An ultrasound positioning optimization method, system and storage medium based on deep learning

Abstract

This invention belongs to the field of ultrasound diagnostic technology, specifically relating to a deep learning-based ultrasound localization optimization method, system, and storage medium. The method includes the following steps: real-time acquisition of ultrasound images; preprocessing the ultrasound images; inputting the preprocessed ultrasound images into a deep learning model, whereby the deep learning model identifies the main organ types in the ultrasound images; outputting optimal parameter prediction values ​​for the main organ types and the current scanning depth through a regression network; applying the optimal parameter prediction values ​​to the ultrasound device, adjusting them according to the type of optimal parameters, and then performing image processing or re-acquiring the ultrasound images. This invention also provides a system for implementing the above method. The method and system of this invention can optimize various ultrasound detection parameters without the need for external visual localization equipment, thereby improving the image quality of ultrasound imaging and showing great application potential in the field of ultrasound detection.

Description

Technical Field

[0001] This invention belongs to the field of ultrasound diagnostic technology, specifically relating to a deep learning-based ultrasound localization optimization method, system, and storage medium. Background Technology

[0002] Ultrasound diagnostic technology generates real-time images based on the reflection, scattering, and attenuation characteristics of sound waves in human tissues. Due to its advantages such as being radiation-free, portable, and repeatable, it has become a core imaging method for clinical screening and interventional guidance. The diagnostic quality of ultrasound diagnosis is highly dependent on the dynamic adaptability of equipment parameters, including transmission frequency, gain, time gain compensation (TGC), depth of focus, and Doppler range. These parameters need to be finely adjusted according to the patient's body size, lesion location, tissue characteristics, and blood flow status.

[0003] In traditional ultrasound diagnosis, organ identification and ultrasound parameter optimization still heavily rely on the operator's experience: physicians must manually identify target organs by observing tissue morphology and echo characteristics in ultrasound images, and then adjust parameters such as gain, depth, and dynamic range based on experience to optimize image quality. This process not only demands extremely high professional competence from operators, resulting in a long training period, but also suffers from the inherent subjectivity of manual judgment and parameter adjustment, easily leading to inconsistent ultrasound data acquired by different operators. Inexperienced operators struggle to quickly and accurately complete organ identification and data optimization, thus affecting diagnostic efficiency and accuracy.

[0004] In existing research on ultrasound diagnostic technology, some studies (such as Chinese patent CN 109770943 B) optimize ultrasound parameters through computer vision positioning. This relies on a camera to collect data on the patient's appearance and probe position, which is then combined with a three-dimensional coordinate model to determine the organs being examined. However, this approach has significant limitations: First, the visual data collected by the camera is prone to errors due to factors such as ambient light, patient clothing obstruction, and body posture, leading to fluctuations in positioning accuracy. Second, it requires additional configuration of visual acquisition hardware such as cameras and support frames, increasing equipment complexity and cost. Third, visual positioning and ultrasound imaging are independent data links, and their spatiotemporal synchronization requires additional calibration; otherwise, it can easily affect the collaborative efficiency of organ identification and parameter adjustment.

[0005] Therefore, there is an urgent need in this field for a technical solution that can automatically identify organs based directly on the characteristics of ultrasound images without relying on external computer vision equipment, and optimize ultrasound data in conjunction with these features, so as to simplify equipment structure, lower the operating threshold, and improve the stability and reliability of ultrasound data. Summary of the Invention

[0006] To address the problems of existing technologies, this invention provides a deep learning-based ultrasonic positioning optimization method, system, and storage medium.

[0007] A deep learning-based ultrasonic localization optimization method includes the following steps: Step 1: Acquire ultrasound images in real time and preprocess the ultrasound images; Step 2: Input the preprocessed ultrasound image into the deep learning model, which identifies the main organ types in the ultrasound image; Step 3: Output the optimal parameter prediction values ​​for the main organ type and the current scan depth through a regression network; Step 4: Apply the predicted optimal parameter values ​​to the ultrasound equipment, adjust them according to the type of optimal parameters, and then perform image processing or re-acquire the ultrasound image.

[0008] Preferably, the mode of the ultrasound image is selected from at least one of B-mode and color Doppler mode; And / or, in step 1, the preprocessing is selected from at least one of the following operations: denoising, enhancement, and region of interest extraction.

[0009] Preferably, in step 3, the optimal parameters include at least one of the following parameters: depth, total gain, dynamic range, time gain compensation curve coefficient, and grayscale mapping value.

[0010] Preferably, in step 4, the method for adjusting the optimal parameter includes at least one of the following methods: 1) The method for adjusting the depth is as follows: D_setting = max(D_min, min(D_predicted, D_max)) Where D_setting is the adjusted depth, D_predicted is the depth output by the regression network, D_min is the shallowest depth at which the probe can achieve effective imaging, and D_max is the deepest depth at which the probe can achieve effective imaging. 2) The method for adjusting time gain compensation is to compensate for the attenuation of ultrasound waves with increasing depth using the TGC curve: The TGC curve is as follows: TGC(z) = C0 + C1×z + C2×z 2 Where C0, C1, and C2 are the coefficients of the time gain compensation curve, and z is the depth; The attenuation formula is: I(z) = I0 ×e^{-μz} Where I(z) is the acoustic intensity at depth z, I0 is the initial acoustic intensity, and μ is the average attenuation coefficient of the tissue; 3) The method for adjusting the total gain is as follows: G_setting = G_predicted×TGC(z) Where G_setting is the adjusted total gain, and G_predicted is the total gain of the regression network output; 4) The method for adjusting the dynamic range is as follows: DR_setting = DR_predicted Where DR_setting is the adjusted dynamic range, and DR_predicted is the dynamic range of the regression network output; 5) The method for adjusting the grayscale curve is as follows: V_out = A×(V_in)^γ_predict Where A is the scaling factor, V_in is the normalized input pixel value, V_out is the normalized output pixel value, and γ_predict is the grayscale mapping value output by the regression network.

[0011] Preferably, the following steps are also included: Step 5: After applying the predicted optimal parameters to the ultrasound device, the image quality of the generated new image is evaluated. If the image quality does not meet expectations, the fine-tuning of the deep learning model, the fine-tuning of the regression network, or the user prompt to adjust the probe position is triggered.

[0012] Preferably, in step 5, the image quality assessment index is at least one of the following: sharpness score, feature integrity, and signal-to-noise ratio.

[0013] Preferably, the formula for the sharpness score is: SS = ΣΣ (G(x, y))² Wherein, the coordinate point (x, y) belongs to the region of interest, and G(x, y) is the gradient magnitude; The method for calculating the feature integrity is as follows: A pre-trained lightweight anatomical feature detection model is used to identify key anatomical structures that should appear in the image, and feature integrity is calculated using the following formula: FC = Number of critical structures detected / Total number of critical structures × 100%.

[0014] This invention also provides a deep learning-based ultrasonic positioning optimization system, comprising: The data acquisition module is configured to acquire ultrasound images in real time; The preprocessing module is configured to preprocess ultrasound images; An organ identification module is configured to input a preprocessed ultrasound image into a deep learning model, the deep learning model identifying the main organ types in the ultrasound image; The parameter prediction module is configured to output optimal parameter prediction values ​​for the major organ type and the current scan depth through a regression network. The parameter optimization module is configured to apply the predicted optimal parameter values ​​to the ultrasound device, adjust the parameters according to their type, and then perform image processing or re-acquire the ultrasound image.

[0015] Preferred options also include: The training module is configured to train the deep learning model and the regression network; The quality assessment module is configured to apply the optimal parameter prediction values ​​to the ultrasound device and then assess the image quality of the generated new images. If the image quality does not meet expectations, it will trigger fine-tuning of the deep learning model, fine-tuning of the regression network, or issue a user prompt to adjust the probe position.

[0016] The present invention also provides a computer-readable storage medium having stored thereon a computer program for implementing the above-described deep learning-based ultrasonic positioning optimization method.

[0017] This invention constructs a deep learning-based ultrasound localization optimization method and system. Through an integrated process of "ultrasound image feature extraction - automatic organ identification - dynamic parameter matching - data optimization output," it directly uses the raw ultrasound images acquired by the ultrasound probe as input, eliminating the need for external visual localization equipment. This achieves automatic target organ identification and precise optimization of ultrasound data. Specifically, it includes an ultrasound image preprocessing module, an organ identification model module, a parameter optimization module, and a quality assessment feedback module. These modules work together to automate the optimization of ultrasound data. This invention simplifies the required equipment and achieves excellent ultrasound localization optimization results, effectively improving the quality of ultrasound images, thus demonstrating promising application prospects.

[0018] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0019] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Detailed Implementation

[0020] It should be noted that the algorithms for data acquisition, transmission, storage and processing steps not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described, can all be implemented using content already disclosed in the prior art.

[0021] Example 1: Deep Learning-Based Ultrasonic Localization Optimization Method and System The system in this embodiment includes: The data acquisition module is configured to acquire ultrasound images in real time; The preprocessing module is configured to preprocess ultrasound images; An organ identification module is configured to input a preprocessed ultrasound image into a deep learning model, the deep learning model identifying the main organ types in the ultrasound image; The parameter prediction module is configured to output optimal parameter prediction values ​​for the major organ type and the current scan depth through a regression network. The parameter optimization module is configured to apply the predicted optimal parameter values ​​to the ultrasound device, adjust the parameters according to their type, and then perform image processing or re-acquire the ultrasound image. The training module is configured to train the deep learning model and the regression network; The quality assessment module is configured to apply the optimal parameter prediction values ​​to the ultrasound device and then assess the image quality of the generated new images. If the image quality does not meet expectations, it will trigger fine-tuning of the deep learning model, fine-tuning of the regression network, or issue a user prompt to adjust the probe position.

[0022] The specific methods for optimizing ultrasonic positioning using this system include the following steps: I. Offline Training Phase Step S101: Construct a multi-organ ultrasound image dataset.

[0023] Collect a large number of labeled ultrasound images from different models of ultrasound equipment. Each image needs to be labeled: Organ tag (y_organ): such as liver, heart, thyroid, kidney, fetus, etc.

[0024] The expert-optimized parameter set (y_params) records the scanning parameters corresponding to each organ image when it is judged as a "high-quality image" by experts. These parameters serve as the "gold standard" for model learning. Key parameters include: total gain, depth, dynamic range, temporal gain compensation curve coefficients, and grayscale mapping values.

[0025] Step S102: Design and train the organ classification model.

[0026] Model selection: Use a convolutional neural network (CNN), such as a variant of ResNet, DenseNet, or EfficientNet.

[0027] Input: Standardized preprocessed ultrasound images.

[0028] Output: A probability vector representing the probability of an image belonging to each organ type and its feature vector.

[0029] Loss function: The standard cross-entropy loss function for classification tasks. The expression is: L_class = - Σ y_organ×log(p_organ) Where p_organ is the organ probability distribution predicted by the model, and y_organ is the actual organ type.

[0030] Step S103: Design and train a parametric regression model.

[0031] Model selection: The feature vectors extracted by the classifier network are then connected to fully connected layers (FC layers) for regression prediction.

[0032] Input: Ultrasound image feature vector.

[0033] Output: A vector of continuous values, directly corresponding to the parameters that need to be optimized.

[0034] For example, a preferred vector representation is: [Gain, Depth, DR, C0, C1, C2,γ] Where C0, C1, and C2 are the polynomial coefficients (i.e., time gain compensation curve coefficients) of the fitted TGC curve, and for each depth z, the gain TGC(z) = C0 + C1 × z + C2 × z 2 Gain is the total gain, Depth is the depth, DR is the dynamic range, and γ is the grayscale mapping value.

[0035] Loss function: Mean Squared Error Loss (MSE) or Huber Loss, used to measure the difference between the predicted parameters and the expert parameters. The expression is: L_reg = Σ (y_params - p_params)^2 Where p_params is the probability distribution of the parameters predicted by the model, and y_params is the actual parameters.

[0036] Step S104: Model Deployment.

[0037] The trained model is integrated into the image processing system of the ultrasound equipment.

[0038] II. Online Inference and Optimization Phase Step 201: Acquire ultrasound images in real time and preprocess the ultrasound images.

[0039] The ultrasound image mode is selected from either B-mode or color Doppler mode.

[0040] The preprocessing is selected from at least one of the following operations: denoising, enhancement, and region of interest extraction. As a preferred approach, Gaussian filtering is first used to remove image noise, and then histogram equalization is used to enhance the contrast of echo features of tissues and organs; then, region of interest (ROI) extraction is performed, and invalid background areas at the image edges are removed by threshold segmentation, retaining the core imaging area containing human tissue, thereby reducing the amount of data for subsequent feature extraction.

[0041] Step 202: Input the preprocessed ultrasound image into a deep learning model, which identifies the main organ types in the ultrasound image.

[0042] Step 203: Output the optimal parameter prediction values ​​for the main organ type and the current scan depth through a regression network.

[0043] Step 204: Apply the predicted optimal parameter values ​​to the ultrasound equipment, adjust them according to the type of optimal parameters, and then perform image processing or re-acquire the ultrasound image.

[0044] The methods for adjusting the optimal parameters include: 1) The method for adjusting the depth is as follows: D_setting = max(D_min, min(D_predicted, D_max)) Where D_setting is the adjusted depth, D_predicted is the depth output by the regression network (i.e., the depth in the output vector), D_min is the shallowest depth at which the probe can achieve effective imaging, and D_max is the deepest depth at which the probe can achieve effective imaging. D_min and D_max are limited by the physical properties of the probe. The depth setting must ensure that the target organ is fully displayed.

[0045] 2) The method for adjusting time gain compensation is to compensate for the attenuation of ultrasound waves with increasing depth using the TGC curve: The TGC curve is as follows: TGC(z) = C0 + C1×z + C2×z 2 Where C0, C1, and C2 are the polynomial coefficients used to fit the TGC curve, and z is the depth; The attenuation formula is: I(z) = I0×e^{-μz} Where I(z) is the acoustic intensity at depth z, I0 is the initial acoustic intensity, and μ is the average attenuation coefficient of the tissue (unit: dB / cm / MHz).

[0046] The TGC curve learned by the model is essentially a compensation function TGC(z). The model defines this compensation curve by predicting polynomial coefficients, so that similar tissues at different depths are displayed with the same gray level.

[0047] 3) The method for adjusting the total gain is as follows: G_setting = G_predicted×TGC(z) Where G_setting is the adjusted total gain, and G_predicted is the total gain of the regression network output (i.e., the Gain in the output vector).

[0048] The total gain is the overall offset of the TGC curve. The model-predicted total gain G_predicted, in conjunction with TGC, ensures that the overall image brightness reaches the standard level for that organ.

[0049] 4) The method for adjusting the dynamic range is as follows: DR_setting = DR_predicted Where DR_setting is the adjusted dynamic range, and DR_predicted is the dynamic range of the regression network output (i.e., DR in the output vector).

[0050] Dynamic range determines the range of brightness in an image from its darkest to its brightest. The model makes predictions based on the dynamic range of organ echoes.

[0051] 5) The method for adjusting the grayscale curve is as follows: V_out = A ×(V_in)^γ_predict Where A is the scaling factor, V_in is the normalized input pixel value, V_out is the normalized output pixel value, and γ_predict is the gamma exponent of the regression network output (i.e., γ in the output vector).

[0052] A grayscale curve is a pixel value mapping function V_out = LUT(V_in). By adjusting γ, a specific grayscale range can be compressed or expanded to enhance contrast.

[0053] Step 205: After applying the predicted optimal parameters to the ultrasound device, the image quality of the generated new image is evaluated. If the image quality does not meet expectations, the fine-tuning of the deep learning model, the fine-tuning of the regression network, or the user prompt to adjust the probe position is triggered.

[0054] Image quality assessment metrics include at least one of the following: sharpness score, feature integrity, and signal-to-noise ratio.

[0055] For sharpness scoring, a gradient energy-based evaluation function is used. First, the gradient magnitude map G of image I is calculated using the Sobel edge detection operator. Then, the sum of squared gradient magnitudes within the target organ region is calculated as the sharpness score. The formula for the sharpness score is: SS = ΣΣ (G(x, y))² In this model, the coordinate point (x, y) belongs to the region of interest (ROI), and G(x, y) is the gradient magnitude. Finally, SS is normalized to a score of 0-100. A high score indicates sharp edges and a clear image.

[0056] The method for calculating the feature integrity is as follows: A pre-trained lightweight anatomical feature detection model (such as a small CNN or U-Net) is used to identify key anatomical structures that should appear in an image. Feature integrity is defined as the ratio of the number of detected key structures to the total number of key structures, and is calculated using the following formula: FC = Number of critical structures detected / Total number of critical structures × 100%.

[0057] Image quality assessment metrics can be set with thresholds as needed. For example, a preferred threshold setting is: sharpness score ≥ 85 points, signal-to-noise ratio ≥ 20dB, and feature integrity ≥ 90%.

[0058] As can be seen from the above embodiments, the present invention provides a method and system for optimizing ultrasonic testing parameters. This method and system can optimize various ultrasonic testing parameters without the need for external visual positioning equipment, thereby improving the image quality of ultrasonic imaging and showing great promise for application in the field of ultrasonic testing.

Claims

1. A deep learning-based ultrasound positioning optimization method, characterized in that, Includes the following steps: Step 1: Acquire ultrasound images in real time and preprocess the ultrasound images; Step 2: Input the preprocessed ultrasound image into the deep learning model, which identifies the main organ types in the ultrasound image; Step 3: Output the optimal parameter prediction values ​​for the main organ type and the current scan depth through a regression network; Step 4: Apply the predicted optimal parameter values ​​to the ultrasound equipment, adjust them according to the type of optimal parameters, and then perform image processing or re-acquire the ultrasound image.

2. The deep learning based ultrasound positioning optimization method of claim 1, wherein: The ultrasound image mode is selected from at least one of B-mode and color Doppler mode; And / or, in step 1, the preprocessing is selected from at least one of the following operations: denoising, enhancement, and region of interest extraction.

3. The deep learning based ultrasound positioning optimization method of claim 1, wherein: In step 3, the optimal parameters include at least one of the following parameters: depth, total gain, dynamic range, time gain compensation curve coefficient, and grayscale mapping value.

4. The deep learning based ultrasound positioning optimization method of claim 3, wherein: In step 4, the method for adjusting the optimal parameters includes at least one of the following methods: 1) The method for adjusting the depth is as follows: D_setting = max(D_min, min(D_predicted, D_max)) Where D_setting is the adjusted depth, D_predicted is the depth output by the regression network, D_min is the shallowest depth at which the probe can achieve effective imaging, and D_max is the deepest depth at which the probe can achieve effective imaging. 2) The method for adjusting time gain compensation is to compensate for the attenuation of ultrasound waves with increasing depth using the TGC curve: The TGC curve is as follows: TGC(z) = C0+ C1x z + C2x z 2 Where C0, C1, and C2 are the coefficients of the time gain compensation curve, and z is the depth; The attenuation formula is: I(z) = I0 ×e^{-μz} Where I(z) is the acoustic intensity at depth z, I0 is the initial acoustic intensity, and μ is the average attenuation coefficient of the tissue; 3) The method for adjusting the total gain is as follows: G_setting = G_predicted×TGC(z) Where G_setting is the adjusted total gain, and G_predicted is the total gain of the regression network output; 4) The method for adjusting the dynamic range is as follows: DR_setting = DR_predicted Where DR_setting is the adjusted dynamic range, and DR_predicted is the dynamic range of the regression network output; 5) The method for adjusting the grayscale curve is as follows: V_out = A ×(V_in)^γ_predict Where A is the scaling factor, V_in is the normalized input pixel value, V_out is the normalized output pixel value, and γ_predict is the grayscale mapping value output by the regression network.

5. The deep learning-based ultrasonic positioning optimization method according to claim 1, characterized in that, It also includes the following steps: Step 5: After applying the predicted optimal parameters to the ultrasound device, the image quality of the generated new image is evaluated. If the image quality does not meet expectations, the fine-tuning of the deep learning model, the fine-tuning of the regression network, or the user prompt to adjust the probe position is triggered.

6. The deep learning based ultrasound positioning optimization method of claim 5, wherein: In step 5, the image quality assessment metrics are at least one of the following: sharpness score, feature integrity, and signal-to-noise ratio.

7. The deep learning-based ultrasonic positioning optimization method according to claim 6, characterized in that: The formula for the sharpness score is: SS = ΣΣ (G(x, y))² Wherein, the coordinate point (x, y) belongs to the region of interest, and G(x, y) is the gradient magnitude; The method for calculating the feature integrity is as follows: A pre-trained lightweight anatomical feature detection model is used to identify key anatomical structures that should appear in the image, and feature integrity is calculated using the following formula: FC = Number of critical structures detected / Total number of critical structures × 100%.

8. A deep learning based ultrasound positioning optimization system, characterized by, include: The data acquisition module is configured to acquire ultrasound images in real time; The preprocessing module is configured to preprocess ultrasound images; An organ identification module is configured to input a preprocessed ultrasound image into a deep learning model, the deep learning model identifying the main organ types in the ultrasound image; The parameter prediction module is configured to output optimal parameter prediction values ​​for the major organ type and the current scan depth through a regression network. The parameter optimization module is configured to apply the predicted optimal parameter values ​​to the ultrasound device, adjust the parameters according to their type, and then perform image processing or re-acquire the ultrasound image.

9. The deep learning based ultrasound positioning optimization system of claim 8, wherein, Also includes: The training module is configured to train the deep learning model and the regression network; The quality assessment module is configured to apply the optimal parameter prediction values ​​to the ultrasound device and then assess the image quality of the generated new images. If the image quality does not meet expectations, it will trigger fine-tuning of the deep learning model, fine-tuning of the regression network, or issue a user prompt to adjust the probe position.

10. A computer-readable storage medium, characterized in that, It stores a computer program for implementing the deep learning-based ultrasonic positioning optimization method according to any one of claims 1-7.