Ophthalmology ultrasonic RF signal region optimization method, model training method and device

Abstract

The invention discloses an ophthalmologic ultrasound RF signal region optimization method and device and a model training method and device, and relates to the technical field of ophthalmologic ultrasound and the technical field of deep learning, and the method comprises the steps: obtaining to-be-optimized RF signal data of ophthalmologic ultrasound of a target object; the method comprises the following steps: inputting to-be-optimized RF signal data into a pre-trained semantic perception segmentation model, identifying eye tissues to which sampling points in the to-be-optimized RF signal data belong, and outputting a probability graph by the semantic perception segmentation model; and inputting the RF signal data to be optimized and the probability graph into a pre-trained signal optimization generation model to obtain optimized RF signal data output by the signal optimization generation model. By adopting the scheme, the missing of key tissue edges and details in ophthalmology ultrasonic imaging is reduced, and the image quality of the ultrasonic image is improved.

Description

Technical Field

[0001] This application relates to the fields of ophthalmic ultrasound technology and deep learning technology, and in particular to an ophthalmic ultrasound RF signal region optimization method, model training method and device. Background Technology

[0002] Ultrasound imaging is an indispensable non-invasive examination method in modern medical diagnosis. Its basic principle is to use the piezoelectric effect of the ultrasound probe to convert (A / D conversion) the reflected, scattered and attenuated signals generated when the mechanical sound waves emitted by the probe penetrate human tissue into analog or digital electrical signals.

[0003] The imaging algorithm then reads these time-series RF signals and reconstructs the image through a series of processes, such as beamforming, envelope detection, logarithmic compression, and scan conversion. Finally, the algorithm maps these electrical signals into B-mode grayscale images based on the intensity of the tissue echo signals, for clinicians to use in diagnosis. In this image, different tissues exhibit different brightness levels based on their acoustic impedance differences: for example, fluids (such as aqueous humor in the eye) typically appear as echo-free black areas because they reflect almost no sound waves; while denser tissues (such as muscles, ligaments, or diseased tissues) appear as white or gray areas of varying brightness due to their stronger reflectivity.

[0004] Different human tissues and organs have drastically different acoustic characteristics and morphological structures. Therefore, general-purpose ultrasound signal processing algorithms suitable for deep, low-frequency imaging (such as abdominal and obstetric imaging) cannot directly meet the needs of all clinical departments, especially in the field of imaging superficial organs requiring high resolution, such as ophthalmology.

[0005] Ophthalmic ultrasound imaging, particularly ultrasound biomicroscopy (UBM) of the anterior segment, faces unique challenges. Clinically, the diagnosis and surgical planning of diseases such as glaucoma require precise observation of delicate, superficial structures such as the suspensory ligaments and ciliary body surrounding the lens. These structures are tiny and have low acoustic contrast, placing extremely high demands on the resolution and contrast of imaging algorithms.

[0006] The main drawback of existing technical solutions is that conventional imaging algorithms cannot meet the high-resolution requirements of ophthalmology. Currently, the post-processing of conventional ultrasound imaging algorithms widely used in clinical practice usually focuses on general signal processing steps, such as beamforming (e.g., delay-summation, DAS) and averaging of the overall signal, basic dynamic range compression, and general noise reduction (e.g., speckle suppression).

[0007] While these conventional algorithms are sufficient for gynecological and abdominal ultrasound imaging needs due to the larger target tissues and relatively lower resolution requirements, when applied to ocular structures, they still employ a global approach to signal processing, blurring crucial tissue edges and details. The complex echo signals generated by the unique anatomical structures of the eye (such as the aqueous humor, lens, and suspensory ligaments) cannot be clearly reproduced using these conventional algorithms, resulting in poor image quality that fails to meet the clinical diagnostic needs of ophthalmology, particularly the anterior segment.

[0008] In summary, existing technologies suffer from the problem of missing key tissue details and low image quality in ophthalmic ultrasound imaging. Summary of the Invention

[0009] This application provides an ophthalmic ultrasound RF signal region optimization method, model training method, and apparatus to solve the problem in the prior art that key tissue edges and details are easily missing in ophthalmic ultrasound imaging, resulting in low image quality.

[0010] This application provides a method for optimizing ophthalmic ultrasound RF signal regions, including:

[0011] Acquire the RF signal data of ophthalmic ultrasound of the target subject to be optimized;

[0012] By inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model, the ocular tissue to which each sampling point in the RF signal data to be optimized belongs is identified. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various ocular tissues. The first training data of the semantic perception segmentation model includes: the first training RF signal data collected, and the tissue label of the sampling point in the first training RF signal data, wherein the tissue label represents the ocular tissue to which the sampling point belongs.

[0013] The RF signal data to be optimized and the probability map are input into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

[0014] Furthermore, after acquiring the RF signal data to be optimized from the ophthalmic ultrasound of the target object, the method further includes:

[0015] For each sampling point in the RF signal data to be optimized, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the RF signal data of the sampling point and each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0016] The step of inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs includes:

[0017] By inputting the multidimensional RF tensor of each sampling point in the RF signal data to be optimized into a pre-trained semantic perception segmentation model, the ocular tissue to which each sampling point in the RF signal data to be optimized belongs is identified. The semantic perception segmentation model outputs a probability map of each multidimensional RF tensor, which contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues.

[0018] The step of inputting the RF signal data to be optimized and the probability map into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model includes:

[0019] The multidimensional RF tensor and the probability map of each sampling point in the RF signal data to be optimized are input into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0020] Furthermore, the tissue labels of the sampling points in the first training RF signal data are determined using the following steps:

[0021] Based on the first training RF signal data, an initial B-mode grayscale image is generated;

[0022] The initial B-mode grayscale image is semantically segmented to generate an anatomical structure mask map, which includes multiple regions divided according to the ocular tissue structure and tissue labels corresponding to each region.

[0023] For each sampling point in the first training RF signal data, the pixel point corresponding to the sampling point in the initial B-mode grayscale image is determined according to the pre-constructed coordinate lookup table. The coordinate lookup table is a non-linear mapping relationship between the pixel points in the B-mode grayscale image and the sampling points in the RF signal data.

[0024] The tissue label of the pixel corresponding to the sampling point in the anatomical structure mask image is determined as the tissue label of the sampling point.

[0025] Furthermore, the coordinate lookup table is constructed using the following steps:

[0026] Obtain the equipment parameters of the hardware used for ophthalmic ultrasound examination, as well as the acquisition parameters used for signal acquisition;

[0027] Based on the device parameters and the acquisition parameters, a Cartesian coordinate system for the B-mode grayscale image is established.

[0028] For each pixel in a B-mode grayscale image The pixel is calculated using the following formula. The corresponding sampling point in the RF signal data :

[0029] ,in, For the depth information of the sampling points, Sampling frequency, This represents the speed at which ultrasound waves propagate through the tissues of the eye. The ordinate value of the pixel;

[0030] ,in, For the band information of the sampling point, The element spacing of the probe in the hardware device. The x-coordinate value of the pixel;

[0031] Based on the calculation of each pixel With sampling points Establish the correspondence between them and construct a coordinate lookup table.

[0032] Furthermore, the first loss function used for training the semantic-aware segmentation model is expressed by the following formula:

[0033] ;

[0034] ;

[0035] in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. Indicates the tissue label of the i-th sampling point;

[0036] The second loss function used for training the signal optimization generation model is expressed by the following formula:

[0037] ;

[0038] in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform;

[0039] The overall loss function of the semantic-aware segmentation model and the signal optimization generation model is expressed by the following formula:

[0040] ;

[0041] in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively.

[0042] This application also provides a method for training a signal optimization model for optimizing ophthalmic ultrasound RF signal regions. The signal optimization model includes a semantic-aware segmentation model and a signal optimization generation model. The method includes:

[0043] Acquire first training data for training a semantic-aware segmentation model. The first training data includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the ocular tissue to which the sampling points belong.

[0044] The first training RF signal data is used as the input to the semantic-aware segmentation model to train the semantic-aware segmentation model;

[0045] Acquire second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, as target reference RF signal data corresponding to the second training RF signal data.

[0046] The second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data is input into the signal optimization generation model to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

[0047] Furthermore, the step of using the first training RF signal data as input to the semantic-aware segmentation model to train the semantic-aware segmentation model includes:

[0048] The first training RF signal data is input into the semantic perception segmentation model to obtain the probability map output by the semantic perception segmentation model. The probability map contains the probability that each sampling point in the first training RF signal data belongs to various ocular tissues.

[0049] Based on the probabilities contained in the probability map and the organization labels of the sampling points in the first training RF signal data, the first loss function is calculated;

[0050] The semantic-aware segmentation model is trained based on the calculation results of the first loss function.

[0051] Furthermore, after obtaining the first training data for training the semantically aware segmentation model, the method further includes:

[0052] For each sampling point in the first training RF signal data, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of that sampling point.

[0053] The step of inputting the first training RF signal data into the semantic-aware segmentation model to obtain the probability map output by the semantic-aware segmentation model includes:

[0054] The multidimensional RF tensor of each sampling point in the first training RF signal data is input into the semantic perception segmentation model to obtain the probability map of each multidimensional RF tensor output by the semantic perception segmentation model. The probability map contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues.

[0055] The calculation of the first loss function based on the probabilities contained in the probability map and the organization labels of the sampling points in the first training RF signal data includes:

[0056] Based on the probabilities contained in the probability map, and the organization labels of each sampling point in the multidimensional RF tensor of each sampling point in the first training RF signal data, the first loss function is calculated using the following formula:

[0057] ;

[0058] ;

[0059] in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. This represents the tissue label of the i-th sampling point.

[0060] Furthermore, after acquiring the second training data for training the signal optimization generation model, the method further includes:

[0061] For each sampling point in the second training RF signal data, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0062] The step of using the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data has been input into the signal optimization generation model to train the signal optimization generation model includes:

[0063] The multidimensional RF tensor of each sampling point in the second training RF signal data is input into the semantic perception segmentation model after training to obtain the probability map output by the semantic perception segmentation model.

[0064] The multidimensional RF tensor of each sampling point in the second training RF signal data, and the probability map output after inputting into the semantic perception segmentation model, are input into the signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0065] Based on the optimized RF signal data and the target reference RF signal data, the second loss function is calculated using the following formula:

[0066] ;

[0067] in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform;

[0068] The signal optimization generation model is trained based on the calculation results of the second loss function.

[0069] Furthermore, training the signal optimization generation model based on the calculation results of the second loss function includes:

[0070] The overall loss function of the semantic-aware segmentation model and the signal optimization generation model is calculated using the following formula:

[0071] ;

[0072] in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively;

[0073] The signal optimization generation model is trained based on the calculation results of the overall loss function.

[0074] This application also provides an ophthalmic ultrasound RF signal region optimization device, including:

[0075] The signal data acquisition module is used to acquire the RF signal data to be optimized from ophthalmic ultrasound of the target object;

[0076] An eye tissue identification module is used to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs by inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various eye tissues. The first training data of the semantic perception segmentation model includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the eye tissue to which the sampling points belong.

[0077] The signal optimization module is used to input the RF signal data to be optimized and the probability map into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

[0078] This application embodiment also provides a signal optimization model training device for ophthalmic ultrasound RF signal region optimization, wherein the signal optimization model includes a semantic-aware segmentation model and a signal optimization generation model, and the device includes:

[0079] The training data acquisition module is used to acquire first training data for training the semantic perception segmentation model. The first training data includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the ocular tissue to which the sampling points belong.

[0080] The semantic awareness training module is used to train the semantic awareness segmentation model by taking the first training RF signal data as input.

[0081] The training data acquisition module is also used to acquire second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

[0082] The signal optimization training module is used to take the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data has been input into the signal optimization generation model as input to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

[0083] This application also provides an electronic device, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor, and the processor is prompted by the machine-executable instructions to: implement any of the above-described ophthalmic ultrasound RF signal region optimization methods, or implement any of the above-described signal optimization model training methods.

[0084] This application also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-described ophthalmic ultrasound RF signal region optimization methods, or implements any of the above-described signal optimization model training methods.

[0085] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the above-described ophthalmic ultrasound RF signal region optimization methods, or to execute any of the above-described signal optimization model training methods.

[0086] The beneficial effects of this application include:

[0087] In the method provided in this application embodiment, the RF signal data to be optimized from ophthalmic ultrasound is input into a semantic-aware segmentation model to identify the ocular tissue to which each sampling point belongs. The resulting probability map contains the probability that each sampling point belongs to various ocular tissues. Furthermore, the RF signal data to be optimized and the probability map output by the semantic-aware segmentation model are input into a signal optimization generation model for adaptive region optimization. Since the probability map contains the probability that each sampling point belongs to various ocular tissues, it is possible to distinguish the RF signal regions of different ocular tissues, thereby achieving targeted enhancement processing for the RF signal regions of different ocular tissues. Compared with the prior art, which uses the same enhancement for the entire signal, it can better reflect the key tissue edges and details, thereby improving the image quality of the ultrasound image subsequently generated based on the optimized RF signal data.

[0088] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0089] The accompanying drawings are provided to further illustrate the present application and form part of the specification. They are used together with the embodiments of the present application to explain the application and do not constitute a limitation thereof. In the drawings:

[0090] Figure 1 A flowchart illustrating the ophthalmic ultrasound RF signal region optimization method provided in this application embodiment;

[0091] Figure 2 A flowchart of the signal optimization model training method provided in the embodiments of this application;

[0092] Figure 3 A flowchart of an ophthalmic ultrasound RF signal region optimization method provided in another embodiment of this application;

[0093] Figure 4 A flowchart of a signal optimization model training method provided in another embodiment of this application;

[0094] Figure 5 This is a flowchart illustrating the process of determining tissue labels at sampling points in an embodiment of this application;

[0095] Figure 6 This is a schematic diagram of the ophthalmic ultrasound RF signal region optimization device provided in the embodiments of this application;

[0096] Figure 7 This is a schematic diagram of the structure of an ophthalmic ultrasound RF signal region optimization device provided in another embodiment of this application;

[0097] Figure 8 This is a schematic diagram of the signal optimization model training device provided in the embodiments of this application;

[0098] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0099] To reduce the loss of critical tissue edges and details in ophthalmic ultrasound imaging and improve the image quality of ultrasound images, this application provides an ophthalmic ultrasound RF signal region optimization method, model training method, and apparatus. The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the embodiments and features described herein can be combined with each other without conflict.

[0100] This application provides a method for optimizing the RF signal region in ophthalmic ultrasound, such as... Figure 1 As shown, it includes:

[0101] Step 11: Obtain the RF signal data to be optimized from ophthalmic ultrasound of the target subject;

[0102] Step 12: By inputting the RF signal data to be optimized into the pre-trained semantic perception segmentation model, the eye tissue to which each sampling point in the RF signal data to be optimized belongs is identified. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various eye tissues. The first training data of the semantic perception segmentation model includes: the first training RF signal data collected, and the tissue label of the sampling point in the first training RF signal data. The tissue label indicates the eye tissue to which the sampling point belongs.

[0103] Step 13: Input the RF signal data to be optimized and the probability map into the pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as the target reference RF signal data corresponding to the second training RF signal data.

[0104] Accordingly, embodiments of this application also provide a signal optimization model training method for ophthalmic ultrasound RF signal region optimization. The signal optimization model includes a semantically perceptual segmentation model and a signal optimization generation model, such as... Figure 2 As shown, the method includes:

[0105] Step 21: Obtain the first training data for training the semantic perception segmentation model. The first training data includes: the first training RF signal data collected, and the tissue labels of the sampling points in the first training RF signal data. The tissue labels represent the eye tissue to which the sampling points belong.

[0106] Step 22: Use the first training RF signal data as input to train the semantic-aware segmentation model;

[0107] Step 23: Obtain second training data for training the signal optimization generation model. The second training data includes: the second training RF signal data collected, and higher quality RF signal data collected by other methods, which serve as the target reference RF signal data corresponding to the second training RF signal data.

[0108] Step 24: Input the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data is completed as input to the signal optimization generation model to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

[0109] The ophthalmic ultrasound RF signal region optimization method provided in this application involves inputting the ophthalmic ultrasound RF signal data to be optimized into a semantic-aware segmentation model to identify the ocular tissue to which each sampling point belongs. The resulting probability map contains the probability that each sampling point belongs to various ocular tissues. Furthermore, the RF signal data to be optimized and the probability map output by the semantic-aware segmentation model are input into a signal optimization generation model for adaptive region optimization. Since the probability map contains the probability that each sampling point belongs to various ocular tissues, it is possible to distinguish the RF signal regions of different ocular tissues, thereby achieving targeted enhancement processing for the RF signal regions of different ocular tissues. Compared with the prior art, which applies the same enhancement to the entire signal, this method can better reflect key tissue edges and details, thereby improving the image quality of the ultrasound image subsequently generated based on the optimized RF signal data.

[0110] In one embodiment of this application, in order to more accurately distinguish similar ocular tissues, such as the cornea and the anterior capsule of the lens, a semantic-aware segmentation method is proposed that combines spatial and temporal neighborhood data of RF signal data. By introducing neighborhood information, the semantic-aware segmentation model can not only analyze the echo intensity of the current sampling point, but also analyze the continuity of the wavefront and the spatial distribution statistical characteristics of the speckle, thereby providing a more accurate physical basis for distinguishing different ocular tissues.

[0111] The following section provides a detailed description of the scheme for optimizing RF signal regions by incorporating spatiotemporal neighborhood information.

[0112] Another embodiment of this application provides a method for optimizing the RF signal region in ophthalmic ultrasound, such as... Figure 3 As shown, it includes:

[0113] Step 31: Obtain the RF signal data to be optimized from ophthalmic ultrasound of the target object.

[0114] In this step, the RF signal data to be optimized can be the raw RF signal data obtained after performing ophthalmic ultrasound scanning on the patient's eyeball and converting the sampled analog signal into digital data, or it can be the data after initial preprocessing of the raw RF signal data.

[0115] Step 32: For each sampling point in the RF signal data to be optimized, generate a multidimensional RF tensor for that sampling point. The multidimensional RF tensor includes the RF signal data of the sampling point and each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0116] In this step, for a sampling point in the RF signal data to be optimized, multiple adjacent acoustic beamlines (A-lines) around the sampling point are extracted, as well as sampling points at multiple times before and after the current time in the sampling time sequence. The sampling points contained in an acoustic beamline include sampling points at all depths of the corresponding channel. These sampling points are obtained to form a multidimensional RF tensor of size H×W×C, where H and W represent the height and width, respectively, represent the spatial neighborhood range, and C represents the number of times in the sampling time sequence.

[0117] A multidimensional RF tensor can be understood as an RF data block (RF Patch), which contains RF signal data of a certain range of sampling points in the spatial and temporal neighborhoods of the sampling point.

[0118] Step 33: Input the multidimensional RF tensor of each sampling point in the RF signal data to be optimized into the pre-trained semantic perception segmentation model to obtain the probability map of each multidimensional RF tensor output by the semantic perception segmentation model.

[0119] In this step, the semantic-aware segmentation model is used to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs. The probability map of each multidimensional RF tensor output by the semantic-aware segmentation model contains the probability that each sampling point of the multidimensional RF tensor belongs to various eye tissues.

[0120] For example, the probability map contains the probability that the RF region represented by the sampling point belongs to the vitreous body (high noise region) and the probability that it belongs to the suspensory ligament (high frequency detail region).

[0121] In the embodiments of this application, various ocular tissues may include: cornea, lens, vitreous body, retina, choroid, and suspensory ligaments.

[0122] In this step, a semantic-aware segmentation model is used, which can calculate the probability distribution of the current anatomical position within milliseconds. For example, it is detected that the current sampling block is located behind the vitreous body. A sampling block contains multiple sampling points, which can be understood as multiple sampling points contained within a certain depth range of a sound beam.

[0123] By utilizing spatial and temporal context, semantic-aware segmentation models can accurately distinguish tissues with similar echo features but different anatomical locations.

[0124] The training process of the semantic-aware segmentation model used in this step will be explained in detail later.

[0125] Step 34: Input the multidimensional RF tensor of each sampling point in the RF signal data to be optimized and the probability map output by the semantic perception segmentation model into the pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0126] For each input multidimensional RF tensor at each sampling point, the signal optimization generation model outputs an optimized multidimensional RF tensor, which contains the optimized RF signal data corresponding to the sampling point.

[0127] In this step, the signal optimization generation model can be used to perform targeted signal optimization processing on the RF signal data generated in the region belonging to the corresponding eye tissue, based on the probability of each sampling point belonging to various eye tissues as shown in the probability map.

[0128] The model is generated through signal optimization, and the weights of the convolution kernels are automatically adjusted according to the probability distribution. For the vitreous region, the model network can exhibit low-pass filtering characteristics for noise reduction, while for regions such as the cornea, lens, retina, choroid, and suspensory ligaments, the model network can exhibit high-pass filtering and gain compensation characteristics for enhancement.

[0129] For example, for regions indicated as vitreous bodies, the variance of the RF signal in that region can be forced to approach 0, thereby achieving noise reduction.

[0130] For key areas such as the lens and suspensory ligament, edge sharpness can be enhanced.

[0131] The training process of the signal optimization generation model used in this step will be explained in detail later.

[0132] In steps 33 and 34 above, an adaptive region optimization of RF signal data is achieved through a dual-branch attention network model, which includes a semantic branch and an enhancement branch, corresponding to the semantic segmentation model and the signal optimization generation model, respectively. This enables intelligent identification of the anatomical region from which the RF signal originates and the application of different processing methods. For example, the suspensory ligament region to be observed is sharpened and enhanced, while the vitreous region is subjected to strong noise reduction. The two processes do not interfere with each other, thereby improving the signal-to-noise ratio and contrast of the image.

[0133] Step 35: Based on the obtained optimized RF signal data, perform image reconstruction to obtain a B-mode grayscale image.

[0134] In this step, envelope detection, logarithmic compression, and coordinate transformation can be performed on the optimized RF signal data to reconstruct high-precision images, thereby obtaining and displaying high-quality ophthalmic ultrasound images that combine "clean dark areas (vitreous body)" and "sharp edges (key tissue boundaries)".

[0135] The above-described embodiments provided in this application Figure 3 The ophthalmic ultrasound RF signal region optimization method shown can intelligently identify the anatomical region from which the RF signal originates and apply different processing accordingly. For example, it can sharpen and enhance key tissue regions such as the suspensory ligaments that need to be observed, while strongly reducing noise in the vitreous region. The two processes do not interfere with each other, improving the signal-to-noise ratio and contrast of the image and reducing the loss of key details inside the eye.

[0136] Furthermore, all the complex computations in this scheme (such as feature learning) are performed during the offline training phase of the model. During the real-time inference phase based on the model, the non-iterative forward propagation computation speed of the aforementioned semantic perception segmentation model and signal optimization generation model is extremely fast, reaching tens or even hundreds of frames per second, which can meet the needs of real-time clinical scanning, imaging, display and viewing.

[0137] Furthermore, this solution performs adaptive region optimization processing on the original RF signal data. Since RF signal data contains the most original information of phase and amplitude, the amount of information it contains far exceeds that of B-mode grayscale images. Directly processing RF signal data can preserve and restore the real acoustic features of eye tissue to the greatest extent, achieving a higher precision enhancement effect than image post-processing.

[0138] The training process of the semantic-aware segmentation model and signal optimization generation model used in the above-mentioned ophthalmic ultrasound RF signal region optimization method is described in detail below.

[0139] In one embodiment of this application, a method for training a signal optimization model for optimizing ophthalmic ultrasound RF signal regions is provided. This signal optimization model includes a semantically perceptual segmentation model and a signal optimization generation model, such as... Figure 4 As shown, its training process includes the following steps:

[0140] Step 41: Obtain the first training data for training the semantic-aware segmentation model.

[0141] In this embodiment of the application, the first training data may include: the collected first training RF signal data, and the tissue label of the sampling point in the first training RF signal data, wherein the tissue label represents the ocular tissue to which the sampling point belongs.

[0142] In practical applications, depending on the specific type of deep learning network used in the semantic perception segmentation model, the Attention U-Net model can be adopted. Attention gating units are introduced into the skip connections of the traditional U-Net, which can automatically suppress background noise in the RF signal data and focus on areas with obvious key echo features such as the retina and lens.

[0143] Alternatively, the Swing Transformer model, such as Swing Transformer V2, can be used to solve the consistency judgment of large-scale structures of the eye (such as the entire vitreous body) by leveraging its powerful long-range dependency capture capability, and it has a natural adaptability to the sliding window processing of RF tensors.

[0144] Alternatively, the DeepLabV3+ (with MobilenetV2 backbone) model can be used, which can employ Spatial Pyramid Pooling with Holes (ASPP) to expand the receptive field without significantly increasing computational cost, making it suitable for identifying subtle edge textures in RF signals, such as suspensory ligaments.

[0145] Step 42: For each sampling point in the first training RF signal data, generate a multidimensional RF tensor for that sampling point. The multidimensional RF tensor includes the RF signal data of the sampling point and each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0146] Step 43: Input the multidimensional RF tensor of each sampling point in the first training RF signal data into the semantic perception segmentation model to be trained, and obtain the probability map of each multidimensional RF tensor output by the semantic perception segmentation model. The probability map contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues.

[0147] Step 44: Based on the probabilities contained in the probability map and the organization labels of each sampling point of the multidimensional RF tensor of each sampling point in the first training RF signal data, calculate the first loss function.

[0148] In this embodiment of the application, the first loss function can be calculated using the following formula:

[0149] ;

[0150] ;

[0151] in, The result of the first loss function calculation is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target tissue. This represents the tissue label of the i-th sampling point.

[0152] Step 45: Based on the calculation results of the first loss function, train the semantic-aware segmentation model.

[0153] In this step, if the calculation result of the first loss function does not meet the first preset condition, it means that the first loss function has not converged yet. The network parameters of the semantic perception segmentation model are adjusted, and the next iteration of training is started.

[0154] When the calculation result of the first loss function meets the first preset condition, it indicates that the first loss function has converged and the training of the semantic perception segmentation model is completed.

[0155] The first preset condition can be flexibly set according to the needs of actual application. For example, the accuracy represented by the calculation result of the first loss function reaches a preset threshold.

[0156] Based on steps 41-45 above, after completing the training of the semantic perception segmentation model, step 46 is initiated to begin training the signal optimization generation model.

[0157] Step 46: Obtain second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

[0158] In the embodiments of this application, other methods that can acquire higher quality RF signal data can be various feasible methods. For example, it can be RF signal data generated offline by a high-precision reconstruction algorithm, or it can be noise-free standard RF signal data generated by a sound field simulation system.

[0159] In practical applications, depending on the specific type of deep learning network used in the signal optimization generation model, a generative adversarial network (CGAN) can be used to make the optimized RF signal data output by the signal optimization generation model closely resemble the target reference RF signal data in terms of texture.

[0160] Alternatively, the Swing Transformer model used in semantic perception segmentation models can be used, while the Restormer model can be used for signal optimization and generation, thus maintaining architectural consistency. Furthermore, a multi-scale hierarchical structure can be adopted, with efficient MDTA (Multiple Transposed Attention) and GDFN (Gated Feedforward Network) at its core. Noise in RF signals often has specific directions or global distributions, and the Restormer model can capture these long-range dependencies, removing point noise in the vitreous body while maintaining the continuity of the lens capsule.

[0161] Alternatively, a Conditioned Diffusion Model can be used to recover signals from noise through an iterative back-diffusion process. When dealing with retinal detachment or subtle lesions, the edge details generated by the Conditioned Diffusion Model are usually superior to those generated by the GAN Model.

[0162] Step 47: For each sampling point in the second training RF signal data, generate a multidimensional RF tensor for that sampling point. The multidimensional RF tensor includes the RF signal data of the sampling point and each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0163] Step 48: Input the multidimensional RF tensor of each sampling point in the second training RF signal data into the completed semantic perception segmentation model to obtain the probability map output by the semantic perception segmentation model.

[0164] Step 49: Input the multidimensional RF tensor of each sampling point in the second training RF signal data, and the probability map output by the semantic perception segmentation model, into the signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0165] Step 410: Calculate the second loss function based on the obtained optimized RF signal data and target reference RF signal data.

[0166] In this embodiment of the application, the second loss function can be calculated using the following formula:

[0167] ;

[0168] in, This is the result of calculating the second loss function. The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point. For the target reference RF signal data of the i-th sampling point, This represents the probability that the i-th sampling point output by the semantic-aware segmentation model belongs to the eye tissue that needs denoising. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each eye tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail reproduction of waveforms.

[0169] The eye tissues that need noise reduction can be the vitreous body, while the eye tissues that need structural reinforcement can be key tissues such as the cornea, lens, retina, choroid, and suspensory ligaments.

[0170] In the formula for calculating the second loss function mentioned above, For noise reduction, As a structural enhancement term, when the probability map indicates that the sampling point belongs to the vitreous body, the second loss function forces the output signal to tend to 0, thereby achieving noise reduction. When the probability map indicates that the sampling point belongs to the critical tissue, the second loss function forces the output signal to structurally approximate the target reference RF signal data, thereby achieving enhancement.

[0171] Step 411: Based on the calculation results of the second loss function, train the signal optimization generation model.

[0172] In this step, the convergence of the second loss function can be determined solely based on its calculation results. If the second loss function has not converged, the network parameters of the signal optimization generation model can be adjusted to proceed to the next iteration of training. If the second loss function converges, the training of the signal optimization generation model can be considered complete.

[0173] The overall loss function for the semantic-aware segmentation model and the signal optimization generation model can also be calculated using the following formula:

[0174] ;

[0175] in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively;

[0176] The signal optimization generation model is trained based on the calculation results of the overall loss function.

[0177] In other words, based on the calculation results of the overall loss function, it is determined whether the overall loss function has converged. If it has not converged, the network parameters of the signal optimization generation model are adjusted and the next iteration of training is started. If it converges, the training of the signal optimization generation model is completed.

[0178] When the signal optimization generation model is trained based on the calculation results of the overall loss function, dual-branch joint training is achieved, namely, joint training of the semantic perception branch and the signal optimization branch, realizing the synergistic effect of semantic perception guiding signal optimization and signal optimization feeding back to semantic perception.

[0179] Semantic awareness-guided signal optimization: Signal optimization can learn how to switch its processing model based on the semantic awareness segmentation results through attention gating.

[0180] Signal optimization feedback semantic perception: In joint training, the gradient adjustment of signal optimization will also be backpropagated to semantic perception. If the optimized RF signal data is clearer at the edge of the eye tissue, semantic perception can more accurately define the boundary of the eye tissue.

[0181] In the training of the above signal optimization model, the first training data includes the organization labels of the sampling points in the first training RF signal data. When training is performed using a multidimensional RF tensor of the sampling points, the first training data includes the organization labels of each sampling point of the multidimensional RF tensor of each sampling point in the first training RF signal data.

[0182] In this embodiment of the application, the organization labels of the sampling points in the first training RF signal data can be pre-set, such as... Figure 5 As shown, the tissue labels of sampling points in the first training RF signal data can be determined using the following steps:

[0183] Step 51: Generate an initial B-mode grayscale image based on the first training RF signal data.

[0184] In this step, various known and feasible ultrasound image generation algorithms can be used to generate an initial B-mode grayscale image based on the first training RF signal data.

[0185] Step 52: Perform semantic segmentation on the initial B-mode grayscale image to generate an anatomical structure mask map. The anatomical structure mask map includes multiple regions divided according to the eye tissue structure, and tissue labels corresponding to each region.

[0186] In this step, an organization label is marked for each region, which is the organization label of the pixels belonging to that region.

[0187] In practical applications, the eye tissues can be divided based on the need for adaptive region optimization of RF signals. For example, they can include the cornea, lens, vitreous body, retina, choroid, and suspensory ligaments.

[0188] In this step, a pre-trained segmentation network can be used to perform semantic segmentation on the initial B-mode grayscale image. In practical applications, various known and feasible segmentation networks can be used, which will not be described in detail here.

[0189] Step 53: For each sampling point in the first training RF signal data, determine the pixel point in the initial B-mode grayscale image corresponding to that sampling point according to the pre-constructed coordinate lookup table, wherein the coordinate lookup table is a non-linear mapping relationship between the pixel points in the initial B-mode grayscale image and the sampling points in the first training RF signal data.

[0190] In this step, the coordinate lookup table can be pre-built for the scanning device, and the building process can include the following steps:

[0191] The first step is to obtain the equipment parameters of the hardware used for ophthalmic ultrasound examination, as well as the acquisition parameters used for signal acquisition.

[0192] It can obtain the geometric parameters of the probe of the hardware device, including the number of array elements (e.g., 512 acoustic beamlines) and the element spacing (Pitch).

[0193] It can also obtain sampling parameters, including sampling frequency (fs) and biological tissue sound velocity (c), which is the speed at which ultrasound waves propagate in the tissues of the eye. Ophthalmology usually takes 1532 m / s or 1550 m / s.

[0194] It can also obtain scan range parameters, including scan depth and scan width;

[0195] It can also obtain the target image resolution, which is the resolution of the B-mode grayscale image to be generated. This is equivalent to the image size displayed on the doctor's end, which is generally an image of 1982×2170 pixels.

[0196] The second step is to establish a Cartesian coordinate system for the B-mode grayscale image based on the above equipment parameters and acquisition parameters.

[0197] Assign the actual coordinates in physical space to each pixel (x, y), that is, where you want the optimized RF signal data to be finally displayed on the screen.

[0198] A Cartesian coordinate grid, where each point represents how far it is from the center of the probe (x-axis) and how deep it is from the probe surface (y-axis).

[0199] Step 3: For each pixel in the B-mode grayscale image The pixel is calculated using the following formula. The corresponding sampling point in the RF signal data :

[0200] ,in, For depth information in the sampling points, Sampling frequency, This represents the speed at which ultrasound waves propagate through the tissues of the eye. The ordinate value of the pixel;

[0201] The calculation belongs to the depth mapping, that is, based on the physical distance of the pixel's y-distance from the probe surface, the total time required for the sound wave to travel from the probe to that depth and then back is calculated, and this time is multiplied by the sampling frequency to convert it into the index number of the RF signal on the depth information. The maximum value of the depth information can be the Depth.

[0202] ,in, This refers to the band information (Channel) of the sampling point. This refers to the element spacing of the probe in the hardware device. The x-coordinate value of the pixel;

[0203] The calculation belongs to the horizontal (channel) mapping. Based on the physical position of pixel x, the physical width x is divided by the element spacing to calculate how many probe elements the pixel spans horizontally. If the number of elements is 512, then... This indicates the index of the sampling point among the 512 acoustic beamlines, which is equivalent to calculating its corresponding A-line index, i.e., determining which original acoustic beamline it belongs to.

[0204] The third step involves inverse coordinate transformation, which reverses the coordinate transformation of each pixel in the B-mode grayscale image to determine its corresponding position in the RF signal data.

[0205] Step 4: Based on the calculated values ​​of each pixel With sampling points Establish the correspondence between them and construct a coordinate lookup table.

[0206] Based on the calculation method in step three above, traverse all 1982×2170 pixels and calculate the value of each pixel. With sampling points The correspondence between them is established and stored in a table to obtain a coordinate lookup table.

[0207] For example, by storing the pixel coordinates as the key and the data index of the sampling point as the value in a table, a coordinate lookup table (LUT) can be obtained.

[0208] Step 54: Determine the tissue label of the pixel corresponding to the sampling point in the anatomical structure mask image as the tissue label of the sampling point.

[0209] In other words, the tissue label of the region to which the pixel corresponding to the sampling point belongs is determined as the tissue label of the sampling point.

[0210] The ophthalmic ultrasound RF signal region optimization method provided in this application has the following beneficial effects:

[0211] This application enables simultaneous bright and dark areas to be clearly visible in ultrasound images. For example, physicians will observe a clear and sharp-edged suspensory ligament structure, while the vitreous region remains clean and free of noise artifacts. This is crucial for clinical needs that rely on suspensory ligament morphology for glaucoma diagnosis or surgical path planning.

[0212] In the prior art, high-quality algorithms, represented by SA-VS, take 10 to 20 minutes to process a single frame of an image. Compared with the SA-VS algorithm, this application compresses the computation time for high-quality reconstruction from minutes to milliseconds, and can easily achieve clinical real-time imaging at more than 30 frames per second (FPS).

[0213] This application performs signal-to-signal processing directly at the front end of the signal chain on the raw RF signal with the most complete information. Because it optimizes the raw RF signal data containing all acoustic information (amplitude and phase), its enhancement and noise reduction effects are far superior to post-processing of B-mode images. This enables the generation of images that more closely approximate the true acoustic characteristics of tissues, with fewer artifacts and higher detail reproduction.

[0214] Based on the same inventive concept, and according to the ophthalmic ultrasound RF signal region optimization method provided in the above embodiments of this application, another embodiment of this application also provides an ophthalmic ultrasound RF signal region optimization device, the structural schematic diagram of which is shown below. Figure 6 As shown, it specifically includes:

[0215] The signal data acquisition module 61 is used to acquire the RF signal data to be optimized from ophthalmic ultrasound of the target object;

[0216] The eye tissue identification module 62 is used to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs by inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various eye tissues. The first training data of the semantic perception segmentation model includes: the first training RF signal data collected, and the tissue label of the sampling point in the first training RF signal data, wherein the tissue label represents the eye tissue to which the sampling point belongs.

[0217] The signal optimization module 63 is used to input the RF signal data to be optimized and the probability map into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

[0218] Furthermore, the signal data acquisition module 61 is also used to generate a multidimensional RF tensor for each sampling point in the ophthalmic ultrasound data of the target object after acquiring the RF signal data to be optimized. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0219] The eye tissue identification module 62 is specifically used to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs by inputting the multidimensional RF tensor of each sampling point in the RF signal data to be optimized into a pre-trained semantic perception segmentation model. The semantic perception segmentation model outputs a probability map of each multidimensional RF tensor, which contains the probability that each sampling point of the multidimensional RF tensor belongs to various eye tissues.

[0220] The signal optimization module 63 is specifically used to input the multidimensional RF tensor and the probability map of each sampling point in the RF signal data to be optimized into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0221] Furthermore, such as Figure 7 As shown, it also includes:

[0222] Tissue label determination module 64 is used to determine the tissue label of the sampling points in the first training RF signal data by the following steps:

[0223] Based on the first training RF signal data, an initial B-mode grayscale image is generated;

[0224] The initial B-mode grayscale image is semantically segmented to generate an anatomical structure mask map, which includes multiple regions divided according to the ocular tissue structure and tissue labels corresponding to each region.

[0225] For each sampling point in the first training RF signal data, the pixel point corresponding to the sampling point in the initial B-mode grayscale image is determined according to the pre-constructed coordinate lookup table. The coordinate lookup table is a non-linear mapping relationship between the pixel points in the B-mode grayscale image and the sampling points in the RF signal data.

[0226] The tissue label of the pixel corresponding to the sampling point in the anatomical structure mask image is determined as the tissue label of the sampling point.

[0227] Furthermore, the organization label determination module 64 is also used to construct the coordinate lookup table using the following steps:

[0228] Obtain the equipment parameters of the hardware used for ophthalmic ultrasound examination, as well as the acquisition parameters used for signal acquisition;

[0229] Based on the device parameters and the acquisition parameters, a Cartesian coordinate system for the B-mode grayscale image is established.

[0230] For each pixel in a B-mode grayscale image The pixel is calculated using the following formula. The corresponding sampling point in the RF signal data :

[0231] ,in, For the depth information of the sampling points, Sampling frequency, This represents the speed at which ultrasound waves propagate through the tissues of the eye. The ordinate value of the pixel;

[0232] ,in, For the band information of the sampling point, The element spacing of the probe in the hardware device. The x-coordinate value of the pixel;

[0233] Based on the calculation of each pixel With sampling points Establish the correspondence between them and construct a coordinate lookup table.

[0234] Furthermore, the first loss function used for training the semantic-aware segmentation model is expressed by the following formula:

[0235] ;

[0236] ;

[0237] in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. Indicates the tissue label of the i-th sampling point;

[0238] The second loss function used for training the signal optimization generation model is expressed by the following formula:

[0239] ;

[0240] in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform;

[0241] The overall loss function of the semantic-aware segmentation model and the signal optimization generation model is expressed by the following formula:

[0242] ;

[0243] in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively.

[0244] Based on the same inventive concept, and according to the signal optimization model training method for ophthalmic ultrasound RF signal region optimization provided in the above embodiments of this application, correspondingly, another embodiment of this application also provides a signal optimization model training device for ophthalmic ultrasound RF signal region optimization. The signal optimization model includes a semantic perception segmentation model and a signal optimization generation model, the structural schematic diagram of which is shown below. Figure 8 As shown, it specifically includes:

[0245] The training data acquisition module 81 is used to acquire first training data for training the semantic perception segmentation model. The first training data includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the ocular tissue to which the sampling points belong.

[0246] The semantic perception training module 82 is used to train the semantic perception segmentation model by taking the first training RF signal data as input to the semantic perception segmentation model.

[0247] The training data acquisition module 81 is also used to acquire second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, as target reference RF signal data corresponding to the second training RF signal data.

[0248] The signal optimization training module 83 is used to take the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data is input as the input of the signal optimization generation model to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

[0249] Furthermore, the semantic perception training module 82 is specifically used to input the first training RF signal data into the semantic perception segmentation model to obtain the probability map output by the semantic perception segmentation model. The probability map contains the probability that each sampling point in the first training RF signal data belongs to various ocular tissues.

[0250] Based on the probabilities contained in the probability map and the organization labels of the sampling points in the first training RF signal data, the first loss function is calculated;

[0251] The semantic-aware segmentation model is trained based on the calculation results of the first loss function.

[0252] Furthermore, the training data acquisition module 81 is also used to generate a multidimensional RF tensor for each sampling point in the first training RF signal data after acquiring the first training data for training the semantic perception segmentation model. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0253] The semantic perception training module 82 is specifically used to input the multidimensional RF tensor of each sampling point in the first training RF signal data into the semantic perception segmentation model to obtain the probability map of each multidimensional RF tensor output by the semantic perception segmentation model. The probability map contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues.

[0254] Based on the probabilities contained in the probability map, and the organization labels of each sampling point in the multidimensional RF tensor of each sampling point in the first training RF signal data, the first loss function is calculated using the following formula:

[0255] ;

[0256] ;

[0257] in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. This represents the tissue label of the i-th sampling point.

[0258] Furthermore, the training data acquisition module 81 is also used to generate a multidimensional RF tensor for each sampling point in the second training RF signal data after acquiring the second training data for training the signal optimization generation model. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point.

[0259] The signal optimization training module 83 is specifically used to input the multidimensional RF tensor of each sampling point in the second training RF signal data into the semantic perception segmentation model after training, and obtain the probability map output by the semantic perception segmentation model.

[0260] The multidimensional RF tensor of each sampling point in the second training RF signal data, and the probability map output after inputting into the semantic perception segmentation model, are input into the signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

[0261] Based on the optimized RF signal data and the target reference RF signal data, the second loss function is calculated using the following formula:

[0262] ;

[0263] in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform;

[0264] The signal optimization generation model is trained based on the calculation results of the second loss function.

[0265] The signal optimization training module 83 is specifically used to calculate the overall loss function of the semantic-aware segmentation model and the signal optimization generation model using the following formula:

[0266] ;

[0267] in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively;

[0268] The signal optimization generation model is trained based on the calculation results of the overall loss function.

[0269] The functions of the above modules can be corresponding to Figures 1 to 5 The corresponding processing steps in the process shown will not be repeated here.

[0270] The ophthalmic ultrasound RF signal region optimization device and signal optimization model training apparatus provided in the embodiments of this application can be implemented by a computer program. Those skilled in the art should understand that the above-described module division method is only one of many module division methods. Whether divided into other modules or not divided into modules, as long as the ophthalmic ultrasound RF signal region optimization device and signal optimization model training apparatus have the above-described functions, they should all be within the protection scope of this application.

[0271] This application also provides an electronic device, such as... Figure 9 As shown, it includes a processor 91 and a machine-readable storage medium 92, the machine-readable storage medium 92 storing machine-executable instructions that can be executed by the processor 91, the processor 91 being prompted by the machine-executable instructions to: implement any of the above-described ophthalmic ultrasound RF signal region optimization methods, or implement any of the above-described signal optimization model training methods.

[0272] This application also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-described ophthalmic ultrasound RF signal region optimization methods, or implements any of the above-described signal optimization model training methods.

[0273] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the above-described ophthalmic ultrasound RF signal region optimization methods, or to execute any of the above-described signal optimization model training methods.

[0274] The machine-readable storage medium in the aforementioned electronic device may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0275] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0276] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of devices, electronic devices, computer-readable storage media, and computer program products are basically similar to the method embodiments, and therefore the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0277] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0278] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0279] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0280] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0281] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for optimizing the RF signal region in ophthalmic ultrasound, characterized in that, include: Acquire the RF signal data of ophthalmic ultrasound of the target subject to be optimized; By inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model, the ocular tissue to which each sampling point in the RF signal data to be optimized belongs is identified. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various ocular tissues. The first training data of the semantic perception segmentation model includes: the first training RF signal data collected, and the tissue label of the sampling point in the first training RF signal data, wherein the tissue label represents the ocular tissue to which the sampling point belongs. The RF signal data to be optimized and the probability map are input into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

2. The method as described in claim 1, characterized in that, After acquiring the RF signal data to be optimized from ophthalmic ultrasound of the target object, the method further includes: For each sampling point in the RF signal data to be optimized, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the RF signal data of the sampling point and each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point. The step of inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs includes: By inputting the multidimensional RF tensor of each sampling point in the RF signal data to be optimized into a pre-trained semantic perception segmentation model, the ocular tissue to which each sampling point in the RF signal data to be optimized belongs is identified. The semantic perception segmentation model outputs a probability map of each multidimensional RF tensor, which contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues. The step of inputting the RF signal data to be optimized and the probability map into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model includes: The multidimensional RF tensor and the probability map of each sampling point in the RF signal data to be optimized are input into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model.

3. The method as described in claim 1, characterized in that, The tissue labels of the sampling points in the first training RF signal data are determined using the following steps: Based on the first training RF signal data, an initial B-mode grayscale image is generated; The initial B-mode grayscale image is semantically segmented to generate an anatomical structure mask map, which includes multiple regions divided according to the ocular tissue structure and tissue labels corresponding to each region. For each sampling point in the first training RF signal data, the pixel point corresponding to the sampling point in the initial B-mode grayscale image is determined according to the pre-constructed coordinate lookup table. The coordinate lookup table is a non-linear mapping relationship between the pixel points in the B-mode grayscale image and the sampling points in the RF signal data. The tissue label of the pixel corresponding to the sampling point in the anatomical structure mask image is determined as the tissue label of the sampling point.

4. The method as described in claim 3, characterized in that, The coordinate lookup table is constructed using the following steps: Obtain the equipment parameters of the hardware used for ophthalmic ultrasound examination, as well as the acquisition parameters used for signal acquisition; Based on the device parameters and the acquisition parameters, a Cartesian coordinate system for the B-mode grayscale image is established. For each pixel in a B-mode grayscale image The pixel is calculated using the following formula. The corresponding sampling point in the RF signal data : ,in, For the depth information of the sampling points, Sampling frequency, This represents the speed at which ultrasound waves propagate through the tissues of the eye. The ordinate value of the pixel; ,in, For the band information of the sampling point, The element spacing of the probe in the hardware device. The x-coordinate value of the pixel; Based on the calculation of each pixel With sampling points Establish the correspondence between them and construct a coordinate lookup table.

5. The method as described in claim 1, characterized in that, The first loss function used for training the semantic-aware segmentation model is expressed by the following formula: ； ； in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. Indicates the tissue label of the i-th sampling point; The second loss function used for training the signal optimization generation model is expressed by the following formula: ； in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform; The overall loss function of the semantic-aware segmentation model and the signal optimization generation model is expressed by the following formula: ； in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively.

6. A method for training a signal optimization model for optimizing ophthalmic ultrasound RF signal regions, characterized in that, The signal optimization model includes a semantic-aware segmentation model and a signal optimization generation model, and the method includes: Acquire first training data for training a semantic-aware segmentation model. The first training data includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the ocular tissue to which the sampling points belong. The first training RF signal data is used as the input to the semantic-aware segmentation model to train the semantic-aware segmentation model; Acquire second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, as target reference RF signal data corresponding to the second training RF signal data. The second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data is input into the signal optimization generation model to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

7. The method as described in claim 6, characterized in that, The step of using the first training RF signal data as input to the semantic-aware segmentation model to train the semantic-aware segmentation model includes: The first training RF signal data is input into the semantic perception segmentation model to obtain the probability map output by the semantic perception segmentation model. The probability map contains the probability that each sampling point in the first training RF signal data belongs to various ocular tissues. Based on the probabilities contained in the probability map and the organization labels of the sampling points in the first training RF signal data, the first loss function is calculated; The semantic-aware segmentation model is trained based on the calculation results of the first loss function.

8. The method as described in claim 7, characterized in that, After obtaining the first training data for training the semantically aware segmentation model, the method further includes: For each sampling point in the first training RF signal data, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of that sampling point. The step of inputting the first training RF signal data into the semantic-aware segmentation model to obtain the probability map output by the semantic-aware segmentation model includes: The multidimensional RF tensor of each sampling point in the first training RF signal data is input into the semantic perception segmentation model to obtain the probability map of each multidimensional RF tensor output by the semantic perception segmentation model. The probability map contains the probability that each sampling point of the multidimensional RF tensor belongs to various ocular tissues. The calculation of the first loss function based on the probabilities contained in the probability map and the organization labels of the sampling points in the first training RF signal data includes: Based on the probabilities contained in the probability map, and the organization labels of each sampling point in the multidimensional RF tensor of each sampling point in the first training RF signal data, the first loss function is calculated using the following formula: ； ； in, The result of the calculation of the first loss function is given, where K is the total number of target tissues. The output of the loss function of the semantic-aware segmentation model represents the probability value that the i-th sampling point belongs to the k-th target organization. This represents the tissue label of the i-th sampling point.

9. The method as described in claim 8, characterized in that, After obtaining the second training data for training the signal optimization generation model, the method further includes: For each sampling point in the second training RF signal data, a multidimensional RF tensor is generated for that sampling point. The multidimensional RF tensor includes the sampling point and the RF signal data of each sampling point in the preset spatial neighborhood and preset temporal neighborhood of the sampling point. The step of using the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data has been input into the signal optimization generation model to train the signal optimization generation model includes: The multidimensional RF tensor of each sampling point in the second training RF signal data is input into the semantic perception segmentation model after training to obtain the probability map output by the semantic perception segmentation model. The multidimensional RF tensor of each sampling point in the second training RF signal data, and the probability map output after inputting into the semantic perception segmentation model, are input into the signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. Based on the optimized RF signal data and the target reference RF signal data, the second loss function is calculated using the following formula: ； in, The result of calculating the second loss function is... The output of the signal optimization generation model represents the optimized RF signal data at the i-th sampling point; For the target reference RF signal data of the i-th sampling point, The probability that the i-th sampling point output by the semantic perception segmentation model belongs to the eye tissue that needs noise reduction. The sum of probabilities that the i-th sampling point output by the semantic-aware segmentation model belongs to each ocular tissue requiring structural enhancement. Structural similarity loss is used to measure the degree of detail restoration of the waveform; The signal optimization generation model is trained based on the calculation results of the second loss function.

10. The method as described in claim 9, characterized in that, The process of training the signal optimization generation model based on the calculation results of the second loss function includes: The overall loss function of the semantic-aware segmentation model and the signal optimization generation model is calculated using the following formula: ； in, and These are the weights of the preset semantic perception segmentation model and the signal optimization generation model, respectively; The signal optimization generation model is trained based on the calculation results of the overall loss function.

11. An ophthalmic ultrasound RF signal region optimization device, characterized in that, include: The signal data acquisition module is used to acquire the RF signal data to be optimized from ophthalmic ultrasound of the target object; An eye tissue identification module is used to identify the eye tissue to which each sampling point in the RF signal data to be optimized belongs by inputting the RF signal data to be optimized into a pre-trained semantic perception segmentation model. The semantic perception segmentation model outputs a probability map, which contains the probability that each sampling point belongs to various eye tissues. The first training data of the semantic perception segmentation model includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the eye tissue to which the sampling points belong. The signal optimization module is used to input the RF signal data to be optimized and the probability map into a pre-trained signal optimization generation model to obtain the optimized RF signal data output by the signal optimization generation model. The second training data of the signal optimization generation model includes: sampled second training RF signal data and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data.

12. A signal optimization model training device for ophthalmic ultrasound RF signal region optimization, characterized in that, The signal optimization model includes a semantic-aware segmentation model and a signal optimization generation model, and the device includes: The training data acquisition module is used to acquire first training data for training the semantic perception segmentation model. The first training data includes: first training RF signal data collected, and tissue labels of sampling points in the first training RF signal data, wherein the tissue labels represent the ocular tissue to which the sampling points belong. The semantic awareness training module is used to train the semantic awareness segmentation model by taking the first training RF signal data as input. The training data acquisition module is also used to acquire second training data for training the signal optimization generation model. The second training data includes: second training RF signal data collected, and higher quality RF signal data collected by other methods, which serve as target reference RF signal data corresponding to the second training RF signal data. The signal optimization training module is used to take the second training RF signal data and the probability map output by the semantic perception segmentation model after the second training RF signal data has been input into the signal optimization generation model as input to train the signal optimization generation model. The probability map contains the probability that each sampling point in the second training RF signal data belongs to various ocular tissues.

13. An electronic device, characterized in that, The device includes a processor and a machine-readable storage medium storing machine-executable instructions that can be executed by the processor, the processor being prompted by the machine-executable instructions to: implement the method of any one of claims 1-5, or implement the method of any one of claims 6-10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-5, or implements the method of any one of claims 6-10.

Images