Wireless ultrasonic probe real-time deformation compensation method and system based on AI

Abstract

The invention provides an AI-based wireless ultrasonic probe real-time deformation compensation method and system. The method comprises the following steps: acquiring original image data and original pressure data, establishing an image structure characteristic path and a pressure response path, respectively extracting structural and physical deformation trends, and generating an image structure change response diagram and a probe pressure induction response diagram; fusing the image structure change response graph and the probe pressure induction response graph to construct a fusion feature graph; based on the fused feature map, constructing a convolutional network model, predicting a non-rigid deformation field of the current frame image, and outputting the non-rigid deformation field; and applying the non-rigid deformation field to a current frame image, executing space coordinate reflection and image resampling operation, and outputting an image after deformation compensation. According to the method, the non-rigid deformation of the wireless ultrasonic image is quickly corrected within the edge computing power through ambiguity-guided adaptive fusion and direction and density regularization deformation prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image processing, and particularly relates to an AI-based wireless ultrasonic probe real-time deformation compensation method and system. BACKGROUND

[0002] Wireless ultrasonic probes are widely used in bedside emergency, remote consultation and field patrol due to their advantages of portability, wirelessness and multi-modal imaging. However, under the condition of lack of wired power supply and high bandwidth, the image stability and real-time performance are particularly prominent. Due to the limitations of device size and power consumption, large image processing hardware cannot be integrated into such probes. Meanwhile, Wi-Fi transmission jitter can cause frame rate to drop sharply, image blur and frame loss. Micro-vibration, sliding and non-constant pressing caused by doctor's handheld operation can also cause complex non-rigid deformation of tissues. Most of the existing methods rely on full-image high-quality registration or only estimate displacement from a single image signal, which cannot give reliable results in weak texture areas and cannot perceive the key physical factor of probe external force. When the image reference frame is missing or the boundary is blurred, the algorithm accuracy decreases sharply and the delay is too high, which cannot meet the clinical needs of real-time imaging in wireless scenarios. SUMMARY

[0003] The purpose of the present application is to design an AI-based wireless ultrasonic probe real-time deformation compensation method and system, which can complete deformation correction within frame-level delay without high-quality reference frames, maintain stable compensation effect for blurred, lost frames and pressure areas, and be directly integrated into portable wireless probes to realize low-power, high-robustness real-time imaging stabilization, significantly improving the reliability and accuracy of ultrasonic diagnosis in complex scenarios.

[0004] To achieve the above purpose, an AI-based wireless ultrasonic probe real-time deformation compensation method is provided in the first aspect of the present application, which comprises:

[0005] Collecting original image data and original pressure data, establishing image structure feature paths and pressure response paths, extracting structural and physical deformation trends respectively, and generating image structure change response maps and probe pressure-induced response maps;

[0006] Fusing the image structure change response maps and the probe pressure-induced response maps to construct a fusion feature map; wherein the fusion feature map is obtained based on a blur map; the blur map is obtained by local Sobel variance normalization of the original image data;

[0007] Based on the fusion feature map, a convolutional network model is constructed to predict the non-rigid deformation field of the current frame image and output the non-rigid deformation field; wherein the tissue density feature map calculated based on the original image data is also introduced in the prediction; the direction consistency term, boundary slip suppression term and tissue density adaptive term are also introduced in the non-rigid deformation field.

[0008] apply the non-rigid deformation field to the current frame image, perform spatial coordinate back mapping and image resampling operations, and output a deformation compensated image; wherein the spatial coordinate back mapping includes applying a deformation displacement of the non-rigid deformation field to each original image data pixel, calculating a back mapping coordinate corresponding to the deformation displacement, and then taking the back mapping coordinate as a sampling target to obtain a new pixel value from the original image data to obtain a back sampling coordinate map.

[0009] Further, the original image data and the original pressure data specifically include: a current frame image, which is a two-dimensional grayscale image collected by an image collection module of the probe; a previous frame image, which is an image frame collected last time and stored in a RAM buffer inside the device; and a probe pressure distribution map, which is collected by a capacitive pressure sensor array arranged on the device shell.

[0010] Further, generating the image structure change response map and the probe pressure induced response map specifically includes: the image channel adopts an inter-frame difference strategy to extract non-rigid structure changes, and generates the image structure change response map through a light convolution structure from the image frame difference; and the pressure channel is used to supplement the deficiency of the image structure detection capability, and generates the probe pressure induced response map through a light perception path after the interpolation and normalization of the pressure array map.

[0011] Further, fusing the image structure change response map and the probe pressure induced response map to construct a fusion feature map specifically includes: fusing the blur map and the image structure change response map, and introducing a boundary preservation regularization term and a tissue response consistency constraint term, the boundary preservation regularization term is used to suppress the image structure mutation after fusion, and the tissue response consistency constraint term is used to enhance the consistency of the tissue deformation mode and the pressure map spatial response in the fusion map.

[0012] Further, the tissue response consistency constraint term is a gradient dot product of the image structure change response map and the probe pressure induced response map.

[0013] Further, the convolutional network model is a three-layer convolutional network model: the first layer is a 3x3 convolution with an input channel of 1, an output channel of 32, and a ReLU activation; the second layer is a 3x3 convolution with a channel number of 32→16, a ReLU activation; and the third layer is a 3x3 convolution with a channel number of 16→2, without activation, and is used to output the non-rigid deformation field.

[0014] Further, the direction consistency term, the boundary slip inhibition term and the tissue density adaptive term introduced in the non-rigid deformation field specifically include: the direction consistency term punishes the pixel points with inconsistent deformation direction and gradient direction of the fusion feature to avoid multi-directional cross deformation field; the boundary slip inhibition term forces the edge deformation vector change to remain smooth and limits the deformation jump in the image edge area; and the tissue density adaptive term weakens the deformation prediction strength of the model in the image density higher area.

[0015] Further, the image resampling operation is a bilinear interpolation on the reverse sampling coordinate map, the bilinear interpolation is an interpolation on any floating point coordinate, and four adjacent integer coordinate points of the floating point coordinate are and the upper, lower, left and right combinations thereof, and the corresponding interpolation weights are distributed in a standard ratio.

[0016] Further, when the sampling point falls outside the image boundary, the image resampling operation adopts an edge value extension method for processing, and the edge value extension method is to take the nearest boundary value for compensation.

[0017] In a second aspect of the present application, an AI-based wireless ultrasonic probe real-time deformation compensation system is provided, and the system comprises:

[0018] A feature extraction unit is configured to collect original image data and original pressure data, establish an image structure feature path and a pressure response path, extract structural and physical deformation trends respectively, and generate an image structure change response map and a probe pressure-induced response map.

[0019] A feature fusion unit is configured to fuse the image structure change response map and the probe pressure-induced response map to construct a fusion feature map, wherein the fusion feature map is obtained based on a blur map; and the blur map is obtained by local Sobel variance normalization of the original image data.

[0020] A deformation prediction unit is configured to construct a convolutional network model based on the fusion feature map, predict a non-rigid deformation field of a current frame of image, and output the non-rigid deformation field; wherein a tissue density feature map calculated based on the original image data is also introduced in the prediction; and a direction consistency term, a boundary slip inhibition term and a tissue density adaptive term are also introduced in the non-rigid deformation field.

[0021] An image compensation unit is configured to apply the non-rigid deformation field to the current frame of image, perform spatial coordinate reverse mapping and image resampling operation, and output the image after deformation compensation; wherein the spatial coordinate reverse mapping includes applying deformation displacement of the non-rigid deformation field to each pixel of the original image data, calculating reverse mapping coordinates corresponding to the deformation displacement, obtaining new pixel values from the original image data by taking the reverse mapping coordinates as sampling targets, and obtaining a reverse sampling coordinate map.

[0022] The beneficial technical effects of the present invention are at least as follows:

[0023] To address the aforementioned issues, this invention provides an AI simulation teaching method and system for holographic interactive teaching scenarios. This invention generates alignable structural-physical candidate deformation information by acquiring differential features between the current and previous frames and a pressure array map at the bottom of the probe. Subsequently, based on local ambiguity, it adaptively weights and fuses features from two pathways, and suppresses jumps through boundary-maintaining and tissue-response consistency regularization. On this basis, it outputs a pixel-level two-dimensional displacement field using a shallow convolutional network with orientation consistency, edge smoothing, and density adjustment terms. Finally, this displacement field is applied to the original image to perform reverse coordinate mapping, and orientation-aware interpolation is used to generate a compensated image. The entire system can complete deformation correction within frame-level time delay without requiring a high-quality reference frame, maintaining stable compensation effects for blurred, dropped, and pressure-affected areas. It can be directly integrated into portable wireless probes to achieve low-power, highly robust real-time imaging stabilization. Attached Figure Description

[0024] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0025] Figure 1 This is a flowchart of the AI-based real-time deformation compensation method for wireless ultrasonic probes according to the present invention.

[0026] Figure 2 This is a framework diagram of the AI-based wireless ultrasonic probe real-time deformation compensation system of the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0028] In one or more embodiments, such as Figure 1 As shown, a real-time deformation compensation method for wireless ultrasound probes based on AI is disclosed, the method comprising the following:

[0029] S1: Collect raw image data and raw pressure data, establish image structure feature pathway and pressure response pathway, extract structural and physical deformation trends respectively, and generate image structure change response map and probe pressure-induced response map;

[0030] Specifically, in this embodiment, this step aims to extract candidate deformation features from two independent channels of the wireless ultrasound device: one is the structural changes in the image frame itself, and the other is the pressure response on the probe contact surface. In practical use, wireless ultrasound is easily affected by changes in the doctor's handholding posture, light pressure on the probe, or rotational movements, causing non-uniform and unpredictable deformation in the image. Considering that image structural changes may not be sufficient to cover all deformation areas, especially in low-texture areas, to enhance the system's perception coverage, this step establishes an image structural feature pathway and a pressure response pathway to extract structural and physical deformation trends respectively, providing an aligned and interpretable basic input for subsequent fusion.

[0031] This step receives the following three inputs, all of which are directly acquired by the device hardware:

[0032] The current frame image I(x,y) is acquired by the probe's image acquisition module. It is a two-dimensional grayscale image with a size of 512×512, a sampling bit width of 8 bits, a data type of unsigned integer, and a frame rate of approximately 30fps.

[0033] The previous frame image I′(x,y) is stored in the device's internal RAM buffer. The time difference is about 33ms. It is the most recently acquired image frame and its format is the same as the current frame.

[0034] The probe pressure distribution map F(x,y) is collected by a 16×16 capacitive pressure sensor array deployed at the bottom of the device housing, synchronized to the main control chip via the I2C interface, linearly interpolated to 512×512 by hardware, and normalized to the [0,1] interval.

[0035] The image path employs an inter-frame difference strategy to extract non-rigid structural changes and enhances the response region through a lightweight convolutional network. In practice, when the doctor slides the probe across the abdomen, the B-mode image produces continuous deformation regions, where tissue edge structures are often blurred or even misaligned due to fine adjustments in the probe angle. Since full-frame registration alone is insufficient to determine the region's response location, the following method is used for image response modeling:

[0036] In this embodiment, the image frame difference I(x,y)-I′(x,y) is calculated, then passed through two convolutional layers (3×3 convolutional kernels, with 16 and 32 channels respectively), using the ReLU activation function in between, followed by a 2×2 max pooling operation, to output the tensor Z. I (x,y) represents the structural change response graph of the current frame.

[0037] Furthermore, the pressure pathway is used to supplement the shortcomings of image structure detection capabilities, especially in tissue blocks with insufficient image structural information and low signal-to-noise ratio (such as the periphery of the liver, smooth boundaries of the bladder, etc.). After interpolating and normalizing the original pressure map F(x,y), it is input into a two-layer perceptual pathway: first, a 1×1 convolution (8 output channels) is used to enhance the local response, and then a 3×3 average pooling layer is used to extract the regional pressure trend, resulting in the tensor Z. F (x,y) represents the local stress-induced response diagram.

[0038] The overall calculation process is expressed as follows:

[0039] Z I (x,y)=Pool(ReLU(Conv2(Conv1(I(x,y)-I′(x,y)))))

[0040] Z F (x,y)=AvgPool(Conv 1×1 (Interp(F(x,y))))

[0041] Among them: Z I (x,y) represents the image structure change response map, which is a two-dimensional tensor obtained by applying two layers of 3×3 convolution (16 and 32 channels), ReLU activation, and 2×2 max pooling to the image frame difference; Z F (x,y) represents the pressure-induced response map, which is a two-dimensional tensor obtained by interpolating and normalizing the original pressure map F(x,y), followed by 1×1 convolution (8 channels) and 3×3 average pooling; I(x,y) and I′(x,y) are the two-dimensional grayscale images of the current frame and the previous frame, respectively, with a resolution of 512×512; F(x,y) is the original two-dimensional image acquired by the device pressure array, with an original size of 16×16, which is aligned with the image after interpolation; Conv1 and Conv2 are the first and second 3×3 convolution operations, respectively; Interp(·) represents linearly interpolating the low-resolution pressure map to the image size; all operations maintain two-dimensional spatial consistency and use zero-padding for edge processing.

[0042] The key to this step is that traditional methods often only model the structural changes of the entire image frame, failing to handle areas in the image without obvious structural boundaries or textures. This step constructs two pathways: the image pathway focuses on the deformation of visible structures, while the pressure pathway infers potential deformation areas not shown in the image through physical contact information. This enables the system to more accurately detect non-rigid displacements during common wireless ultrasound operations such as slight probe sliding, rotation, or pressure.

[0043] Specifically, in clinical bladder measurement scenarios, when a doctor presses the probe hard to clearly display the anterior wall contour, the contour of that area in the image will be compressed and blurred. However, the pressure pathway can reflect the high intensity of contact in that area in real time and activate the corresponding features, thereby providing a real physical response reference for subsequent compensation paths.

[0044] S2: The image structure change response map and the probe pressure-induced response map are fused to construct a fused feature map; wherein, the fused feature map is constructed based on the ambiguity map; the ambiguity map is obtained by local Sobel variance normalization of the original image data;

[0045] Specifically, in this embodiment, this step aims to transform the image structure change response map Z obtained in step S1 into... I (x,y) and pressure response diagram Z F (x, y) are fused to construct a unified candidate deformation feature map with region-aware capabilities. In actual clinical scenarios, wireless ultrasound images are often affected by minor probe slippage, local pressure, and differences in tissue echoes, resulting in blurred regions, fracture boundaries, and discontinuous tissue structures. The structural signals in these image regions are often unstable or of low reliability, while pressure information has more direct physical response characteristics. Therefore, to ensure that the system can adaptively perceive and assign weights based on local image quality, this step constructs a fusion mechanism guided by ambiguity awareness. A fusion optimization model based on boundary-preserving regularization and tissue response consistency terms is designed, taking into account the characteristics of ultrasound images. This model not only performs region-weighted fusion but also optimizes the spatial consistency of the fusion field, ensuring its interpretability and structural stability for clinical applications.

[0046] In this embodiment, the local blur map β(x,y) of the image is calculated. The core idea is to measure the reliability of the image region in terms of structural representation. We calculate the gradient strength variance of the Sobel operator in its 5×5 neighborhood at each pixel and normalize it to the [0,1] interval for the entire image. The β(x,y) value of highly blurred regions is close to 0, while that of clear regions is close to 1.

[0047] Specifically, during bladder capacity monitoring, if the image boundaries are blurred and the tissue contours are severely distorted, Z... I The structural response provided by (x,y) may mislead subsequent deformation predictions in these regions. And Z... F (x,y) serves as a pressure-induced physical response map, which can more stably cover these weakly textured areas. Therefore, the fusion must dynamically adjust the weights based on the local sharpness of the image.

[0048] To achieve this goal, this step constructs the fusion representation as follows:

[0049]

[0050] Where: β(x,y) is the fuzzy map, obtained through local Sobel variance normalization of the image; the first term is the explicit fusion expression based on confidence; the second term... It is a boundary preservation regularization term used to suppress abrupt structural changes in the fused image. Let μ represent the two-dimensional Laplace operator; the third term μ·ψ(x,y) is an organizational response consistency constraint term, used to enhance the consistency between the organizational deformation pattern and the spatial response of the stress map in the fusion graph, defined as:

[0051]

[0052] in: and These are the gradients of the two feature maps, respectively; when ψ(x,y) increases, it indicates that the structure map and the pressure map are aligned in direction, and the fusion map should not penalize the differences; otherwise, structure alignment is performed; λ and μ are two regularization parameters, usually set in the range of [0.01, 0.1], and are tuned by the deployment platform performance.

[0053] This fusion expression possesses strong structural and physical consistency control capabilities. In areas with severe wireless image blur (such as the liver surface and intercostal sliding surfaces), Z... I Traditional direct weighted fusion can amplify discontinuities, which are prone to abrupt changes in response. By introducing a boundary preservation term, the system automatically penalizes regions of significant change in the fused graph, ensuring a smoother spatial distribution and more consistent orientation. Furthermore, an organization response consistency term enhances regions by comparing the gradient direction consistency between the two graphs, making the fused graph more consistent with the expected continuity of physical anatomical structures.

[0054] Specifically, in the assessment of diaphragmatic slippage, the image is clear in the middle but blurry at the edges and under significant stress. This mechanism can preserve the structural features of the image in the middle and automatically enhance the pressure-induced response at the edges, thereby constructing a fusion field that better reflects the actual deformation.

[0055] This step outputs the fused feature map Z. fusion (x, y) represents a unified deformation candidate map obtained by integrating image structure credibility, external force distribution, and boundary physical continuity, used for non-rigid deformation field prediction in the next step. This tensor has a size of 512×512, and its numerical range is normalized after fusion. This step addresses the problem of invalid structural features in blurred image regions by supplementing them with physical pressure information, suppressing image jump risks through regularization terms, and strengthening structural consistency through gradient collaboration.

[0056] S3: Based on the fused feature map, a convolutional network model is constructed to predict the non-rigid deformation field of the current frame image and output the non-rigid deformation field; wherein the prediction also incorporates a tissue density feature map calculated based on the original image data; the non-rigid deformation field also incorporates an orientation consistency term, a boundary slip suppression term, and a tissue density adaptive term;

[0057] Specifically, in this embodiment, the fused feature map Z output in step two is... fusion Based on (x,y), predict the non-rigid deformation field of the current frame image. As the core computational step of this scheme, this step determines how the final image is compensated and whether the compensation possesses tissue structural continuity, edge stability, and physical consistency. Due to the handheld operation of wireless ultrasound probes, complex manipulation behaviors such as slight rotation, non-uniform pressure, and local sliding often occur, resulting in multi-directional and spatially inconsistent tissue deformation regions in the image. Traditional methods typically rely on full-image registration or fixed-pattern prediction, which struggles to handle such dynamically complex and directionally discontinuous scenarios. Therefore, this step proposes a structure-guided deformation field prediction method based on fused feature maps. In the model structure design, directional consistency regularization, edge slip suppression terms, and tissue density adaptive adjustment terms are introduced. The deformation field is simultaneously modeled from three dimensions: spatial structure, mechanical response, and image attributes, improving the system's adaptability to highly dynamic clinical ultrasound scenarios.

[0058] Furthermore, a tissue density feature map ρ(x,y) is calculated from the original image I(x,y). This map is used to adjust the response intensity of the deformation field in different tissue regions and avoid over-prediction of displacement in high-density regions (such as bone or boundary reflection areas).

[0059] Specifically, this step constructs a lightweight three-layer convolutional network model, with the following structure:

[0060] The first layer is a 3×3 convolution with 1 input channel and 32 output channels, activated by ReLU.

[0061] The second layer is a 3×3 convolution, with the number of channels changing from 32 to 16, and ReLU activation;

[0062] The third layer is a 3×3 convolution with 2 channels instead of 16, and no activation. It is used to output the final two-dimensional deformation vector field (Δx, Δy).

[0063] Input is Z fusion (x, y), the output is This represents the coordinate displacement that should occur at each pixel. To address issues such as tissue reversal risk, boundary drift, and high-frequency structural breakage in wireless ultrasound images, this step introduces the following structure into the network optimization objective:

[0064]

[0065] in: This is an L1 regularization term used to limit excessive deformation estimation; The directional consistency term is defined as follows:

[0066]

[0067] This penalty applies to pixels whose deformation direction is inconsistent with the direction of the fused feature gradient, thus avoiding multi-directional cross deformation fields.

[0068] The boundary slip suppression term restricts deformation abruptness in image edge regions, and is defined as follows:

[0069]

[0070] in For the 5-pixel area outside the image edge, this item forces the edge deformation vector change to remain smooth, adapting to the blurring and drift phenomenon commonly found at the frame edges of wireless ultrasound images.

[0071] This is the tissue density adaptive term, whose goal is to allow the model to automatically reduce the deformation prediction intensity in areas with high image density (such as bone reflection areas and strong tissue interfaces). Its specific definition is as follows:

[0072]

[0073] Where ρ(x,y) is the image density map, which is obtained by normalizing the brightness statistics within the local maximum response window of I(x,y). The larger the value, the higher the tissue density.

[0074] These three losses together constitute a three-dimensional structure-aware constraint system encompassing orientation, boundary, and tissue. This system maintains overall orientation accuracy and boundary stability while automatically adjusting the displacement prediction intensity in different regions. In practical applications, when the probe slides laterally along the edge of the liver region, the image at the boundary is blurred, but the internal tissue density is high. Without a density adjustment term, the model might mispredict high-amplitude displacements in these areas, leading to structural misalignment. This mechanism automatically suppresses displacement intensity at the edges and in high-density areas, achieving structural stability consistent with reality.

[0075] The two-dimensional non-rigid deformation field obtained in this step The tensor is 512×512×2, which is the same size as the image. Each pixel contains a displacement vector in the form of (Δx, Δy). After tanh activation, the range is limited to [-25, 25] pixels, which can be directly used for the next step of image coordinate inversion and resampling.

[0076] S4: Apply the non-rigid deformation field to the current frame image, perform spatial coordinate inverse mapping and image resampling operations, and output the image after deformation compensation; wherein, the spatial coordinate inverse mapping includes applying the deformation displacement of the non-rigid deformation field to each original image data pixel, calculating the inverse mapping coordinates corresponding to the deformation displacement, and then obtaining new pixel values ​​from the original image data with the inverse mapping coordinates as the sampling target to obtain the inverse sampling coordinate map.

[0077] Specifically, in this embodiment, this step aims to transform the non-rigid deformation vector field obtained in the previous stage... Applied to the current frame image I(x,y), perform spatial coordinate inverse mapping and image resampling operations, and output the deformed compensated image I. corrected (x,y).

[0078] The core of image compensation is to apply deformation displacement to each pixel (x,y) of the original image. Calculate the corresponding reverse mapping coordinates T(x,y), and then use these as the sampling target to obtain new pixel values ​​from the original image.

[0079] Define the backsampling coordinate graph as follows:

[0080] T(x,y)=(x-Δx(x,y),y-Δy(x,y))

[0081] Where (Δx(x,y),Δy(x,y)) comes from the output of the previous step, and the unit is pixels.

[0082] Perform bilinear interpolation on T(x,y) to generate a new image I. corrected (x,y). In the implementation, if the sampling point falls outside the image boundary (i.e., T(x,y) has coordinates outside the [0,511] interval), the edge value extension method is used to process it, that is, the nearest boundary value is taken for compensation.

[0083] The interpolation process uses bilinear weighting for calculation. For any floating-point coordinate (u, v), its four adjacent integer coordinates are... The interpolation weights, along with their combinations (up, down, left, and right), are allocated according to a standard ratio (the closer the distance, the higher the weight). The entire interpolation process is executed using floating-point matrices, and vectorized processing supports parallel acceleration via GPU or NEON SIMD.

[0084] Specifically, during liver B-mode imaging, doctors often use rapid lateral movement to obtain the complete liver lobe structure. If the probe is slightly tilted at this time, the lower structure is easily compressed and the upper boundary is blurred in the image. Step three predicts... This process generates a large negative vertical displacement in the lower region of the image, while outputting a smaller or even positive vertical displacement in the upper boundary region. In this step, this deformation field directly acts on the image spatial coordinates. After the image pixels are remapped, the liver tissue contour is restored to a near-natural state, and the boundary coherence is significantly improved.

[0085] This step outputs the compensated image I. corrected (x,t) is a grayscale image with dimensions of 512×512 pixels, and the pixel type is 8-bit unsigned integer, consistent with the original image. This image is the current frame after resampling compensation under the guidance of the deformation field, possessing a more stable organizational structure and more accurate spatial layout, and can be directly used for device display, storage, or further analysis.

[0086] In one or more embodiments, such as Figure 2 As shown, an AI-based real-time deformation compensation system for wireless ultrasound probes is disclosed, the system comprising:

[0087] The feature extraction unit is used to acquire raw image data and raw pressure data, establish image structure feature pathways and pressure response pathways, extract structural and physical deformation trends respectively, and generate image structure change response maps and probe pressure-induced response maps.

[0088] The feature fusion unit is used to fuse the image structure change response map and the probe pressure-induced response map to construct a fused feature map; wherein, the fused feature map is constructed based on a fuzzy map; the fuzzy map is obtained by local Sobel variance normalization of the original image data;

[0089] The deformation prediction unit is used to construct a convolutional network model based on the fused feature map, predict the non-rigid deformation field of the current frame image, and output the non-rigid deformation field; wherein the prediction also incorporates a tissue density feature map calculated based on the original image data; the non-rigid deformation field also incorporates an orientation consistency term, a boundary slip suppression term, and a tissue density adaptive term;

[0090] The image compensation unit is used to apply the non-rigid deformation field to the current frame image, perform spatial coordinate inverse mapping and image resampling operations, and output the deformation-compensated image; wherein, the spatial coordinate inverse mapping includes applying the deformation displacement of the non-rigid deformation field to each original image data pixel, calculating the inverse mapping coordinates corresponding to the deformation displacement, and then obtaining new pixel values ​​from the original image data with the inverse mapping coordinates as the sampling target to obtain the inverse sampling coordinate map.

[0091] It is worth noting that the specific workflow of the AI-based wireless ultrasound probe real-time deformation compensation system provided in this embodiment is the same as that of the AI-based wireless ultrasound probe real-time deformation compensation method described in the above embodiment, and will not be repeated here.

[0092] This invention also provides an AI-based real-time deformation compensation device for wireless ultrasound probes, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiments of the AI-based real-time deformation compensation method for wireless ultrasound probes, for example... Figure 1 The steps S1 to S4 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0093] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the AI-based wireless ultrasound probe real-time deformation compensation device.

[0094] The AI-based real-time deformation compensation device for wireless ultrasound probes can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the AI-based real-time deformation compensation device for wireless ultrasound probes may also include input / output devices, network access devices, buses, etc.

[0095] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, 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, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the AI-based wireless ultrasound probe real-time deformation compensation device, connecting all parts of the device via various interfaces and lines.

[0096] The memory can be used to store the computer program and / or modules. The processor realizes various functions of the AI-based wireless ultrasonic probe real-time deformation compensation device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the operation of the air conditioner controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card (SMC), secure digital card (SD), flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0097] The module integrated into the AI-based wireless ultrasound probe real-time deformation compensation device, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0098] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0099] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A real-time deformation compensation method for wireless ultrasonic probes based on AI, characterized in that, The method includes: Raw image data and raw pressure data are collected, image structure feature pathways and pressure response pathways are established, structural and physical deformation trends are extracted respectively, and image structure change response map and probe pressure-induced response map are generated. The image structure change response map and the probe pressure-induced response map are fused to construct a fused feature map; wherein, the fused feature map is constructed based on the ambiguity map; the ambiguity map is obtained by local Sobel variance normalization of the original image data; Based on the fused feature map, a convolutional network model is constructed to predict the non-rigid deformation field of the current frame image and output the non-rigid deformation field; wherein the prediction also incorporates a tissue density feature map calculated based on the original image data; the non-rigid deformation field also incorporates an orientation consistency term, a boundary slip suppression term, and a tissue density adaptive term; The non-rigid deformation field is applied to the current frame image, and spatial coordinate inverse mapping and image resampling operations are performed to output the image after deformation compensation. The spatial coordinate inverse mapping includes applying the deformation displacement of the non-rigid deformation field to each original image data pixel, calculating the inverse mapping coordinates corresponding to the deformation displacement, and then obtaining new pixel values ​​from the original image data with the inverse mapping coordinates as the sampling target to obtain the inverse sampling coordinate map.

2. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The raw image data and raw pressure data specifically include: the current frame image, which is a two-dimensional grayscale image acquired by the probe's image acquisition module; the previous frame image, which is stored in the device's internal RAM buffer and is the most recently acquired image frame; and the probe pressure distribution map, which is acquired by the capacitive pressure sensor array deployed on the device housing.

3. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The generation of the image structure change response map and the probe pressure-induced response map specifically includes: the image path uses an inter-frame difference strategy to extract non-rigid structural changes, and generates the image structure change response map from the image frame difference through a lightweight convolution structure; the pressure path is used to supplement the insufficient image structure detection capability, and generates the probe pressure-induced response map by passing the interpolated and normalized pressure array map through a lightweight sensing path.

4. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The specific steps of fusing the image structure change response map and the probe pressure-induced response map to construct a fused feature map include: fusing the fuzzy map with the image structure change response map, and introducing a boundary preservation regularization term and a tissue response consistency constraint term. The boundary preservation regularization term is used to suppress abrupt changes in the image structure after fusion, and the tissue response consistency constraint term is used to enhance the consistency between the tissue deformation pattern and the spatial response of the pressure map in the fused map.

5. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 4, characterized in that, The tissue response consistency constraint is the gradient dot product of the image structure change response map and the probe pressure-induced response map.

6. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The convolutional network model is a three-layer convolutional network model: the first layer is a 3×3 convolution with 1 input channel and 32 output channels, and ReLU activation; the second layer is a 3×3 convolution with the number of channels changing from 32 to 16, and ReLU activation; the third layer is a 3×3 convolution with the number of channels changing from 16 to 2, and no activation, used to output non-rigid deformation fields.

7. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The orientation consistency term, boundary slip suppression term, and tissue density adaptive term introduced in the non-rigid deformation field specifically include: the orientation consistency term penalizes pixels whose deformation direction is inconsistent with the direction of the fused feature gradient, avoiding multi-directional cross deformation fields; the boundary slip suppression term forces the edge deformation vector change to remain smooth, limiting deformation jumps in image edge regions; and the tissue density adaptive term reduces the deformation prediction intensity of the model in areas with high image density.

8. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 1, characterized in that, The image resampling operation involves performing bilinear interpolation on the backsampled coordinate map. The bilinear interpolation is performed on any floating-point coordinate, where the four adjacent integer coordinate points are... The interpolation weights for the above, below, left, and right combinations are allocated according to standard proportions.

9. The AI-based real-time deformation compensation method for wireless ultrasonic probes according to claim 8, characterized in that, When the image resampling operation falls outside the image boundary, it uses an edge value extension method, which involves taking the nearest boundary value for compensation.

10. A real-time deformation compensation system for a wireless ultrasonic probe based on AI, characterized in that, The system includes: The feature extraction unit is used to acquire raw image data and raw pressure data, establish image structure feature pathways and pressure response pathways, extract structural and physical deformation trends respectively, and generate image structure change response maps and probe pressure-induced response maps. The feature fusion unit is used to fuse the image structure change response map and the probe pressure-induced response map to construct a fused feature map; wherein, the fused feature map is constructed based on a fuzzy map; the fuzzy map is obtained by local Sobel variance normalization of the original image data; The deformation prediction unit is used to construct a convolutional network model based on the fused feature map, predict the non-rigid deformation field of the current frame image, and output the non-rigid deformation field; wherein the prediction also incorporates a tissue density feature map calculated based on the original image data; the non-rigid deformation field also incorporates an orientation consistency term, a boundary slip suppression term, and a tissue density adaptive term; The image compensation unit is used to apply the non-rigid deformation field to the current frame image, perform spatial coordinate inverse mapping and image resampling operations, and output the deformation-compensated image; wherein, the spatial coordinate inverse mapping includes applying the deformation displacement of the non-rigid deformation field to each original image data pixel, calculating the inverse mapping coordinates corresponding to the deformation displacement, and then obtaining new pixel values ​​from the original image data with the inverse mapping coordinates as the sampling target to obtain the inverse sampling coordinate map.

Images