A deep learning-based thyroid nodule ultrasound image segmentation method

Abstract

The application relates to the technical field of image processing, and discloses a thyroid nodule ultrasonic image segmentation method based on deep learning, which comprises the following steps: performing analytical signal decomposition on an ultrasonic echo radio frequency matrix and an orthogonal transformation matrix to extract an instantaneous amplitude matrix and a phase matrix; generating a spatial phase gradient matrix by extracting the partial derivative of the instantaneous phase matrix along a spatial coordinate system; performing translation normalization to construct a local anisotropy matrix in combination with a spatial window size constant; determining a Hessian matrix determinant and a local Hessian mean matrix based on the instantaneous amplitude matrix, deriving a singularity attenuation matrix, and modulating to generate a regularized anisotropy matrix; extracting a local gradient variance matrix according to the instantaneous amplitude matrix and comprehensively establishing a capsule interface probability matrix; then, a capsule maintaining projection matrix is fused and generated; and finally, the instantaneous amplitude matrix and the capsule maintaining projection matrix are subjected to numerical modulation, and an enhanced image feature matrix is reconstructed and output.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a deep learning-based method for segmenting ultrasound images of thyroid nodules. Background Technology

[0002] Deep learning-based ultrasound image segmentation techniques are commonly used for the auxiliary diagnosis of thyroid nodules. Existing ultrasound image segmentation networks often use the amplitude features of ultrasound B-mode as direct input. However, during ultrasound detection, due to the small difference in acoustic impedance between thyroid nodules and surrounding normal glandular tissue, conventional methods for extracting boundaries based on amplitude features are easily affected by acoustic impedance gradient spoofing. This spoofing increases the difficulty for the segmentation network to extract the objective physical boundaries. To overcome the influence of amplitude attenuation, some existing techniques introduce spatial phase features. However, in actual scenarios, there are numerous interwoven fiber bundles within thyroid tissue. These interwoven fiber bundles lead to anisotropic fiber distortion. This causes disordered refraction distortion of the simple phase gradient at the fiber interface, affecting the continuity of local feature structures. To reduce the error caused by this directional refraction, local features usually require smoothing processes such as spatial normalization. Furthermore, microcalcifications are commonly found in thyroid nodules. These calcifications produce local high-frequency point reflections. During spatial normalization, high-frequency reflections can cause singularity perturbations in microcalcifications, thereby disrupting the continuity of the feature matrix and generating false boundaries. When indiscriminate geometric regularization is used to eliminate singularity perturbations, it often leads to excessive smoothing of the capsule interface. Indiscriminate smoothing causes the loss of true high-order gradient features of the thyroid capsule, ultimately resulting in underfitting of the segmentation network at the nodule capsule.

[0003] In summary, a deep learning-based method for thyroid nodule ultrasound image segmentation is needed to overcome defects such as acoustic impedance masquerading, fiber distortion, calcification perturbation, and excessive capsule smoothing in the input feature construction process. Summary of the Invention

[0004] This invention provides a deep learning-based method for segmenting thyroid nodules from ultrasound images, which helps to solve the problems mentioned in the background section.

[0005] This invention provides the following technical solution: a deep learning-based method for segmenting thyroid nodules from ultrasound images, comprising:

[0006] Analytical signal decomposition is performed on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal;

[0007] The spatial phase gradient matrix is ​​generated by extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system.

[0008] By combining the spatial window size constant with the spatial phase gradient matrix, a spatial translation normalization is performed to construct a local anisotropic matrix;

[0009] The determinant of the Hessian matrix is ​​determined based on the second-order differential characteristics of the instantaneous amplitude matrix, and the local Hessian mean matrix is ​​summarized within the range of the constant size of the spatial window.

[0010] The singularity decay matrix is ​​derived by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and then the local anisotropy matrix is ​​modulated by the singularity decay matrix to generate a regularized anisotropy matrix.

[0011] The amplitude gradient matrix is ​​determined based on the first-order differential characteristics of the instantaneous amplitude matrix, and the local gradient variance matrix is ​​extracted in conjunction with the spatial window size constant.

[0012] The envelope interface probability matrix is ​​established by combining the magnitude gradient matrix and the local gradient variance matrix, and then the envelope retention projection matrix is ​​generated by fusing the local anisotropy matrix and the regularized anisotropy matrix.

[0013] The instantaneous amplitude matrix and the envelope-preserving projection matrix are numerically modulated to generate a phase-compensated boundary feature matrix, which is then reconstructed with the instantaneous amplitude matrix to output an enhanced image feature matrix.

[0014] The enhanced image feature matrix is ​​input into a deep learning segmentation network, which outputs a thyroid nodule segmentation mask.

[0015] Optionally, the step of performing analytical signal decomposition on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal includes:

[0016] Set up and start the ultrasound linear array probe to transmit continuous ultrasound radiofrequency pulses with a determined center frequency to the thyroid target area;

[0017] The system hardware receiving channel is activated to prepare for receiving ultrasonic echo signals. A spatial coordinate system is established with the horizontal physical arrangement direction of the ultrasonic probe array elements as the horizontal axis and the vertical physical penetration depth direction of ultrasonic wave propagation as the vertical axis.

[0018] Obtain the two-dimensional ultrasonic echo radio frequency matrix obtained by the ultrasonic probe acquisition and conversion, and obtain the orthogonal transformation matrix generated by the system after performing a discrete Hilbert transformation on the radio frequency matrix along the depth physical direction;

[0019] The square of the two-dimensional ultrasonic echo radio frequency matrix is ​​added to the square of the orthogonal transformation matrix, and the square root operation is performed on the result to generate the instantaneous amplitude matrix.

[0020] The ratio is obtained by dividing the orthogonal transformation matrix by the two-dimensional ultrasonic echo radio frequency matrix, and the instantaneous phase matrix is ​​generated by performing an arctangent operation on the ratio.

[0021] Optionally, the step of extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system to generate the spatial phase gradient matrix includes:

[0022] Calculate the first-order partial derivatives of the instantaneous phase matrix along the horizontal and vertical axes of the spatial coordinate system, respectively.

[0023] Squaring the first-order partial derivatives along the horizontal axis and the vertical axis respectively;

[0024] The results of squaring the two first-order partial derivatives are added together, and the square root of the sum is taken to generate the spatial phase gradient matrix.

[0025] Optionally, the step of performing spatial translation normalization on the spatial phase gradient matrix in conjunction with the spatial window size constant to construct a local anisotropy matrix includes:

[0026] Set the space window size constant;

[0027] Within a local spatial window determined by the spatial window size constant, a spatial translation summation operation is performed on the spatial phase gradient matrix. The result of the spatial translation summation operation is divided by the square of the spatial window size constant to obtain the mean value of the local translation gradient.

[0028] The spatial phase gradient matrix is ​​added to the mean of the local translation gradient to obtain the denominator term, and the spatial phase gradient matrix is ​​divided by the denominator term to generate the local anisotropy matrix.

[0029] Optionally, the step of determining the determinant of the Hessian matrix based on the second-order differential characteristic of the instantaneous amplitude matrix, and summarizing the local Hessian mean matrix within the range of the spatial window size constant, includes:

[0030] Calculate the second-order pure partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively, and calculate the second-order mixed partial derivatives of the instantaneous amplitude matrix;

[0031] Multiply the second-order pure partial derivatives in the horizontal direction with the second-order pure partial derivatives in the vertical direction to obtain the first product term. Squaring the second-order mixed partial derivatives to obtain the second squared term. Subtracting the second squared term from the first product term generates the determinant of the Hessian matrix.

[0032] Within a local spatial window determined by the spatial window size constant, the determinant of the Hessian matrix is ​​subjected to spatial translation and summation. The result of the summation of the Hessian matrix determinant is divided by the square of the spatial window size constant to generate a local Hessian mean matrix.

[0033] Optionally, the step of deriving the singularity decay matrix by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and using it to modulate the local anisotropy matrix to generate a regularized anisotropy matrix, includes:

[0034] Divide the determinant of the Hessian matrix by the local Hessian mean matrix to obtain the median ratio, take the negative value of the median ratio, and perform an exponential operation with the base of the natural logarithm as the base and the median ratio after taking the negative value as the exponent to generate a singular decay matrix.

[0035] The local anisotropy matrix is ​​multiplied by the singularity decay matrix to generate a regularized anisotropy matrix.

[0036] Optionally, the step of determining the amplitude gradient matrix based on the first-order differential feature of the instantaneous amplitude matrix and extracting the local gradient variance matrix in conjunction with the spatial window size constant includes:

[0037] Calculate the first-order partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Square the first-order partial derivatives along the horizontal and vertical axes, respectively. Add the results of squaring the two first-order partial derivatives, and take the square root of the sum to generate the amplitude gradient matrix.

[0038] Within a local spatial window determined by the spatial window size constant, the squared values ​​of the magnitude gradient matrix are spatially translated and summed. The result of the spatial translation and summation of the squared magnitude gradient values ​​is divided by the square of the spatial window size constant to obtain the mean of the squared terms.

[0039] Within the same local spatial window, the magnitude gradient matrix is ​​subjected to spatial translation and summation. The result of spatial translation and summation of the magnitude gradient is divided by the square of the spatial window size constant to obtain the gradient mean. The gradient mean is squared to obtain the mean square term. The mean of the square term is subtracted from the mean square term to generate the local gradient variance matrix.

[0040] Optionally, the step of establishing the envelope interface probability matrix by integrating the magnitude gradient matrix and the local gradient variance matrix, and then fusing the local anisotropy matrix and the regularized anisotropy matrix to generate the envelope preservation projection matrix, includes:

[0041] Set a minimum constant;

[0042] The local gradient variance matrix is ​​added to the minimum constant to obtain the variance correction term. The magnitude gradient matrix is ​​squared to obtain the magnitude gradient squared term. The magnitude gradient squared term is divided by the variance correction term to obtain the correction ratio. The correction ratio is negative, and the exponential decay term is obtained by exponential operation with the base of the natural logarithm as the base and the negative correction ratio as the exponent. The membrane interface probability matrix is ​​generated by subtracting the exponential decay term from the value 1.

[0043] Multiplying the membrane interface probability matrix with the local anisotropy matrix yields the original feature retention term. Subtracting the membrane interface probability matrix from the value 1 yields the smoothing probability term. Multiplying the smoothing probability term with the regularized anisotropy matrix yields the regularized feature retention term. Adding the original feature retention term and the regularized feature retention term generates the membrane preservation projection matrix.

[0044] Optionally, the step of numerically modulating the instantaneous amplitude matrix and the envelope-preserving projection matrix to generate a phase-compensated boundary feature matrix, and reconstructing it with the instantaneous amplitude matrix to output an enhanced image feature matrix, includes:

[0045] The phase-compensated boundary feature matrix is ​​generated by multiplying the membrane-preserving projection matrix with the instantaneous amplitude matrix.

[0046] Multiply the instantaneous amplitude matrix with the phase compensation boundary feature matrix to obtain the numerator fusion term, square the instantaneous amplitude matrix to obtain the amplitude square term, square the phase compensation boundary feature matrix to obtain the feature square term, add the amplitude square term and the feature square term, take the square root of the squared result to obtain the denominator normalization term, and divide the numerator fusion term by the denominator normalization term to generate the enhanced image feature matrix;

[0047] The enhanced image feature matrix is ​​input into a trained deep learning segmentation network, which then outputs a thyroid nodule segmentation mask.

[0048] The present invention has the following beneficial effects:

[0049] 1. By performing analytical signal decomposition on the ultrasound echo radio frequency matrix and orthogonal transformation matrix, this scheme extracts the instantaneous amplitude matrix and instantaneous phase matrix. Combined with features such as the spatial phase gradient matrix, local anisotropy matrix, Hessian matrix determinant, and local gradient variance matrix, a capsule-preserving projection matrix is ​​constructed. Finally, this matrix is ​​reconstructed with the instantaneous amplitude matrix to output the enhanced image feature matrix. In the specific environment of thyroid nodule ultrasound image processing, the acoustic impedance difference between nodules and normal tissue is small, and there are interwoven fiber bundles and microcalcifications within the tissue. Conventional methods are easily affected by fiber refraction distortion and calcification reflection disturbances during feature extraction, and cause excessive smoothing of the actual capsule interface during smoothing and noise reduction. This scheme utilizes a local anisotropy matrix to reduce numerical errors caused by fiber refraction and a singularity attenuation matrix to suppress feature continuity disruption caused by calcifications. Simultaneously, it establishes the capsule interface probability matrix by combining the amplitude gradient matrix and the local gradient variance matrix, preserving capsule features when performing regularization on the local anisotropy matrix. This provides deep learning networks with input features containing structural boundary information and reduces the interference of complex glandular tissue environment on ultrasound feature extraction.

[0050] 2. By setting and activating the ultrasound linear array probe to emit continuous ultrasound radio frequency pulses towards the thyroid target area, a spatial coordinate system is established with the transverse physical arrangement direction of the probe array elements as the horizontal axis and the longitudinal physical penetration depth direction of the ultrasound propagation as the vertical axis. Then, a two-dimensional ultrasound echo radio frequency matrix and an orthogonal transformation matrix generated by performing discrete Hilbert transformation along the depth physical direction are obtained. Based on this, the square root operation is performed on the sum of the squares of the radio frequency matrix and the orthogonal transformation matrix to generate an instantaneous amplitude matrix, and the arctangent operation is performed on the ratio of the two matrices to generate an instantaneous phase matrix. This step decouples the original low-order high-frequency ultrasound echo radio frequency signal into amplitude characteristics reflecting macroscopic acoustic impedance differences and phase characteristics reflecting microscopic tissue continuity, realizing the physical separation of the envelope signal and phase fluctuations. This provides a spatial data source with basic analytical characteristics for subsequent extraction of morphological boundaries that are not affected by local acoustic attenuation.

[0051] 3. The first-order partial derivatives of the instantaneous phase matrix are calculated along the horizontal and vertical axes of the established spatial coordinate system, and the first-order partial derivatives in both directions are squared. After adding the results of the squared first-order partial derivatives, the square root operation is performed to generate the spatial phase gradient matrix. This step uses the spatial directional derivative of phase information to characterize the physical microscopic changes of the tissue interface. Since the phase information does not undergo baseband drift with the exponential decay of acoustic wave energy in deep tissue, it can effectively capture the weak structural misalignments that cannot be represented by conventional amplitude images, suppress the acoustic impedance gradient camouflage caused by the extremely small difference in acoustic impedance between the nodule boundary and the surrounding normal glandular tissue, and improve the contrast of morphological boundary structural features at the level of spatial change rate.

[0052] 4. By setting a spatial window size constant, and performing a spatial translation summation operation on the spatial phase gradient matrix within the local spatial window determined by this constant, the summation result is divided by the square of the spatial window size constant to obtain the mean of the local translation gradient. Then, the spatial phase gradient matrix is ​​added to this mean of the local translation gradient to obtain the denominator. Finally, the spatial phase gradient matrix is ​​divided by the denominator to generate the local anisotropy matrix. This step addresses the anisotropic fiber distortion problem caused by a large number of interlaced fiber bundles inside the thyroid gland. Adaptive spatial normalization constraints are performed in the local spatial region to transform the gradient features with strong directionality into dimensionless structural coherence indices. This eliminates the numerical variation caused by the directional refraction of sound waves at the interface of non-uniform media and maintains the local continuity of the characteristic structure inside the tissue.

[0053] 5. By calculating the second-order pure partial derivatives and the second-order mixed partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, the first product term is obtained by multiplying the second-order pure partial derivatives, and the second square term of the second-order mixed partial derivatives is subtracted to generate the determinant of the Hessian matrix. Then, the determinant of the Hessian matrix is ​​spatially translated and summed within a local spatial window determined by the constant size of the spatial window, and divided by the square of the constant to generate the local Hessian mean matrix. This step utilizes the second-order differential property of the amplitude image to quantify the local curvature characteristics of the spatial signal. Since the microcalcifications exhibit extremely strong point-like physical scattering in the ultrasonic field, their curvature envelope will produce drastic abrupt changes in the axial and lateral directions. Therefore, this step can effectively separate the point-like high-frequency and high-brightness reflection areas from the low-frequency solid background, providing a spatial comparison benchmark for subsequent evaluation of local singularity perturbations.

[0054] 6. By dividing the determinant of the Hessian matrix by the local Hessian mean matrix to obtain an intermediate ratio, the intermediate ratio is negative and exponentially operated on with the base of the natural logarithm to generate a singularity decay matrix. Finally, the local anisotropy matrix is ​​multiplied by this singularity decay matrix to generate a regularized anisotropy matrix. This step repairs the local field strength distortion caused by high-frequency calcification points. Using a nonlinear suppression mask constructed from spatial curvature contrast, the structural feature values ​​of the region where the strong scattering source is located are adaptively shrunk through a negative exponential mapping relationship, while maintaining high transmittance in the spatially flat region. This effectively performs local strong smoothing regularization on the numerical abrupt changes in the local anisotropy matrix caused by calcification points, eliminates false boundaries caused by small calcification singularity perturbations, and blocks the erroneous propagation of speckle noise in the spatial structure.

[0055] 7. An amplitude gradient matrix is ​​generated by calculating the first-order partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, summing their squares, and taking the square root. Then, within a local window determined by the constant size of the spatial window, the squared value and the original value of the gradient matrix are spatially translated and summed, and divided by the square of the constant to obtain the mean of the squared terms and the mean of the gradient. Finally, the mean of the squared terms is subtracted from the square of the mean of the gradient to generate the local gradient variance matrix. This step uses the statistical dispersion of the first-order derivative to quantify the spatial fluctuation characteristics of the boundary signal. Since the real thyroid capsule interface exhibits continuous and dramatic amplitude steps in the ultrasound echo, its local variance will be significantly higher than that of disordered fibrous tissue or flat parenchyma. Therefore, this step achieves a spatial measurement of the deterministic nature of the real boundary, providing a high-order data foundation for the accurate separation of fibers and capsule.

[0056] 8. A variance correction term is obtained by adding a minimum constant to the local gradient variance matrix. The squared magnitude gradient term is divided by the variance correction term and the negative value is taken. After natural logarithmic exponentiation, the term is subtracted by 1 to generate the capsule interface probability matrix. Then, the probability matrix is ​​cross-multiplied with the local anisotropy matrix and the smoothed probability term (after subtracting the probability matrix from 1) and the regularized anisotropy matrix. The original feature retention term and the regularized feature retention term are added to generate the capsule-preserving projection matrix. This step establishes a high-order adaptive spatial projection mechanism. In regions where the probability of capsule formation is high, unsmoothed anisotropic features are forcibly retained. In calcification noise regions, regularized smoothed features are introduced. This overcomes the technical defect of excessive smoothing of the capsule interface caused by conventional geometric smoothing and prevents underfitting of the segmentation network at the nodule capsule boundary.

[0057] 9. A phase-compensated boundary feature matrix is ​​generated by multiplying the envelope-preserving projection matrix with the instantaneous amplitude matrix. The numerator fusion term is obtained by multiplying the instantaneous amplitude matrix with this boundary feature matrix. Simultaneously, the two matrices are squared, added, and square-rooted to generate a denominator normalization term. Finally, the numerator fusion term is divided by the denominator normalization term to generate an enhanced image feature matrix. This step reverses the dimensionless coherent structural features after multidimensional correction and physically modulates them back into the amplitude image with the physical dimensions of acoustic energy. The effective reconstruction of physical properties is achieved through nonlinear square root mapping, ensuring that the output tensor retains high-fidelity high-order phase boundary information while having an energy dynamic range isomorphic to the original ultrasound image. This eliminates the dimensional conflict of the analytical multi-channel input, enabling it to directly match the input of a multi-layer convolutional neural network. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the basic process of the present invention.

[0059] Figure 2 This is a schematic diagram illustrating the analytical signal decomposition and spatial phase gradient construction of this invention.

[0060] Figure 3 This is a schematic diagram illustrating the local anisotropy regularization and singularity suppression of the present invention.

[0061] Figure 4 This is a schematic diagram of the coating retention projection and enhanced feature reconstruction of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1, refer to Figure 1 A deep learning-based method for segmenting thyroid nodules from ultrasound images, comprising:

[0064] Analytical signal decomposition is performed on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal;

[0065] The spatial phase gradient matrix is ​​generated by extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system.

[0066] By combining the spatial window size constant with the spatial phase gradient matrix, a spatial translation normalization is performed to construct a local anisotropic matrix;

[0067] The determinant of the Hessian matrix is ​​determined based on the second-order differential characteristics of the instantaneous amplitude matrix, and the local Hessian mean matrix is ​​summarized within the range of the constant size of the spatial window.

[0068] The singularity decay matrix is ​​derived by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and then the local anisotropy matrix is ​​modulated by the singularity decay matrix to generate a regularized anisotropy matrix.

[0069] The amplitude gradient matrix is ​​determined based on the first-order differential characteristics of the instantaneous amplitude matrix, and the local gradient variance matrix is ​​extracted in conjunction with the spatial window size constant.

[0070] The envelope interface probability matrix is ​​established by combining the magnitude gradient matrix and the local gradient variance matrix, and then the envelope retention projection matrix is ​​generated by fusing the local anisotropy matrix and the regularized anisotropy matrix.

[0071] The instantaneous amplitude matrix and the envelope-preserving projection matrix are numerically modulated to generate a phase-compensated boundary feature matrix, which is then reconstructed with the instantaneous amplitude matrix to output an enhanced image feature matrix.

[0072] The enhanced image feature matrix is ​​input into a deep learning segmentation network, which outputs a thyroid nodule segmentation mask.

[0073] Reference Figure 2 The figure shows the correspondence between ultrasound probe acquisition, radio frequency matrix analytical decomposition, and spatial phase gradient generation.

[0074] The step of performing analytical signal decomposition on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal includes:

[0075] The purpose of this step is to use analytical signal theory to decouple the original probe signal into amplitude characteristics reflecting macroscopic acoustic impedance differences and phase characteristics reflecting microscopic organizational continuity, thereby providing basic data support for overcoming acoustic impedance gradient camouflage in the future.

[0076] Set up and start the ultrasound linear array probe to transmit continuous ultrasound radiofrequency pulses with a determined center frequency to the thyroid target area;

[0077] The system hardware receiving channel is activated to prepare for receiving ultrasonic echo signals. A spatial coordinate system is established with the horizontal physical arrangement direction of the ultrasonic probe array elements as the horizontal axis and the vertical physical penetration depth direction of ultrasonic wave propagation as the vertical axis.

[0078] Obtain the two-dimensional ultrasonic echo radio frequency matrix obtained from the ultrasonic probe acquisition and conversion;

[0079] Obtain the orthogonal transformation matrix generated by the system after performing a discrete Hilbert transformation on the radio frequency matrix along the depth physical direction;

[0080] To quantify the pure energy envelope of sound wave transmission without phase fluctuations, based on the aforementioned data, the instantaneous amplitude matrix is ​​first calculated. The calculation steps are as follows:

[0081] The square of the two-dimensional ultrasonic echo radio frequency matrix is ​​added to the square of the orthogonal transformation matrix, and the square root operation is performed on the result to generate the instantaneous amplitude matrix.

[0082] Next, in order to obtain pure structural phase information independent of local signal attenuation, the instantaneous phase matrix characterizing the offset of the acoustic physical interface is calculated using the same known acquired data. The calculation steps are as follows:

[0083] The ratio is obtained by dividing the orthogonal transformation matrix by the two-dimensional ultrasonic echo radio frequency matrix, and the instantaneous phase matrix is ​​generated by performing an arctangent operation on the ratio.

[0084] By setting and activating an ultrasound linear array probe to emit continuous ultrasound radio frequency pulses toward the thyroid target area, and establishing a spatial coordinate system with the transverse physical arrangement direction of the probe array elements as the horizontal axis and the longitudinal physical penetration depth direction of the ultrasound propagation as the vertical axis, a two-dimensional ultrasound echo radio frequency matrix and an orthogonal transformation matrix generated by performing a discrete Hilbert transform along the depth physical direction are obtained. Based on this, the square root operation is performed on the sum of the squares of the radio frequency matrix and the orthogonal transformation matrix to generate an instantaneous amplitude matrix, and the arctangent operation is performed on the ratio of the two matrices to generate an instantaneous phase matrix. This step decouples the original low-order high-frequency ultrasound echo radio frequency signal into amplitude characteristics reflecting macroscopic acoustic impedance differences and phase characteristics reflecting microscopic tissue continuity, realizing the physical separation of the envelope signal and phase fluctuations, and providing a spatial data source with basic analytical characteristics for subsequent extraction of morphological boundaries that are not affected by local acoustic attenuation.

[0085] The step of extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system to generate the spatial phase gradient matrix includes:

[0086] The purpose of this step is to calculate the spatial directional derivative of the phase along the established spatial coordinate system, obtain morphological boundary features that are completely unaffected by local acoustic attenuation of the tissue, and thus preliminarily solve the basic problem of acoustic impedance gradient camouflage.

[0087] To extract the spatial rate of change of this phase information along the spatial coordinate system and thus accurately locate the physical boundary of the tissue, the spatial phase gradient matrix is ​​directly calculated. The calculation steps are as follows:

[0088] Calculate the first-order partial derivatives of the instantaneous phase matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Square the first-order partial derivatives along the horizontal and vertical axes, respectively. Add the results of the squaring of the two first-order partial derivatives, and perform a square root operation on the sum to generate the spatial phase gradient matrix.

[0089] The spatial phase gradient matrix is ​​generated by calculating the first-order partial derivatives of the instantaneous phase matrix along the horizontal and vertical axes of the established spatial coordinate system, squaring the first-order partial derivatives in both directions, adding the results of the squaring of the two first-order partial derivatives, and then performing a square root operation. This step uses the spatial directional derivatives of phase information to characterize the physical microscopic changes of the tissue interface. Since the phase information does not undergo baseband drift with the exponential decay of acoustic wave energy in deep tissues, it can effectively capture subtle structural misalignments that cannot be represented by conventional amplitude images, suppress the acoustic impedance gradient camouflage caused by the minimal difference in acoustic impedance between nodule boundaries and surrounding normal glandular tissue, and improve the contrast of morphological boundary structural features at the level of spatial change rate.

[0090] The step of performing spatial translation normalization on the spatial phase gradient matrix by combining the spatial window size constant to construct a local anisotropy matrix includes:

[0091] The purpose of this step is to address the anisotropic distortion problem caused by the complex thyroid fibrous structure by performing adaptive spatial normalization calculations on the phase gradient within a local spatial window, thereby eliminating numerical errors caused by directional refraction.

[0092] Set a constant for the spatial window size, the value of which is based on the number of physical pixels corresponding to the wavelength of the ultrasonic probe's emission center frequency, to ensure that it can completely cover one acoustic scattering cycle; if the value is too large, it will cause the features to be over-smoothed and lose true details; if the value is too small, it will not be able to effectively filter out local speckle noise.

[0093] Within a local spatial window determined by the spatial window size constant, a spatial translation summation operation is performed on the spatial phase gradient matrix. The result of the spatial translation summation operation is divided by the square of the spatial window size constant to obtain the mean of the local translation gradient. The spatial phase gradient matrix is ​​added to the mean of the local translation gradient to obtain the denominator. The spatial phase gradient matrix is ​​divided by the denominator to generate the local anisotropy matrix.

[0094] By setting a constant for the spatial window size and performing a spatial translation summation operation on the spatial phase gradient matrix within a local spatial window determined by this constant, the summation result is divided by the square of the spatial window size constant to obtain the mean of the local translation gradient. Then, the spatial phase gradient matrix is ​​added to this mean of the local translation gradient to obtain the denominator term. Finally, the spatial phase gradient matrix is ​​divided by the denominator term to generate the local anisotropy matrix. This step addresses the anisotropic fiber distortion problem caused by a large number of interlaced fiber bundles inside the thyroid gland. It performs adaptive spatial normalization constraints in the local spatial region, transforming the gradient features with strong directionality into dimensionless structural coherence indices. This eliminates the numerical anomalies caused by the directional refraction of sound waves at the interface of a non-homogeneous medium and maintains the local continuity of the characteristic structures inside the tissue.

[0095] Reference Figure 3 The figure shows the construction of the local anisotropic matrix, the determination of the Hessian matrix determinant, and the singularity attenuation modulation relationship.

[0096] The process of determining the determinant of the Hessian matrix based on the second-order differential characteristic of the instantaneous amplitude matrix, and summarizing the local Hessian mean matrix within the constant range of the spatial window size, includes:

[0097] This step aims to utilize the second-order differential property of amplitude to accurately identify the tiny calcification points that cause severe disturbances. Their extremely strong point-like physical scattering will cause local curvature extrema, providing a digital identification basis for subsequent solutions to singularity perturbation problems.

[0098] To accurately locate high-frequency point reflection sources using curvature extremum features, the determinant of the Hessian matrix is ​​first calculated. The calculation steps are as follows:

[0099] Calculate the second-order pure partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Calculate the second-order mixed partial derivatives of the instantaneous amplitude matrix. Multiply the second-order pure partial derivatives along the horizontal and vertical axes to obtain the first product term. Square the second-order mixed partial derivatives to obtain the second squared term. Subtract the second squared term from the first product term to generate the Hessian matrix determinant.

[0100] Subsequently, in order to provide an adaptive spatial comparison benchmark threshold for the judgment of singularity perturbations, a local Hessian mean matrix is ​​calculated as a dynamic benchmark based on a constant space. The calculation steps are as follows:

[0101] Within a local spatial window determined by the spatial window size constant, the determinant of the Hessian matrix is ​​subjected to spatial translation and summation. The result of the summation of the Hessian matrix determinant is divided by the square of the spatial window size constant to generate a local Hessian mean matrix.

[0102] By calculating the second-order pure partial derivatives and the second-order mixed partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, the first product term is obtained by multiplying the second-order pure partial derivatives, and the second square term of the second-order mixed partial derivatives is subtracted to generate the Hessian matrix determinant. Then, the Hessian matrix determinant is spatially translated and summed within a local spatial window determined by the spatial window size constant, and divided by the square of the constant to generate the local Hessian mean matrix. This step utilizes the second-order differential property of the amplitude image to quantify the local curvature characteristics of the spatial signal. Since microcalcifications exhibit extremely strong point-like physical scattering in the ultrasonic field, their curvature envelope will produce drastic abrupt changes in the axial and lateral directions. Therefore, this step can effectively separate the point-like high-frequency, high-brightness reflection area from the low-frequency solid background, providing a spatial comparison benchmark for subsequent evaluation of local singularity perturbations.

[0103] The process of deriving the singularity decay matrix by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and then using it to modulate the local anisotropy matrix to generate a regularized anisotropy matrix, includes:

[0104] The purpose of this step is to use the identified calcification singularity benchmarks to generate a high-precision suppression mask matrix and to perform forced smoothing regularization on the anisotropic matrix containing noise perturbations, thereby completely eliminating the false boundary problem caused by micro-calcifications.

[0105] To generate a weighting coefficient between 0 and 1 that has the reverse singularity suppression physical property, the dimensionless singularity attenuation matrix is ​​first calculated. The calculation steps are as follows:

[0106] Divide the determinant of the Hessian matrix by the local Hessian mean matrix to obtain the median ratio, take the negative value of the median ratio, and perform an exponential operation with the base of the natural logarithm as the base and the median ratio after taking the negative value as the exponent to generate a singular decay matrix.

[0107] Next, in order to apply the suppression weights effectively to the original spatial structure and completely filter out calcification interference, the regularization anisotropy matrix is ​​calculated using the aforementioned attenuation matrix. The calculation steps are as follows:

[0108] The local anisotropy matrix is ​​multiplied by the singularity decay matrix to generate a regularized anisotropy matrix.

[0109] By dividing the determinant of the Hessian matrix by the local Hessian mean matrix to obtain an intermediate ratio, taking the negative value of this intermediate ratio and performing an exponential operation with the base of the natural logarithm to generate a singularity decay matrix, and finally multiplying the local anisotropy matrix with this singularity decay matrix to generate a regularized anisotropy matrix, this step repairs the local field strength distortion caused by high-frequency calcification points. Using a nonlinear suppression mask constructed from spatial curvature contrast, the structural feature values ​​of the region where the strong scattering source is located are adaptively shrunk through a negative exponential mapping relationship, while maintaining high transmittance in the spatially flat region. This effectively performs local strong smoothing regularization on the numerical abrupt changes in the local anisotropy matrix caused by calcification points, eliminates false boundaries caused by small calcification singularity perturbations, and blocks the erroneous propagation of speckle noise in the spatial structure.

[0110] Reference Figure 4 The figure shows the magnitude gradient, local gradient variance, envelope-preserving projection matrix, and the relationship between the enhanced features and the reconstruction.

[0111] The step of determining the amplitude gradient matrix based on the first-order differential feature of the instantaneous amplitude matrix and extracting the local gradient variance matrix in conjunction with the spatial window size constant includes:

[0112] The purpose of this step is to prevent the regularization process from indiscriminately smoothing out the real thyroid capsule. By using the local variance characteristics of the amplitude gradient as an independent discriminant, the continuous and drastic fluctuation characteristics of the physical interface are dynamically separated and evaluated.

[0113] To accurately capture drastic changes in the amplitude of local tissue echo signals for preliminary screening of the true capsule, the amplitude gradient matrix is ​​first calculated. The calculation steps are as follows:

[0114] Calculate the first-order partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Square the first-order partial derivatives along the horizontal and vertical axes, respectively. Add the results of squaring the two first-order partial derivatives, and take the square root of the sum to generate the amplitude gradient matrix.

[0115] Next, in order to quantify the spatial dispersion of this magnitude gradient to further confirm structural continuity, the local gradient variance matrix is ​​calculated in conjunction with the pre-defined spatial window size constant. The calculation steps are as follows:

[0116] Within a local spatial window determined by the spatial window size constant, the squared values ​​of the magnitude gradient matrix are spatially translated and summed. The result of the spatial translation and summation of the squared magnitude gradient values ​​is divided by the square of the spatial window size constant to obtain the mean of the squared terms. Within the same local spatial window, the magnitude gradient matrix is ​​spatially translated and summed. The result of the spatial translation and summation of the magnitude gradient is divided by the square of the spatial window size constant to obtain the mean gradient. The mean gradient is squared to obtain the mean squared term. The mean squared term is subtracted from the mean squared term to generate the local gradient variance matrix.

[0117] The amplitude gradient matrix is ​​generated by calculating the first-order partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, summing their squares, and taking the square root. Then, within a local window determined by the constant size of the spatial window, the squared value and the original value of the gradient matrix are spatially translated and summed, and then divided by the square of the constant to obtain the mean of the squared terms and the mean of the gradient. Finally, the mean of the squared terms is subtracted from the square of the mean of the gradient to generate the local gradient variance matrix. This step uses the statistical dispersion of the first-order derivative to quantify the spatial fluctuation characteristics of the boundary signal. Since the real thyroid capsule interface exhibits continuous and dramatic amplitude steps in the ultrasound echo, its local variance will be significantly higher than that of disordered fibrous tissue or flat parenchyma. Therefore, this step achieves a spatial measurement of the deterministic nature of the real boundary, providing a high-order data foundation for the accurate separation of fibers and capsule.

[0118] The process of establishing the envelope interface probability matrix by combining the magnitude gradient matrix and the local gradient variance matrix, and then fusing the local anisotropy matrix and the regularized anisotropy matrix to generate the envelope preservation projection matrix, includes:

[0119] This step aims to combine the regularization decay results with the capsule variance response results to establish a high-order adaptive projection mapping, which preserves the original features in the capsule region and uses smoothed features in the interference calcification region, thus solving the segmentation underfitting caused by excessive smoothing.

[0120] Set a very small constant whose value is based on the lower limit of machine precision of the current hardware computing architecture. The purpose of setting it is to prevent overflow and crash exceptions where the denominator is zero during subsequent mathematical division operations.

[0121] To dynamically assess the reliability of the current spatial location belonging to the true envelope based on the gradient and its variance index, the envelope interface probability matrix is ​​first calculated based on a set constant and prior data. The calculation steps are as follows:

[0122] The local gradient variance matrix is ​​added to the minimum constant to obtain the variance correction term. The magnitude gradient matrix is ​​squared to obtain the magnitude gradient squared term. The magnitude gradient squared term is divided by the variance correction term to obtain the correction ratio. The correction ratio is negative, and the exponential decay term is obtained by exponential operation with the base of the natural logarithm as the base and the negative correction ratio as the exponent. The membrane interface probability matrix is ​​generated by subtracting the exponential decay term from the value 1.

[0123] Subsequently, in order to achieve a physically smooth transition between preserving minute capsule details and strongly suppressing calcification noise, a capsule preservation projection matrix is ​​constructed based on the above probability matrix, and its calculation steps are as follows:

[0124] Multiplying the membrane interface probability matrix with the local anisotropy matrix yields the original feature retention term. Subtracting the membrane interface probability matrix from the value 1 yields the smoothing probability term. Multiplying the smoothing probability term with the regularized anisotropy matrix yields the regularized feature retention term. Adding the original feature retention term and the regularized feature retention term generates the membrane preservation projection matrix.

[0125] The variance correction term is obtained by adding a minimum constant to the local gradient variance matrix. The squared magnitude gradient term is divided by the variance correction term and the negative value is taken. After the natural logarithm exponential operation, the term is subtracted by 1 to generate the capsule interface probability matrix. Then, the probability matrix is ​​cross-multiplied with the local anisotropy matrix and the smoothed probability term (after subtracting the probability matrix by 1) and the regularized anisotropy matrix. The original feature retention term and the regularized feature retention term are added to generate the capsule-preserving projection matrix. This step establishes a high-order adaptive spatial projection mechanism. In regions where the probability of capsule formation is high, unsmoothed anisotropic features are forcibly retained. In calcification noise regions, regularized smoothed features are introduced. This overcomes the technical defect of excessive smoothing of the capsule interface caused by conventional geometric smoothing and prevents underfitting of the segmentation network at the nodule capsule boundary.

[0126] The step of numerically modulating the instantaneous amplitude matrix and the capsule-preserving projection matrix to generate a phase-compensated boundary feature matrix, and reconstructing it with the instantaneous amplitude matrix to output an enhanced image feature matrix, includes:

[0127] The purpose of this step is to physically modulate the dimensionless coherent structure tensor, which has undergone complex multidimensional processing, back onto the original amplitude data that has the dimension of sound wave energy, thereby generating closed-loop enhancement features that directly match mainstream deep learning segmentation networks.

[0128] To impart realistic acoustic reflection energy properties to the structural features, numerical modulation is first performed using the previously extracted amplitude features, and the phase compensation boundary feature matrix is ​​calculated. The calculation steps are as follows:

[0129] The phase-compensated boundary feature matrix is ​​generated by multiplying the membrane-preserving projection matrix with the instantaneous amplitude matrix.

[0130] Finally, to ensure that the feature matrix input to the deep learning network is completely isomorphic to the original ultrasound image in terms of the dynamic range of physical energy, and to completely eliminate any dimensional conflicts, a dimensionally unified enhanced image feature matrix is ​​constructed and output directly into the deep learning framework. The calculation steps are as follows:

[0131] Multiply the instantaneous amplitude matrix with the phase compensation boundary feature matrix to obtain the numerator fusion term, square the instantaneous amplitude matrix to obtain the amplitude square term, square the phase compensation boundary feature matrix to obtain the feature square term, add the amplitude square term and the feature square term, take the square root of the squared result to obtain the denominator normalization term, and divide the numerator fusion term by the denominator normalization term to generate the enhanced image feature matrix;

[0132] The enhanced image feature matrix is ​​input into a trained deep learning segmentation network, which then outputs a thyroid nodule segmentation mask.

[0133] By multiplying the envelope-preserving projection matrix with the instantaneous amplitude matrix to generate a phase-compensated boundary feature matrix, and then multiplying the instantaneous amplitude matrix with this boundary feature matrix to obtain a numerator fusion term, the two matrices are squared, added, and square-rooted to generate a denominator normalization term. Finally, the numerator fusion term is divided by the denominator normalization term to generate an enhanced image feature matrix. This step reverses the dimensionless coherent structural features after multidimensional correction and physically modulates them back into the amplitude image with the physical dimensions of acoustic energy. Through nonlinear square root mapping, the physical properties are effectively reconstructed, ensuring that the output tensor retains high-fidelity high-order phase boundary information while having an energy dynamic range isomorphic to the original ultrasound image. This eliminates the dimensional conflict of the analytical multi-channel input, enabling it to directly match the input of a multi-layer convolutional neural network.

[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0135] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for segmenting thyroid nodules from ultrasound images, characterized in that, include: Analytical signal decomposition is performed on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal; The spatial phase gradient matrix is ​​generated by extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system. By combining the spatial window size constant with the spatial phase gradient matrix, a spatial translation normalization is performed to construct a local anisotropic matrix; The determinant of the Hessian matrix is ​​determined based on the second-order differential characteristics of the instantaneous amplitude matrix, and the local Hessian mean matrix is ​​summarized within the range of the constant size of the spatial window. The singularity decay matrix is ​​derived by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and then the local anisotropy matrix is ​​modulated by the singularity decay matrix to generate a regularized anisotropy matrix. The amplitude gradient matrix is ​​determined based on the first-order differential characteristics of the instantaneous amplitude matrix, and the local gradient variance matrix is ​​extracted in conjunction with the spatial window size constant. The envelope interface probability matrix is ​​established by combining the magnitude gradient matrix and the local gradient variance matrix, and then the envelope retention projection matrix is ​​generated by fusing the local anisotropy matrix and the regularized anisotropy matrix. The instantaneous amplitude matrix and the envelope-preserving projection matrix are numerically modulated to generate a phase-compensated boundary feature matrix, which is then reconstructed with the instantaneous amplitude matrix to output an enhanced image feature matrix. The enhanced image feature matrix is ​​input into a deep learning segmentation network, which outputs a thyroid nodule segmentation mask.

2. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 1, characterized in that, The step of performing analytical signal decomposition on the ultrasonic echo radio frequency matrix and orthogonal transformation matrix to extract the instantaneous amplitude matrix and instantaneous phase matrix of the corresponding signal includes: Set up and start the ultrasound linear array probe to transmit continuous ultrasound radiofrequency pulses with a determined center frequency to the thyroid target area; The system hardware receiving channel is activated to prepare for receiving ultrasonic echo signals. A spatial coordinate system is established with the horizontal physical arrangement direction of the ultrasonic probe array elements as the horizontal axis and the vertical physical penetration depth direction of ultrasonic wave propagation as the vertical axis. Obtain the two-dimensional ultrasonic echo radio frequency matrix obtained by the ultrasonic probe acquisition and conversion, and obtain the orthogonal transformation matrix generated by the system after performing a discrete Hilbert transformation on the radio frequency matrix along the depth physical direction; The orthogonal transformation matrix is ​​an orthogonal component matrix with the same size as the two-dimensional ultrasonic echo radio frequency matrix, obtained by performing a discrete Hilbert transformation on the two-dimensional ultrasonic echo radio frequency matrix along the depth physical direction. The square of the two-dimensional ultrasonic echo radio frequency matrix is ​​added to the square of the orthogonal transformation matrix, and the square root operation is performed on the result to generate the instantaneous amplitude matrix. The ratio is obtained by dividing the orthogonal transformation matrix by the two-dimensional ultrasonic echo radio frequency matrix, and the instantaneous phase matrix is ​​generated by performing an arctangent operation on the ratio.

3. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 2, characterized in that, The step of extracting the partial derivative of the instantaneous phase matrix along the spatial coordinate system to generate the spatial phase gradient matrix includes: Calculate the first-order partial derivatives of the instantaneous phase matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Squaring the first-order partial derivatives along the horizontal axis and the vertical axis respectively; The results of squaring the two first-order partial derivatives are added together, and the square root of the sum is taken to generate the spatial phase gradient matrix.

4. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 3, characterized in that, The step of performing spatial translation normalization on the spatial phase gradient matrix by combining the spatial window size constant to construct a local anisotropy matrix includes: Set the space window size constant; Within a local spatial window determined by the spatial window size constant, a spatial translation summation operation is performed on the spatial phase gradient matrix. The result of the spatial translation summation operation is divided by the square of the spatial window size constant to obtain the mean value of the local translation gradient. The spatial phase gradient matrix is ​​added to the mean of the local translation gradient to obtain the denominator term, and the spatial phase gradient matrix is ​​divided by the denominator term to generate the local anisotropy matrix.

5. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 4, characterized in that, The process of determining the determinant of the Hessian matrix based on the second-order differential characteristic of the instantaneous amplitude matrix, and summarizing the local Hessian mean matrix within the constant range of the spatial window size, includes: Calculate the second-order pure partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively, and calculate the second-order mixed partial derivatives of the instantaneous amplitude matrix; Multiply the second-order pure partial derivatives in the horizontal direction with the second-order pure partial derivatives in the vertical direction to obtain the first product term. Squaring the second-order mixed partial derivatives to obtain the second squared term. Subtracting the second squared term from the first product term generates the determinant of the Hessian matrix. Within a local spatial window determined by the spatial window size constant, the determinant of the Hessian matrix is ​​subjected to spatial translation and summation. The result of the summation of the Hessian matrix determinant is divided by the square of the spatial window size constant to generate a local Hessian mean matrix.

6. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 5, characterized in that, The process of deriving the singularity decay matrix by combining the determinant of the Hessian matrix and the local Hessian mean matrix, and then using it to modulate the local anisotropy matrix to generate a regularized anisotropy matrix, includes: Divide the determinant of the Hessian matrix by the local Hessian mean matrix to obtain the median ratio, take the negative value of the median ratio, and perform an exponential operation with the base of the natural logarithm as the base and the median ratio after taking the negative value as the exponent to generate a singular decay matrix. The local anisotropy matrix is ​​multiplied by the singularity decay matrix to generate a regularized anisotropy matrix.

7. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 6, characterized in that, The step of determining the amplitude gradient matrix based on the first-order differential feature of the instantaneous amplitude matrix and extracting the local gradient variance matrix in conjunction with the spatial window size constant includes: Calculate the first-order partial derivatives of the instantaneous amplitude matrix along the horizontal and vertical axes of the spatial coordinate system, respectively. Square the first-order partial derivatives along the horizontal and vertical axes, respectively. Add the results of squaring the two first-order partial derivatives, and take the square root of the sum to generate the amplitude gradient matrix. Within a local spatial window determined by the spatial window size constant, the squared values ​​of the magnitude gradient matrix are spatially translated and summed. The result of the spatial translation and summation of the squared magnitude gradient values ​​is divided by the square of the spatial window size constant to obtain the mean of the squared terms. Within the same local spatial window, the magnitude gradient matrix is ​​subjected to spatial translation and summation. The result of spatial translation and summation of the magnitude gradient is divided by the square of the spatial window size constant to obtain the gradient mean. The gradient mean is squared to obtain the mean square term. The mean of the square term is subtracted from the mean square term to generate the local gradient variance matrix.

8. The deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 7, characterized in that, The process of establishing the envelope interface probability matrix by combining the magnitude gradient matrix and the local gradient variance matrix, and then fusing the local anisotropy matrix and the regularized anisotropy matrix to generate the envelope preservation projection matrix, includes: Set a minimum constant; The local gradient variance matrix is ​​added to the minimum constant to obtain the variance correction term. The magnitude gradient matrix is ​​squared to obtain the magnitude gradient squared term. The magnitude gradient squared term is divided by the variance correction term to obtain the correction ratio. The correction ratio is negative, and the exponential decay term is obtained by exponential operation with the base of the natural logarithm as the base and the negative correction ratio as the exponent. The membrane interface probability matrix is ​​generated by subtracting the exponential decay term from the value 1. Multiplying the membrane interface probability matrix with the local anisotropy matrix yields the original feature retention term. Subtracting the membrane interface probability matrix from the value 1 yields the smoothing probability term. Multiplying the smoothing probability term with the regularized anisotropy matrix yields the regularized feature retention term. Adding the original feature retention term and the regularized feature retention term generates the membrane preservation projection matrix.

9. A deep learning-based ultrasound image segmentation method for thyroid nodules according to claim 8, characterized in that, The step of numerically modulating the instantaneous amplitude matrix and the capsule-preserving projection matrix to generate a phase-compensated boundary feature matrix, and reconstructing it with the instantaneous amplitude matrix to output an enhanced image feature matrix, includes: The phase-compensated boundary feature matrix is ​​generated by multiplying the membrane-preserving projection matrix with the instantaneous amplitude matrix. Multiply the instantaneous amplitude matrix with the phase compensation boundary feature matrix to obtain the numerator fusion term, square the instantaneous amplitude matrix to obtain the amplitude square term, square the phase compensation boundary feature matrix to obtain the feature square term, add the amplitude square term and the feature square term, take the square root of the squared result to obtain the denominator normalization term, and divide the numerator fusion term by the denominator normalization term to generate the enhanced image feature matrix; The enhanced image feature matrix is ​​input into a trained deep learning segmentation network, which then outputs a thyroid nodule segmentation mask.

Images