Liver steatosis quantification method, device, system, apparatus, and medium

Abstract

The application discloses a liver steatosis quantitative detection method, device and system. The method comprises the following steps: acquiring an ultrasonic radio frequency signal, reconstructing the ultrasonic radio frequency signal into a fan-shaped B-mode gray image through DSC forward transformation, and pre-storing a mapping index table of pixel points to a radio frequency data domain; segmenting a liver region of the gray image by using a deep learning segmentation model to obtain a fan-shaped binary mask; performing DSC inverse transformation on the fan-shaped binary mask based on the mapping index table to obtain a radio frequency domain mask matrix; screening an effective depth interval and performing frequency spectrum transformation to obtain a power spectrum energy matrix; and then performing frequency normalization and linear fitting to obtain an attenuation coefficient representing a liver steatosis degree. The application completes robust deep learning segmentation in an intuitive and low-dimensional fan-shaped image domain, accurately returns to the radio frequency domain through an invertible index table mapping, and takes into account segmentation robustness, calculation efficiency and signal integrity, so that high-precision liver steatosis quantitative detection is realized under complex clinical conditions.

Description

Technical Field

[0001] This invention relates to the fields of ultrasound medical image processing and quantitative analysis of liver fat, and particularly to a method, device, system, equipment and medium for quantitative detection of liver steatosis based on deep learning and ultrasound radio frequency signals. Background Technology

[0002] Hepatic steatosis is a common pathophysiological phenomenon in liver diseases, and its early diagnosis and severity assessment have significant clinical value. While biochemical analysis is considered the gold standard for measuring hepatic steatosis, it is an invasive procedure and may be subject to sampling bias and sample variability. Ultrasound imaging technology, due to its portability, low cost, and non-invasiveness, offers an attractive solution for non-invasively quantifying liver fat content and accurately detecting hepatic steatosis.

[0003] In recent years, an increasing number of researchers have begun to use ultrasound measurement methods based on RF (Radio-Frequency) signals for the diagnosis and assessment of hepatic steatosis. Because RF signals carry richer information, they can more accurately reflect tissue characteristics, enabling precise measurement and quantitative analysis of hepatic steatosis. Chinese patent CN116671985B discloses an image processing-based ultrasound measurement method for hepatic steatosis, which searches for liver boundaries pixel-by-pixel on a grayscale image and distinguishes interfering signal regions using a threshold curve. Chinese patent CN118512209B discloses a quantitative detection method for hepatic steatosis based on ultrasound radio-frequency signals, which converts the radio-frequency signal into a grayscale image, searches for liver boundaries pixel-by-pixel, generates a mask template, and then performs spectral transformation using a depth interval matrix to calculate the attenuation coefficient.

[0004] However, the existing methods mentioned above all employ traditional gray-level gradient and boundary search, threshold curve, and other algorithms in the liver region identification stage. These methods rely on features designed by human experience and are sensitive to image quality. Under complex clinical acquisition conditions such as poor probe position or angle, poor contact with the skin, and rib occlusion, the segmentation accuracy drops significantly. Furthermore, traditional methods usually rely on only a small number of typical images for parameter tuning, resulting in limited generalization ability. The effectiveness of these algorithms depends on professional operators and a clean liver interface. Under conditions such as rib acoustic shadowing or intestinal gas interference, scattered and fragmented masks are easily generated, making it impossible to calculate the attenuation coefficient with physical significance. Objectively, this limits their clinical applicability to large medical institutions and hinders the widespread adoption of the technology.

[0005] On the other hand, if liver segmentation is performed directly in the radiofrequency signal domain, the high dimensionality (e.g., 64×4096) and large dynamic range (approximately 10) of the radiofrequency signal would be problematic. 4 ~ 10 4The computational demands are enormous (on a scale of several orders of magnitude), and the lack of intuitive visual feedback hinders model training and manual annotation. Furthermore, while deep learning segmentation methods (such as nnU-Net) have been used for medical image segmentation, for example, to automatically segment and select liver parenchyma regions on chest CT images to classify hepatic steatosis, these methods target the CT image domain and do not involve the reversible transformation from ultrasound radiofrequency signals to the image domain. Therefore, they cannot be directly used to calculate radiofrequency attenuation coefficients while maintaining physical scale.

[0006] Therefore, there is a need for a highly robust, efficient quantitative detection method for hepatic steatosis that can maintain signal integrity. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method for quantitative detection of hepatic steatosis, comprising the following steps: Step S1: Acquire at least one set of ultrasound radio frequency signals, reconstruct the ultrasound radio frequency signals into a sector-shaped B-mode grayscale image through digital scan transformation (DSC), and calculate and store a mapping index table from each pixel of the sector-shaped B-mode grayscale image to the ultrasound radio frequency signal data domain during the DSC process. Step S2: Segment the liver region of the sector-shaped B-mode grayscale image using a pre-trained deep learning segmentation model to obtain a sector-shaped binary mask. Step S3: Perform an inverse DSC on the sector-shaped binary mask based on the mapping index table, mapping the sector-shaped binary mask into a radio frequency domain mask matrix of the same size as the ultrasound radio frequency signal. Step S4: Filter effective depth intervals according to the radio frequency domain mask matrix, and perform spectral transformation on the effective depth intervals to obtain a power spectrum energy matrix. Step S5: Perform frequency normalization on the power spectrum energy matrix and fit it to obtain an attenuation coefficient characterizing the degree of hepatic steatosis.

[0008] Further, step S1 includes: calculating the sampling depth scale, the physical height of the sector image, the pixel physical scale, and the coordinates of the sector center based on the sector geometric parameters and the output image size. For each pixel, the physical offset relative to the sector center is calculated, polar coordinate transformation is performed, the sampling point index and scan line index are mapped, interpolation parameters and four nearest neighbor indexes are obtained to form the mapping index table, and interpolation is performed to obtain the gray value of the pixel. Utilizing the mirror symmetry of the sector geometry, only half of the sector's mapping index is calculated and stored, while the other half is obtained through mirror symmetry. In step S3, a zero matrix of the same size as the ultrasound radio frequency signal is created as the radio frequency domain mask matrix. For each pixel in the sector binary mask that has a value of 1 and is located within the effective data area, its corresponding four nearest neighbor indexes are obtained through the mapping index table, and the corresponding index position in the radio frequency domain mask matrix is ​​assigned a value of 1. The digital scan conversion inverse transform only utilizes the mapping index table of half of the sector.

[0009] Furthermore, after step S1, the method further includes performing foreground adaptive histogram mapping on the sector B-mode grayscale image: pixels with grayscale values ​​less than or equal to the background threshold are determined as background and remain unchanged; for foreground pixels, their normalized histogram and cumulative distribution function are calculated, and a target normal distribution is set, and the foreground grayscale is mapped to the normal distribution through histogram matching.

[0010] Furthermore, the deep learning segmentation model is a two-dimensional convolutional neural network segmentation model with an encoder-decoder architecture. The encoder includes a preset number of downsampling stages, and the decoder includes a corresponding upsampling stage. Instance normalization layers and nonlinear activation functions are used to accelerate convergence, and the model is trained using a combination of Dice loss and cross-entropy loss. Before inference, the deep learning segmentation model performs cubic spline interpolation to scale the fan-shaped B-mode grayscale image to a fixed input size. After outputting the mask, it restores the original size through nearest neighbor interpolation. The segmentation results are then labeled with connected components, and the largest connected component is retained to eliminate small-area false positives.

[0011] Further, before step S4, the method includes: based on the radio frequency domain mask matrix, removing low-echo noise regions of blood vessels and bile ducts to obtain an optimized mask matrix. Step S4 includes: dividing the ultrasound radio frequency signal into preset depth intervals; for each radio frequency signal line, if all sampling points within a certain depth interval are located within the optimized mask matrix, then the depth interval is encoded as 1; otherwise, it is encoded as 0, generating a depth interval matrix. A fast Fourier transform is performed on the depth intervals encoded as 1 to calculate the power spectrum and convert it to dB units; the power spectra of different radio frequency signal lines within the same depth interval are averaged to obtain a power spectrum energy matrix; linear interpolation and Gaussian filtering are performed on the missing data in the middle of the power spectrum energy matrix, and the smallest bounding rectangle of the region greater than a preset multiple of its maximum value is selected as the depth range and frequency range used to calculate the attenuation coefficient.

[0012] Further, step S5 includes: Plot energy-depth polygons with energy as the vertical axis and depth measurement value as the horizontal axis. For each point on each polygon, calculate the ratio of energy to the corresponding frequency. Calculate the average of the energy-frequency ratios of points with the same depth measurement value on different polygons. Use the average value and the corresponding depth measurement value to perform linear fitting, and the resulting slope is the attenuation coefficient characterizing the degree of hepatic steatosis.

[0013] Furthermore, the present invention also provides a detection device for implementing the above method, a detection system including an ultrasonic probe and the detection device, an electronic device including a memory and a processor, and a computer-readable storage medium storing a corresponding computer program.

[0014] The beneficial effects of this invention are as follows: First, it exhibits strong segmentation robustness. The deep learning segmentation model maintains high accuracy on large-scale datasets containing rib acoustic shadows, various tissue interfaces, and adverse imaging conditions, fundamentally solving the application bottleneck of traditional methods failing under complex working conditions. Second, it boasts high processing efficiency. This invention pre-calculates a digital scan conversion index table, requiring only table lookup and assignment operations in the forward and inverse mapping processes, resulting in low time complexity. Furthermore, it converts high-dimensional, high-dynamic-range radio frequency signals into low-dimensional, intuitively readable images for segmentation, significantly reducing the feature dimension of artificial intelligence processing. Third, it ensures good signal integrity. The inverse mapping accurately returns to the original radio frequency signal domain, fully preserving the sampling rate, dynamic range, and physical scale, ensuring that spectral analysis and attenuation coefficient calculation are unaffected. Fourth, it provides a fully automated process, eliminating the need for manual delineation of regions of interest and providing end-to-end output of quantitative fat indicators. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram illustrating the steps of a quantitative detection method for hepatic steatosis provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of the processing flow of the quantitative detection method for hepatic steatosis provided in an embodiment of the present invention.

[0017] Figure 3 A comparison diagram of the effects before and after ground foreground histogram mapping is provided for embodiments of the present invention.

[0018] Figure 4 The image shows a typical liver segmentation effect in an embodiment of the present invention.

[0019] Figure 5 This is a schematic diagram of the structure of the quantitative detection device for liver steatosis provided in an embodiment of the present invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings, but the invention can be implemented in many different forms and is not limited to the embodiments described herein. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] like Figure 1 As shown, this embodiment of the invention provides a method for quantitative detection of hepatic steatosis, including the following steps: Step S1: Acquire at least one set of ultrasonic radio frequency signals, reconstruct the ultrasonic radio frequency signals into a fan-shaped B-mode grayscale image through digital scan conversion forward transformation, and calculate and store the mapping index table of each pixel point of the fan-shaped B-mode grayscale image to the ultrasonic radio frequency signal data domain during the digital scan conversion forward transformation process. Step S2: Use a pre-trained deep learning segmentation model to segment the liver region of the fan-shaped B-mode grayscale image to obtain a fan-shaped binary mask; Step S3: Based on the mapping index table, perform digital scan conversion inverse transformation on the sector binary mask to map the sector binary mask into a radio frequency domain mask matrix of the same size as the ultrasonic radio frequency signal; Step S4: Select the effective depth interval according to the radio frequency domain mask matrix, and perform spectral transformation on the effective depth interval to obtain the power spectrum energy matrix; Step S5: The power spectrum energy matrix is ​​frequency normalized and fitted to obtain the attenuation coefficient characterizing the degree of hepatic steatosis.

[0022] refer to Figure 2 The quantitative detection method for hepatic steatosis provided in this embodiment has the following overall processing chain: input the original RF signal → forward DSC (Digital Scan Conversion) to obtain a fan-shaped B-mode grayscale image → deep learning segmentation to obtain a fan-shaped liver binary mask → inverse DSC to obtain radio frequency domain mask matrix 1 → remove noise signals such as blood vessels based on image processing methods to obtain radio frequency domain mask matrix 2 → select effective depth intervals based on the mask matrix to obtain depth interval matrix encoding → step 6 power spectrum calculation, frequency normalization and linear fitting → output the hepatic steatosis attenuation coefficient β (unit: dB / (cm·MHz)).

[0023] Specifically, in some embodiments of the present invention, the acquired single set of raw radio frequency signals includes L lines, with S sampling points per line, for example, L=64 and S=4096, meaning that the effective length of each line is 4096 sampling points. To increase the sampling range and obtain more comprehensive information about the liver, multiple samplings can be performed on each subject, with each sampling containing multiple sets (e.g., 4 sets) of radio frequency signals at a certain time interval; during sampling, the probe can be placed still and rely on the liver's own movement for acquisition, or the probe can be moved for acquisition, and the information from multiple sampling locations can be superimposed for subsequent processing.

[0024] In this embodiment of the invention, in step S1, the overall generation pipeline of the sector-shaped B-mode grayscale image can be divided into the following stages (in order of execution): (1) Read the original RF data and reorganize it into 4 groups, each group consisting of 64 rows × 4096 columns of RF data.

[0025] (2) Preprocessing stage: The RF data of each group is processed line by line in the following order: bandpass filtering → time gain compensation → Hilbert envelope detection → logarithmic compression → normalization.

[0026] (3) DSC forward transform converts the preprocessed rectangular data into a fan-shaped image.

[0027] (4) Adaptive histogram enhancement for cropping region of interest and foreground.

[0028] (5) Output the B-mode grayscale image and save it as a .PNG image.

[0029] Before the DSC forward conversion, the radio frequency signal is preprocessed to condition the signal before it enters the digital scanning conversion. The preprocessing is performed in the following order: bandpass filtering → time gain compensation → Hilbert envelope detection → logarithmic compression → normalization. The flowchart of each step is as follows: (1) Bandpass filtering Using a fourth-order Butterworth bandpass filter for bidirectional filtering, its difference equation can be simplified as follows: , Among them, coefficient , By scipy.signal.butter(4, [ , The function `btype='band'` is generated based on the normalized cutoff frequency, which is: , in, This is the low-end cutoff frequency of the bandpass filter. Let the high-end cutoff frequency of the bandpass filter be , and assume the probe center frequency is . ,but and They are respectively The first preset multiple and the second preset multiple are determined based on the probe bandwidth characteristics. In a preferred embodiment, the first preset multiple is 0.5~0.8, and the second preset multiple is 1.2~1.5. Sampling rate .

[0030] (2) Time gain compensation To compensate for the exponential attenuation of ultrasound waves as they propagate through tissue with increasing depth (time), time gain compensation is applied to the signal:

[0031] in, (n is the sampling point index), α is the preset attenuation compensation coefficient, which can be preset within a reasonable range according to the average attenuation coefficient of the tissue to compensate for the signal attenuation caused by the increase in depth.

[0032] (3) Hilbert envelope detection Constructing analytic signals ,in, The Hilbert transform is represented by its modulus, which gives the envelope: .

[0033] (4) Logarithmic compression To compress the dynamic range, the envelope signal is mapped to the logarithmic domain:

[0034] Where ε is a constant to prevent the logarithmic parameter from being zero; in this embodiment, it is taken as... .

[0035] (5) Normalization Linearly map the logarithmic domain signal to [0, 255]: .

[0036] The DSC forward transform will generate high-dimensional, high-dynamic-range (approximately 10) 4 ~ 10 4 The radio frequency signal is converted into a visually readable image (0-255 grayscale values) for doctors, which facilitates manual annotation and model training, significantly reduces the feature dimension of subsequent artificial intelligence processing, and retains the physical depth and angle information corresponding to each pixel for inverse transformation. To this end, physical scale parameters are pre-calculated based on sector geometric parameters (dead zone distance, sector angle, sampling rate, number of effective sampling points) and output image size. The specific calculation process is as follows: (1) Sampling depth scale, i.e., the physical distance (mm / sampling point) corresponding to each sampling point in the depth direction: , in, The sampling rate is expressed in Hz; 1560 is the reference value for the speed of sound in soft tissue (m / s), which is divided by 2 to represent the round-trip propagation of ultrasound.

[0037] (2) Physical height of the sector image (mm): , in, The total physical height is expressed in mm. Since the top of the ultrasonic fan-shaped image is an arc, this formula calculates the vertical span from the top (center vertex) to the bottom (end of effective sampling depth) of the image. The dead zone distance, measured in mm, is the depth from the probe surface to the start of the effective echo. Due to strong reflections, electrical noise, and probe ringing effects in the near field, the echo signal in this region is unreliable and needs to be skipped during imaging. Clinically, it is typically 5–10 mm. In this embodiment, it is set as follows: =0 indicates no dead zone, but the subsequent RF signal only takes the range of 30mm to 100mm for attenuation estimation; θ is the sector scanning angle in radians, with a default value of 50 / 180×π; This represents the number of valid sampling points.

[0038] (3) Pixel physical scale, i.e., physical spacing between sampling points (mm / pixel):

[0039] in, The default pixel height for the output image is 480 pixels in this embodiment.

[0040] (4) The coordinates (pixels) of the center of the sector in the image:

[0041]

[0042] in, The default pixel width of the output image is 420 pixels in this embodiment; the negative sign in Cy is used to place the virtual probe vertex (or dead zone start point) above the visible area of ​​the image, with the center of the circle located below the fan-shaped vertex (the y-axis is positive downwards).

[0043] During the initialization phase of the DSC forward transform, the four-point index and interpolation weights corresponding to all pixels (x, y) in the RF data domain are pre-calculated based on the sector geometry parameters and the output image size, and stored as a mapping index table. Subsequent forward or inverse transforms of all RF signals are performed directly from this table, avoiding redundant calculations. For a given image pixel P(x, y), its floating-point coordinates F(l, s) mapped to the original scan line data are calculated using the following inverse mapping formula: (1) Physical offset (unit: mm):

[0044]

[0045] (2) Polar coordinate transformation, polar radius r and polar angle θ:

[0046]

[0047] Since the depth direction of ultrasound images is usually downward along the Y-axis, atan2,dy is used to represent the depth direction offset.

[0048] (3) Map to sampling point index (depth direction, maximum 4096):

[0049] (4) Map to scan line index (angle direction, maximum number of scan lines is 64, total sector angle is 50° by default):

[0050] Where φ is the total angle of the sector. This represents the total number of scan lines for the RF signal lines.

[0051] Interpolation parameter calculation: integer part ; weight of the fractional part (quantized to an integer of 0 to 255): ; validity flag uiAlpha: if 0<uiLine<lineCount and 0<uiSample<sampleCount, then uiAlpha=255, otherwise uiAlpha=0.

[0052] The four nearest neighbor point indices (top-left tl, top-right tr, bottom-left bl, bottom-right br) are: , , , .

[0053] Boundary processing: when uiLine=0, tl=bl= 1; when uiSample=0, tl=tr= 1; when uiSample≥sampleCount -1, bl=br= 1, and invalid indices take a value of 0 during interpolation.

[0054] By utilizing the symmetry of sector geometry, it is only necessary to calculate and store the mapping indices for half of the sector region (e.g., the left half sector), and the mapping for pixels in the other half is directly obtained by mirror symmetry copying, thereby further reducing the amount of calculation and storage.

[0055] The forward DSC conversion method is as follows: for each pixel (x, y), look up the table to obtain the original radio frequency intensity of four neighboring points (values converted through a grayscale mapping table, recorded as TL, TR, BL, BR, with values ranging from 0 to 255) and interpolation weights, then perform bilinear interpolation. is used to represent the horizontal offset, which is the fractional offset of the mapping point in the horizontal / column direction (the quantized value is 0~255), use to represent the vertical offset, which is the fractional offset of the mapping point in the vertical / row direction (the quantized value is 0~255), let the pixel values of the four nearest neighbor points be respectively , , , , then the formula for the bilinear interpolation result is:

[0056]

[0057]

[0058] Using >>8 to perform fast fixed-point arithmetic, the quantized weights can be represented as: , , , , Final pixel grayscale gray = min( (255). It should be noted that other interpolation methods (such as nearest neighbor interpolation, bicubic interpolation, etc.) can also achieve the above mapping, and all of them fall within the protection scope of this invention. After the above processing, a fan-shaped B-mode grayscale image (e.g., with a size of 420×480 pixels) is obtained, and the physical depth and angle information corresponding to each pixel is preserved.

[0059] To improve the ability of deep learning segmentation models to identify liver boundaries, this embodiment performs foreground adaptive histogram mapping on the sector-shaped B-mode grayscale image. First, pixels with grayscale values ​​less than or equal to a background threshold (typically 1) are classified as background and left unchanged. For foreground pixels, their normalized histogram and cumulative distribution function are calculated, and a target normal distribution is set (peak grayscale range 64~128, standard deviation automatically calculated based on the standard deviation and dynamic range of the foreground pixels). Histogram matching is then used to map the foreground grayscale values ​​to this normal distribution. This process effectively enhances the contrast between the liver region and surrounding tissues without affecting the background. The specific process is as follows: (1) Let the set of foreground pixels be Its grayscale value ranges from 0 to 255, and the grayscale threshold for distinguishing between background and foreground is 1 (grayscale values ​​≤ 1 are considered background). The foreground normalized histogram (probability density function) is obtained using np.histogram(..., density=True).

[0060] Obtain the foreground cumulative distribution function (CDF) using np.cumsum(): .

[0061] (2) The target distribution is selected as a normal distribution with mean μ and standard deviation σ, and cut off in the interval [0, 255]. Its target probability density function is... ,in, The standard normal density function, and the normalization constant. , making Target cumulative distribution function In this embodiment, the target peak grayscale range is 64–128.

[0062] (3) Histogram matching mapping: For each foreground gray level g, find a new gray level g′ such that .

[0063] (4) Automatic calculation of standard deviation σ: Based on the statistical characteristics of the foreground pixels and the peak range specified by the user, σ is automatically calculated using a weighted linear combination:

[0064] in: is the sample standard deviation of the foreground grayscale values, and w is the width of the target peak interval (i.e., the upper limit of the target peak interval minus the lower limit, which is 64 by default in this embodiment). Let be the dynamic range of the foreground. This formula is an empirical heuristic rule, where: The term ensures that the width portion of the target distribution inherits the characteristics of the original distribution, preventing excessive compression; the term w ensures that the main peak of the target distribution is located within the desired grayscale range, preventing the distribution from becoming too sharp due to excessively small σ. The term, as a measure of dynamic range, can appropriately broaden the distribution in low-contrast images to avoid loss of detail after mapping. α, β, and γ are preset weighting coefficients. They are dynamically determined after a weighted linear combination based on the statistical characteristics of the foreground pixels, and are limited to preset upper and lower thresholds to ensure a visually reasonable histogram shape. To highlight the original distribution characteristics, the weight of α can be greater than that of β and γ, for example, α=0.6, β=0.3, γ=0.1. It should be noted that this combination of weighting coefficients is only one optional implementation of this invention. Those skilled in the art can replace them with equal weights (i.e., α=β=γ) or other arbitrary heuristic combinations according to the actual image contrast, and can also achieve the histogram mapping function of this invention. These variations are all within the protection scope of this invention. σ_min and σ_max are preset upper and lower thresholds, and their specific values ​​can be calibrated or set empirically by those skilled in the art based on the required contrast enhancement effect through a limited number of experiments. Figure 3 The image shown is a comparison of the effects before and after foreground histogram mapping.

[0065] In this embodiment of the invention, a deep learning segmentation model with adaptive hyperparameter optimization (e.g., a model based on nnU-Net 2D configuration) is used in step S2. This model can automatically adjust hyperparameters such as network depth, patch size, and normalization strategy during training, adapting to different datasets and imaging conditions without manual parameter tuning. Before inference, the deep learning segmentation model performs cubic spline interpolation to scale the fan-shaped B-mode grayscale image to a fixed input size (e.g., 256×128 pixels), and after outputting the mask, restores the original size through nearest neighbor interpolation.

[0066] In one specific embodiment, the deep learning segmentation model is a two-dimensional convolutional neural network segmentation model employing an encoder-decoder architecture (e.g., a basic configuration based on nnU-Netv2). The encoder includes a preset number of downsampling stages, and the decoder includes corresponding upsampling stages. As downsampling proceeds, the number of feature map channels increases incrementally (the channel doubling base is preset based on model capacity requirements). Small-sized convolutional kernels are used to extract local features, and instance normalization and non-linear activation functions are employed to accelerate convergence. The loss function uses a combination of Dice loss and cross-entropy loss. Dice loss It directly optimizes the segmentation evaluation metric, is robust to class imbalance, and uses cross-entropy loss. This method provides pixel-level classification supervision and increases training stability; the combination of these two approaches balances segmentation accuracy and training stability. During training, a preset initial learning rate is used, which is gradually reduced to zero using a polynomial decay strategy. A mini-batch training strategy is employed, with the batch size preset based on the GPU memory capacity. The input image is cropped into pixel blocks of a preset size. Data augmentation employs online dynamic strategies, including random rotation within a preset angle range, random scaling within a preset scale range, and grayscale and spatial transformations such as Gaussian noise injection, Gaussian blur, and brightness / contrast adjustment. It should be noted that various deep learning segmentation networks can actually be selected; the model described in this embodiment has achieved good results and does not constitute a limitation of the invention.

[0067] For training data, this embodiment used 10,993 fan-shaped B-mode grayscale images from 284 subjects acquired by three operators to ensure coverage of a wide range of clinical variations. Covered scenarios included, but were not limited to, typical tissue interfaces such as pure lung, hepatopulmonary junction, pure liver, liver-kidney, liver-intestine, and pure intestine; images with rib shadowing interference; and images acquired from different locations, probe angles, and depths. All images were annotated pixel-wise by an ultrasound physician to define the liver region. The dataset was divided into a training set of 7,967 images, a validation set of 1,992 images, and a test set of 1,034 images. The model performance on the test set was as follows: for all images, the Dice coefficient was 0.964, the accuracy was 0.979, and the IoU was 0.942; for only complete liver images, the Dice coefficient was 0.996, the accuracy was 0.996, and the IoU was 0.993. Therefore, this invention extends the segmentation task to large-scale data covering complex scenarios, and proves the model's convergence and high accuracy under extreme disturbances through independent test sets, fundamentally solving the application bottlenecks of traditional methods. Figure 4 The image shown is a liver segmentation result in a typical scenario.

[0068] In this embodiment of the invention, the post-processing of the segmentation results includes: marking connected components and comparing areas of the segmentation results, retaining the largest connected component to eliminate small false positives; optionally, performing morphological closing operations to fill small holes.

[0069] In step S3 of this embodiment, the sector-shaped binary mask (liver region) obtained from deep learning segmentation is reverse-mapped back to the original RF data domain to generate a mask matrix of the same size as the RF signal, which is used for subsequent depth interval filtering and spectral analysis. During the initialization phase of the DSC forward transform, all pixels are pre-calculated based on the sector geometric parameters and the output image size. The corresponding four-point index of the RF data field is stored as an index table. Subsequent forward or inverse transforms of all RF signals are performed directly from this table, avoiding redundant calculations. This mapping is accurate and reversible, with no resampling error.

[0070] The DSC inverse transform does not directly calculate the original sampling coordinates from the sector image coordinates using analytical geometric formulas. Instead, it utilizes a pre-computed forward mapping index table for back projection: mapping each foreground pixel in the sector image back to its four nearest neighbors on the original rectangular data grid and accumulating their labels. Let the forward mapping be the function F:(x,y)→(l,s), and its four nearest neighbor integer sampling points be denoted as N(x,y)={TL,TR,BL,BR}. The inverse transform process is as follows: Input: Binary sector mask image (1 represents the liver), with dimensions of Hpixel × Wpixel; Pre-computed mapping table: For each sector pixel (x,y), store the linear indices N(x,y) = {TL,TR,BL,BR} of its four nearest neighbors in the original rectangular raster. Create an array of all zeros of size lineCount × sampleCount (i.e., L × S). As a radio frequency domain mask matrix. For all masks in the image that satisfy... For a foreground pixel whose value is 1 and uiAlpha > 0 (meaning the pixel is within the valid data area), obtain four neighboring pixel indices idx1, idx2, idx3, and idx4 from the index table T(x,y), and then execute: This involves projecting the pixel onto the four neighboring points of the original rectangular grid and setting them to 1. The inverse digital scan transformation also utilizes only the mapping index table of the left half of the sector. This mapping is accurate and reversible, with no resampling error, and outputs a binary matrix of the same size as the original RF signal, completely preserving the original physical scale without resampling error. As post-processing, morphological closing operations are performed on the generated RF domain mask matrix to fill the holes.

[0071] In this embodiment of the invention, the method further includes the following steps before step S4: Based on the radio frequency domain mask matrix By utilizing local grayscale statistics or morphological operations, low-echo noise regions such as blood vessels and bile ducts are removed, resulting in an optimized mask matrix. The interference signal region refers to an area in the image that is particularly bright or dark compared to the surrounding tissue, caused by blood vessels or other tissue structures within the liver.

[0072] In step S4, the RF signal is divided into preset depth intervals. For each RF line, if all sampling points within a certain depth interval are located... If the depth is within a certain range, the interval is encoded as 1; otherwise, it is 0, generating a depth interval matrix D. Based on the effective depth of the RF domain mask matrix, the sampling start and end points are preset within a reasonable sampling point range. Several overlapping or non-overlapping depth intervals are divided in the depth direction using a preset interval length and a preset sliding step size. In the frequency dimension, based on the probe's effective bandwidth, the frequency band between the lower and upper frequency limits is selected, and frequency measurement points are extracted at preset frequency intervals. Furthermore, the depth interval matrix can be post-processed to remove short rows and columns. For example, each row / column can be traversed, and the number of pixels with a value of 1 can be counted. If the number of pixels in a row / column is less than a preset multiple (e.g., 0.4) of the maximum number of pixels in that direction, then all pixels in that row / column are set to 0.

[0073] In step S5, a Fast Fourier Transform is performed on each depth interval coded as 1 to calculate the power spectrum and take dB units. Since the attenuation characteristics are the same on each RF line, the power spectra of different RF lines within the same depth interval are averaged. Combining the depth measurement points in the depth direction and the frequency measurement points on the power spectrum, the power spectrum energy matrix P(f,d) is obtained. For missing data in the middle of the matrix, linear interpolation is performed using adjacent data, and then Gaussian filtering is used to reduce noise interference. Since the probe frequency is 3.5MHz, experiments show that the energy is concentrated between 1.5MHz and 4.5MHz, while the depth range of the data acquisition is 2cm to 10cm. Since the signal attenuation characteristics are most pronounced in the high-energy region of the energy matrix, the high-energy region can be automatically selected, i.e., a suitable depth and frequency range can be chosen to reduce data processing volume and improve processing speed. First, the global maximum value of the processed energy matrix is ​​found. Then, a preset percentage threshold is set (e.g., 60%~80% of the maximum value, depending on the signal-to-noise ratio requirement). Connected regions with energy values ​​greater than this threshold are selected, and their smallest bounding rectangle is taken as the preferred depth and frequency range for calculating the attenuation coefficient. After selecting the high-energy region from the energy matrix, a polygonal line is plotted with energy as the vertical axis and depth measurement value as the horizontal axis, corresponding to each frequency measurement point. The relationship between power spectrum energy and depth and frequency can be expressed by the following formula: Where f refers to a certain frequency measurement point, and These refer to a depth measurement point at that frequency and its corresponding energy. and This refers to the energy of the adjacent depth measurement point at that frequency, and β refers to the attenuation coefficient. Since the ultrasonic signal travels a round trip, d is multiplied by 2. Let dBf be the value of the data point on the broken line divided by the corresponding frequency, with units of dB / MHz. Combining the above two equations, we get: First, divide the data points on each broken line by the corresponding frequency to obtain... Then, the average data points at the same depth on different broken lines are averaged, and then linear fitting is performed. The final slope obtained is the attenuation coefficient β.

[0074] This invention pre-calculates a digital scan conversion index table, and the inverse mapping process only requires table lookup and assignment operations, with a time complexity of O(sector pixel count), avoiding redundant calculations. The trained deep learning model is exported to ONNX format and inference is performed using the ONNX Runtime to further reduce inference latency and facilitate deployment to edge computing devices. In a Python environment (CPU inference), for a set of radio frequency signals (64 lines × 4096 sampling points / line), the overall time is approximately 400ms, of which AI segmentation accounts for approximately 20ms and inverse mapping accounts for approximately 60ms, meeting the real-time clinical requirements; subsequent deployment of the code to C++ can further improve computational efficiency.

[0075] Furthermore, the current deep learning model can be replaced with a more efficient inference model, or model pruning, model distillation, and other operations can be performed to reduce model parameters and computational load, thereby achieving real-time performance on edge devices.

[0076] refer to Figure 5 The present invention also provides a quantitative detection device for hepatic steatosis, used to implement the above method, comprising: an ultrasound radio frequency signal acquisition module for acquiring at least one set of ultrasound radio frequency signals; an image reconstruction module for reconstructing the ultrasound radio frequency signals into a fan-shaped B-mode grayscale image through digital scan conversion forward transformation, and pre-calculating and storing a mapping index table of each pixel point to the ultrasound radio frequency signal data domain; a deep learning segmentation module for segmenting the liver region of the fan-shaped B-mode grayscale image using a pre-trained deep learning segmentation model to obtain a fan-shaped binary mask; a mask inverse mapping module for performing a digital scan conversion inverse transformation on the fan-shaped binary mask based on the mapping index table to obtain a radio frequency domain mask matrix of the same size as the ultrasound radio frequency signal; a power spectrum transformation module for selecting an effective depth interval according to the radio frequency domain mask matrix and performing a spectrum transformation to obtain a power spectrum energy matrix; and an attenuation coefficient calculation module for performing frequency normalization processing and linear fitting on the power spectrum energy matrix to obtain an attenuation coefficient characterizing the degree of hepatic steatosis.

[0077] The above-described device embodiments and method embodiments belong to the same inventive concept. Details not described in detail in the device embodiments can be found in the above-described method embodiments. Those skilled in the art will understand that the modules of the device in the embodiments can be adaptively changed, combined into one module or split into multiple sub-modules, and installed in one or more devices.

[0078] This invention also provides a quantitative detection system for hepatic steatosis, including an ultrasound probe and the aforementioned detection device.

[0079] This invention also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described method.

[0080] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0081] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. The use of the words first, second, third, etc., does not indicate any order and can be interpreted as names.

Claims

1. A method for quantitative detection of hepatic steatosis, characterized in that, Includes the following steps: Step S1: Acquire at least one set of ultrasonic radio frequency signals, reconstruct the ultrasonic radio frequency signals into a fan-shaped B-mode grayscale image through digital scan conversion forward transformation, and calculate and store the mapping index table of each pixel point of the fan-shaped B-mode grayscale image to the ultrasonic radio frequency signal data domain during the digital scan conversion forward transformation process. Step S2: Use a pre-trained deep learning segmentation model to segment the liver region of the fan-shaped B-mode grayscale image to obtain a fan-shaped binary mask; Step S3: Based on the mapping index table, perform digital scan conversion inverse transformation on the sector binary mask to map the sector binary mask into a radio frequency domain mask matrix of the same size as the ultrasonic radio frequency signal; Step S4: Select the effective depth interval according to the radio frequency domain mask matrix, and perform spectral transformation on the effective depth interval to obtain the power spectrum energy matrix; Step S5: The power spectrum energy matrix is ​​frequency normalized and fitted to obtain the attenuation coefficient characterizing the degree of hepatic steatosis. Step S1 includes: Based on the sector geometric parameters and the output image size, calculate the sampling depth scale, the physical height of the sector image, the pixel physical scale, and the coordinates of the sector center. For each pixel, calculate the physical offset relative to the center of the sector, perform polar coordinate transformation, map to obtain the sampling point index and scan line index, obtain the interpolation parameters and the four nearest neighbor indexes to form the mapping index table, and perform interpolation to obtain the gray value of the pixel. By utilizing the mirror symmetry of sector geometry, only the mapping index of half of the sector is calculated and stored, while the other half is obtained through mirror symmetry; In step S3, a zero matrix of the same size as the ultrasonic radio frequency signal is created as the radio frequency domain mask matrix. For each pixel in the sector binary mask that has a value of 1 and is located in the effective data area, the indexes of its four nearest neighbors are obtained through the mapping index table, and the corresponding index positions in the radio frequency domain mask matrix are assigned a value of 1. The digital scan conversion inverse transform only uses the mapping index table of half of the sector area. Following step S1, the method further includes performing foreground adaptive histogram mapping on the sector-shaped B-mode grayscale image: Pixels with gray values ​​less than or equal to the background threshold are identified as background and remain unchanged; for foreground pixels, their normalized histogram and cumulative distribution function are calculated, and a target normal distribution is set. The foreground gray values ​​are mapped to this normal distribution through histogram matching. The deep learning segmentation model is a two-dimensional convolutional neural network segmentation model with an encoder-decoder architecture. The encoder includes a preset number of downsampling stages, and the decoder includes a corresponding upsampling stage. Instance normalization layers and nonlinear activation functions are used to accelerate convergence, and the model is trained using a combination of Dice loss and cross-entropy loss. Before inference, the deep learning segmentation model performs cubic spline interpolation to scale the sector B-mode grayscale image to a fixed input size. After outputting the mask, it restores the original size through nearest neighbor interpolation. The segmentation results are then labeled with connected components, and the largest connected component is retained to eliminate small-area false positives.

2. The method for quantitative detection of hepatic steatosis according to claim 1, characterized in that, The procedure preceding step S4 also includes: Based on the radio frequency domain mask matrix, the low-echo noise regions of blood vessels and bile ducts are removed to obtain the optimized mask matrix; Step S4 includes: dividing the ultrasonic radio frequency signal into preset depth intervals; for each radio frequency signal line, if all sampling points in a certain depth interval are located within the optimized mask matrix, then the depth interval is encoded as 1, otherwise it is encoded as 0, and a depth interval matrix is ​​generated. A fast Fourier transform is performed on the depth interval coded as 1 to calculate the power spectrum and convert it to dB units. The power spectra of different radio frequency signal lines within the same depth interval are averaged to obtain a power spectrum energy matrix. Linear interpolation and Gaussian filtering are performed on the missing data in the middle of the power spectrum energy matrix. The smallest bounding rectangle of the region that is greater than a preset multiple of its maximum value is selected as the depth range and frequency range for calculating the attenuation coefficient.

3. The method for quantitative detection of hepatic steatosis according to claim 1, characterized in that, Step S5 includes: Plot energy-depth polygons with energy as the vertical axis and depth measurement value as the horizontal axis. For each point on each polygon, calculate the ratio of energy to the corresponding frequency. Calculate the average of the energy-frequency ratios of points with the same depth measurement value on different polygons. Use the average value and the corresponding depth measurement value to perform linear fitting, and the resulting slope is the attenuation coefficient characterizing the degree of hepatic steatosis.

4. A device for quantitative detection of hepatic steatosis, characterized in that, A method for quantitative detection of hepatic steatosis as described in any one of claims 1 to 3, comprising: An ultrasonic radio frequency signal acquisition module is used to acquire at least one set of ultrasonic radio frequency signals; The image reconstruction module is used to reconstruct the ultrasonic radio frequency signal into a fan-shaped B-mode grayscale image through digital scanning conversion positive transformation, and to pre-calculate and store the mapping index table of each pixel point to the ultrasonic radio frequency signal data domain. The deep learning segmentation module is used to segment the liver region of the fan-shaped B-mode grayscale image using a pre-trained deep learning segmentation model to obtain a fan-shaped binary mask. The mask inverse mapping module is used to perform digital scan conversion inverse transformation on the sector binary mask based on the mapping index table to obtain a radio frequency domain mask matrix of the same size as the ultrasonic radio frequency signal; The power spectrum transformation module is used to filter the effective depth interval according to the radio frequency domain mask matrix and perform spectrum transformation to obtain the power spectrum energy matrix. The attenuation coefficient calculation module is used to perform frequency normalization and linear fitting on the power spectrum energy matrix to obtain the attenuation coefficient characterizing the degree of hepatic steatosis.

5. A quantitative detection system for hepatic steatosis, characterized in that, It includes an ultrasound probe and the quantitative detection device for hepatic steatosis as described in claim 4.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the method for quantitative detection of hepatic steatosis as described in any one of claims 1 to 3.

7. 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 steps of the method for quantitative detection of hepatic steatosis as described in any one of claims 1 to 3.

Images