component value, total number of phase jumps, and ratio of amplitude extreme values to amplitude mean are extracted from each signal segment. A four-dimensional vector is constructed and normalized to the unit feature space. The feature vector group of known tissue type samples is loaded, and the location of each signal segment vector is determined. The tissue type number corresponding to each signal segment is generated and summarized to obtain the tissue type classification result;

[0112] S3: Based on the tissue type classification results, a multi-scale convolutional neural network is used to retrieve the image pixel area identified by the corresponding number on the ultrasound image frame. The pixel grayscale average value, grayscale extreme value position density, regional edge direction change value and gradient superposition intensity distribution in the region are extracted in sequence. A regional feature map is constructed in the same space. The gradient difference between adjacent pixel blocks in terms of the four indicators of pixel grayscale average value, grayscale extreme value position density, regional edge direction change value and gradient superposition intensity distribution is compared. The maximum edge transition intensity within each region is calculated, and contrast adjustment and enhancement operations are performed on the image boundary area to obtain tissue boundary comparison results.

[0113] S4: Based on the tissue boundary comparison results, a continuous sequence of tissue regions is selected from adjacent frames. For each region sequence, the sequence of pixel grayscale distribution center points, the main axis of the maximum grayscale gradient direction, the edge area value, and the reflection amplitude trajectory value are extracted. The above values of the same region in different frames are differentially calculated to obtain a sequence of change rates in the time dimension. The sequence matching value between the decoded output and the predefined offset pattern set is combined to determine whether it constitutes an abnormality. The region whose continuous offset exceeds the set standard in multiple frames of images is identified to establish the temporal change results of the lesion.

[0114] S5: Based on the results of temporal changes in lesions, all regional indexes with cross-frame offset anomalies are screened out, and the regional area value, boundary closure, pixel grayscale distribution standard deviation, and contour gradient bidirectional symmetry ratio of the corresponding region in each frame of the image are obtained in turn. A multidimensional vector is constructed and input into the support vector machine. Using the benign and malignant lesion boundary function trained by the model, the position of the current regional vector on both sides of the function is compared, the morphological type is assigned to the current region, and the morphological characteristics of the lesion tissue are obtained.

[0115] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An ultrasonic imaging diagnostic system, characterized in that: The system comprises: The signal recognition module includes: Signal extraction submodule: Acquires ultrasonic echo signals through sensors, performs time-frequency analysis on the received ultrasonic echo signals through Fourier transform, extracts the frequency distribution characteristics of the signal at different time points and the energy distribution of each frequency component on the time axis, calculates the time-frequency distribution of the signal, performs segmented analysis on the signal according to the time window, 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 the corresponding local feature vector and generate reflection feature data; Feature matching submodule: Based on the reflection feature data, the frequency, amplitude, and phase characteristics of each signal segment are compared with known physiological tissue characteristics, matching regions are selected based on feature similarity, feature comparison and screening are performed, and reflection patterns that meet physiological tissue characteristics are obtained to generate tissue reflection feature matching results; Classification judgment submodule: Based on the tissue reflection feature matching results, a deep neural network is used to perform weighted calculations on the frequency, amplitude, phase, and time delay features, and mapped through an activation function to obtain the classification probability. It is then determined whether the matching area of the signal belongs to the normal tissue area or the diseased tissue area, and the categories are divided according to the threshold to generate the signal classification results. Tissue analysis module: Based on the signal classification results, by comparing the frequency, amplitude, and phase changes, it screens out the reflection patterns that meet the physiological tissue characteristics, performs regional matching judgment, extracts the tissue areas that meet the reflection patterns, and generates optimized tissue classification results; an image processing module, which adjusts the brightness and contrast of the target area based on the optimized tissue classification result, calculates the brightness difference between the target area and the surrounding background, adjusts the brightness value of each pixel to make the tissue boundary and the background stand out, then uses a multi-scale convolutional neural network to enhance the edge details of the lesion area in the target area, and then performs image synthesis on the entire target area to 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 abnormal signals through the isolation forest method, identify abnormal areas, and generate abnormal area identification results; Diagnostic decision module: Based on the abnormal area identification results and the processed images, a support vector machine is used to identify the diseased area and the normal area through regional classification and comparison. The morphological characteristics of the identified diseased area are analyzed. Combined with historical data comparison, it is determined whether the image contains benign lesions and malignant lesions, and the lesion diagnosis results are generated.

2. The ultrasonic imaging diagnostic system according to claim 1, wherein: The deep neural network is based on the formula: in: Represents the classification probability, which represents the possibility that the signal area is the diseased tissue 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 frequency reflection characteristics, represents the amplitude reflection characteristic, represents the phase reflection characteristic, represents the time-delayed reflection characteristic, represents the bias term.

3. The ultrasonic imaging diagnostic system according to claim 1, wherein: The tissue analysis module includes: Reflection feature extraction submodule: Based on the signal classification results, extract frequency, amplitude, and phase changes, compare and screen reflection patterns that meet physiological tissue characteristics, perform similarity calculation and screening, and generate feature comparison results; Region screening submodule: Based on the feature comparison results, the matching area is screened, whether the area meets the reflection mode is determined, screening and elimination operations are performed, and the region screening result is generated; Classification optimization submodule: Based on the regional screening results, optimize the tissue classification, reclassify the region according to the clustering of features within the region, and generate optimized tissue classification results.

4. The ultrasonic imaging diagnostic system according to claim 1, wherein: The multi-scale convolutional neural network is based on the formula: in: Represents the detail enhancement result map, represents the total number of convolution scales, Indicates the The fusion weight at each scale is represents the image contrast enhancement factor, a lesion region image input representing the target region, represents the edge mask image, represents a pixel-by-pixel multiplication operation, represents the texture compensation coefficient, Indicates the The local texture feature map extracted at different scales, Indicates the The corresponding convolution kernel at each scale is represents a two-dimensional convolution operation, Represents the weighted fusion operation of feature maps at multiple scales; Execution process: First, input the lesion area image , combined with edge mask Perform pixel-by-pixel multiplication , emphasizing the structure boundary and suppressing the invalid area, and then by contrast enhancement factor Amplify the brightness of the edge area and obtain the first Texture feature map at different scales , and according to the texture compensation coefficient , added to the image enhancement expression, and then the composite image is combined with the Convolution kernel at different scales Perform convolution operation , extracting detail information, and finally, the convolution responses at all scales are fused through the corresponding weights Perform weighted fusion , synthesized into detail enhancement results .

5. The ultrasonic imaging diagnostic system according to claim 1, wherein: The anomaly detection module includes: Time series analysis submodule: Based on the processed image, it extracts the time series characteristics of the signal, analyzes the signal changes at each time point, calculates the amplitude difference, frequency drift and phase mutation of the signal between consecutive time points, extracts the change rate and pattern characteristics, detects signal fluctuations, and obtains the trend of signal changes over time to generate time series feature data; Anomaly identification submodule: Based on the time series feature data, it compares and classifies the changes in the signal, marks the abnormal changes in the signal that are different from the historical data, performs screening and identification of abnormal signals, and generates anomaly identification results; Abnormal area calibration submodule: Based on the abnormal identification result, the detected abnormal area is located and calibrated, and the area corresponding to the abnormal signal is determined by comparing the difference between the abnormal signal and the normal signal, thereby generating the abnormal area identification result.

6. The ultrasonic imaging diagnostic system according to claim 1, wherein: The diagnosis decision module includes: Region classification submodule: Based on the abnormal region identification results and the processed image, a support vector machine is used to extract features from each region in the image. By analyzing the color, texture, and edge information of the region, and comparing the contrast and brightness differences of different regions, classification judgment is made, and the diseased region and normal region are identified to generate the region classification result; Morphological analysis submodule: Based on the regional classification results, the identified lesion area is subjected to boundary analysis, the morphological parameters of the lesion area are measured, the geometric morphological features of the lesion are extracted, and the morphological similarity is calculated to generate morphological feature analysis results; Result confirmation submodule: Based on the morphological feature analysis results and combined with historical data comparison, the benign or malignant nature of the lesion area is judged. By comparing the morphological features of the current lesion area with the lesion features in the historical data, the possible lesion type is determined and the lesion diagnosis result is generated.

7. The ultrasonic imaging diagnostic system according to claim 6, wherein: The support vector machine is based on the formula: in: Indicates the classification result, indicating whether the processed image is a lesion area. represents the Lagrange multiplier, represents the label of the sample point, represents the kernel function, Represents the weight coefficient of color feature, Represents the average brightness value of the area, represents the weight coefficient of texture features, represents the texture complexity index, represents the bias term; Implementation process: First, each region in the ultrasound image is extracted as a feature vector , the feature vector contains the color, texture and edge information of the region, and is obtained through the Lagrange multiplier To weight the contribution of each training sample to the classification decision boundary, the sample label Determine whether the region is a lesion area or a normal area, kernel function Used to measure input features and support vectors The similarity of the color and texture features is introduced at the same time. The color feature is calculated by the average brightness of the region To extract the weight coefficient Determine the impact of features on classification. Texture features are calculated by calculating the texture complexity of the area. To extract the weight coefficient Determines the impact of texture features on classification, and finally the bias term Used to adjust the decision boundary.

Images