Focused ultrasound micro-disruption image segmentation method based on composite energy field guided snake model

Abstract

The application relates to the technical field of image segmentation, and particularly discloses a focused ultrasound micro-damage image segmentation method based on a composite energy field guided Snake model. In view of the problems of single external energy construction, insufficient anti-interference capability, poor time sequence stability and the like of the existing focused ultrasound micro-damage image segmentation method in processing weak boundary, strong noise and dynamic expansion ultrasound images, the external energy item based on regional texture features, boundary structure and dynamic information is integrated into a composite energy field, multi-source joint driving is applied to the contour evolution of a Snake active contour segmentation model from two dimensions of spatial structure and time evolution, the interference of false critical points and the phenomena of contour drift, collapse and border crossing are significantly inhibited in weak boundary, strong noise and damage dynamic expansion ultrasound images, stable and accurate segmentation of the damage region in a complex ultrasound background is realized, and good time sequence consistency is maintained.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to a focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model. Background Technology

[0002] High-intensity focused ultrasound (HIFU), as a non-invasive tumor treatment technology, has been widely used in the clinical treatment of various solid tumors such as uterine fibroids, liver cancer, and pancreatic cancer due to its advantages such as high precision of thermal ablation, no ionizing radiation, and repeatability. HIFU treatment focuses ultrasound energy on the target area in the body, generating instantaneous high temperatures that cause coagulative necrosis of the target tissue, thereby ablating the lesion. During treatment, real-time and accurate monitoring of the formation and expansion of the lesion area is crucial for evaluating treatment effectiveness and ensuring treatment safety. Ultrasound imaging, due to its real-time performance, low cost, and lack of radiation, has become one of the most commonly used monitoring methods in HIFU treatment.

[0003] However, ultrasound images during HIFU treatment present several inherent challenges, severely limiting the accuracy and stability of lesion segmentation. First, tissue damage caused by HIFU irradiation typically appears in ultrasound images as areas with blurred boundaries, low contrast, and insignificant echo changes, especially in micro-lesions or the early stages of lesion formation. The gray-level difference between the lesion area and surrounding normal tissue is extremely weak, making traditional gray-level gradient-based segmentation methods difficult to effectively distinguish. Second, ultrasound images commonly contain strong speckle noise, acoustic shadowing, artifacts, and tissue motion interference. These factors further obscure the true boundaries of the lesion, leading to problems such as contour drift, boundary overshooting, or unstable convergence in segmentation algorithms. Furthermore, HIFU lesions exhibit dynamic expansion during treatment, with the lesion area continuously evolving over time, requiring segmentation methods to possess good temporal consistency and dynamic response capabilities.

[0004] To address the problem of lesion segmentation in ultrasound images, researchers have proposed various image segmentation methods. Threshold-based segmentation methods are simple to implement but struggle with uneven gray-level distribution and blurred boundaries. Region-growing-based segmentation methods are sensitive to the selection of initial seed points and are prone to region leakage under weak boundary conditions. Edge-detection-based segmentation methods are susceptible to noise interference and struggle to form continuous, closed lesion boundaries. Active contour models (Snake), a classic energy-minimization-based segmentation method, can maintain the smoothness and continuity of contours under internal energy constraints and drive the contour evolution towards the target boundary through external energy, making it widely used in medical image segmentation. However, traditional Snake models primarily rely on image gradient information to construct external energy. In HIFU ultrasound images, the gradient at the lesion boundary is weak and background noise is strong, making it difficult for a single gradient external force to provide a stable and effective driving force. The contour is easily affected by false edges during evolution, leading to inaccurate segmentation results or convergence failure.

[0005] To enhance the capture range and noise resistance of the Snake model, researchers proposed the Gradient Vector Flow (GVF) model and its improved methods. GVF expands the effective range of external forces by diffusing the gradient field, guiding the contour towards the concave boundary. However, in weak gradient and noisy regions of ultrasound images, the GVF diffusion process easily generates vector collisions with opposite directions or small amplitudes, forming numerous unstable pseudo-critical points. These pseudo-critical points do not correspond to the true lesion boundary but incorrectly attract the contour evolution, leading to "collapse" or "boundary crossing" in the segmentation results. Although subsequent studies have smoothed and corrected the GVF field by introducing filtering strategies or spherical external forces, it is still difficult to fundamentally eliminate the interference of pseudo-critical points caused by noise and weak gradient regions. Furthermore, during the dynamic expansion of the lesion, existing methods lack effective utilization of temporal information, and the segmentation results are prone to fluctuations or inconsistencies between consecutive frames.

[0006] In summary, existing HIFU damage segmentation methods based on active contour models generally suffer from problems such as limited external energy construction, insufficient anti-interference ability, and poor temporal stability when processing ultrasound images with weak boundaries, strong noise, and dynamic expansion, making it difficult to meet the requirements for accurate segmentation of minor and early-stage damage. Summary of the Invention

[0007] This invention provides a focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model. The technical problem it solves is: how to construct a composite energy field that can fully utilize image region information, boundary structure information and temporal dynamic information, and effectively suppress pseudo-critical point interference caused by noise, so as to achieve stable and accurate segmentation of the damage area under complex ultrasound background.

[0008] To address the above technical problems, this invention provides a focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model, comprising the following steps:

[0009] Input an ultrasound sequence for dynamic detection to obtain damage candidate regions. Each damage candidate region includes multiple spatially adjacent response points.

[0010] An adaptive ROI region is constructed based on the spatial distribution range of response points within the damage candidate region.

[0011] Generate a structurally sound and continuously closed initial contour within the adaptive ROI region;

[0012] Driven by the composite energy field, the initial contour is iteratively updated using the Snake active contour segmentation model to obtain the segmentation result; the composite energy field is composed of external energy terms based on regional texture features, boundary structure and dynamic information of the Snake contour's internal region.

[0013] Furthermore, the composite energy field is constructed as a weighted sum of three external energy terms based on regional texture features, boundary structure, and dynamic information of the Snake contour's internal region.

[0014] Furthermore, the external energy term based on region texture features is calculated as follows:

[0015] After normalizing the internal region of the Snake contour, a local window is selected with each pixel position as the center to extract the texture statistical feature vector. The sixth dimension of the statistical vector is taken as the local texture intensity response and then normalized to obtain the texture intensity field.

[0016] Calculate the spatial gradient magnitude of the texture intensity field for each pixel;

[0017] By performing a line integral along the Snake contour with respect to the gradient magnitude of the texture intensity field within the Snake contour region, an external energy term based on the region's texture features is obtained.

[0018] Furthermore, the external energy term based on the boundary structure is calculated as follows:

[0019] Multi-scale wavelet decomposition of the texture intensity field is performed to extract the high-frequency sub-band coefficients in the horizontal, vertical and diagonal directions at each scale;

[0020] By fusing the high-frequency subband energies at various scales and in various directions, a wavelet energy map is obtained.

[0021] The wavelet energy map is normalized and mapped to a confidence weight function;

[0022] The diffusion term of the gradient vector flow field is weighted and smoothed using the confidence weight function to obtain the weighted gradient vector flow field.

[0023] The weighted gradient vector flow field is used as the external energy term based on the boundary structure.

[0024] Furthermore, the external energy term based on dynamic information is calculated in the following way:

[0025] Optical flow calculations are performed on adjacent ultrasound images to obtain pixel-level optical flow vector fields;

[0026] The optical flow amplitude is calculated based on the horizontal and vertical components of the optical flow vector field.

[0027] The optical flow amplitude is used as a dynamic modulation factor to construct an external energy term based on dynamic information.

[0028] Furthermore, generating a structurally sound and continuously closed initial contour within the adaptive ROI region specifically includes:

[0029] Canny edge detection is performed on the ultrasound image within the adaptive ROI region to extract a set of candidate edge points;

[0030] Based on the response center output during the damage detection phase, the candidate edge point set is spatially consistent and the edge contour that matches the detection location is retained.

[0031] A position offset correction strategy is used to adjust the overall position of the screened edge contours so that their center position is consistent with the damage localization result given in the detection stage, thus obtaining the initial contour.

[0032] Furthermore, an adaptive ROI region is constructed based on the spatial distribution range of response points within the damage candidate region, specifically including:

[0033] The initial ROI region is constructed by using the maximum and minimum coordinates of the response points in the horizontal and vertical directions within the damage candidate region as the boundary coordinates of the initial ROI.

[0034] An adaptive expansion mechanism is introduced based on the initial ROI to obtain an adaptive ROI region;

[0035] The adaptive expansion mechanism is as follows: taking the center point of the initial ROI region as the center, it expands in both the horizontal and vertical directions, and the expanded boundary does not exceed the width and height range of the entire ultrasound image; the expansion range is determined by multiplying the scale of the initial ROI region by a preset expansion coefficient.

[0036] Furthermore, the initial contour is iteratively updated under the joint drive of the composite energy field, including:

[0037] In each iteration, the external force generated by the composite energy field is projected onto the normal direction of the contour point, retaining only the normal component;

[0038] The update direction is determined based on the distance from the contour point to the nearest boundary. When the distance from the contour point to the nearest boundary is less than the preset distance threshold and the update direction shows an outward expansion trend, the outward expansion normal component is removed, and only the part that allows shrinkage or tangential adjustment is retained.

[0039] Furthermore, the iterative update of the initial contour under the joint drive of the composite energy field also includes:

[0040] Apply amplitude limits to the displacement of a single update;

[0041] The centroid of the outline is restored to its initial position after each update.

[0042] When the average normal displacement of all contour points is less than the preset convergence threshold, the contour is determined to have converged and the iteration stops.

[0043] Furthermore, the iterative update of the initial contour under the joint drive of the composite energy field also includes: if the contour area is less than the initial area multiplied by a preset ratio threshold during the iteration process, the process reverts to the previous contour and terminates the iteration.

[0044] This invention provides a focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model. By introducing damage dynamic detection results to construct a spatiotemporally adaptive region of interest and automatically generating an initial contour with center correction, the search space is effectively compressed and the spatial consistency between the contour and the actual damage area is improved. Furthermore, a composite energy field is constructed, consisting of an external energy term based on regional texture features, a boundary structure external energy term based on wavelet confidence-weighted gradient vector flow field, and a dynamic information external energy term based on optical flow amplitude modulation. This field applies multi-source joint driving to the contour evolution from both spatial structure and temporal evolution dimensions. Simultaneously, during the iterative update process, normal projection, directional suppression, displacement limiting, centroid locking, and a convergence criterion for average normal displacement and an early stopping mechanism for the lower limit of area are employed. This significantly suppresses pseudo-critical point interference and contour drift, collapse, and boundary crossing phenomena in ultrasound images with weak boundaries, strong noise, and dynamic damage expansion, achieving stable and accurate segmentation of the damage area under complex ultrasound backgrounds while maintaining good temporal consistency. Attached Figure Description

[0045] Figure 1 This is a flowchart of a focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model provided in an embodiment of the present invention;

[0046] Figure 2This is a diagram illustrating the initial contour generation and center correction process provided in an embodiment of the present invention. Figure 2 Image (a) shows the region of interest in the damage. Figure 2 (b) shows the traditional Snake segmentation result. Figure 2 (c) shows the main edge contours after spatial consistency filtering. Figure 2 (d) represents the initial contour after center correction;

[0047] Figure 3 This invention provides a visual comparison of representative frames from the HIFU ultrasound sequence across three irradiation stages. Figure 3 In the image, (a1), (b1), (c1), (d1), (e1), and (f1) represent the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the early stages of the proposed method, respectively. Figure 3 In the image, (a2), (b2), (c2), (d2), (e2), and (f2) are representative frames in the intermediate stage of the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the proposed method, respectively. Figure 3 In the middle, (a3), (b3), (c3), (d3), (e3), and (f3) are representative frames of the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the late stage of this method, respectively.

[0048] Figure 4 This is a detailed diagram of the composite energy field in the HIFU-damaged region provided in an embodiment of the present invention. Figure 4 (a) is the original ultrasound image. Figure 4 In the middle (b), the original GVF vector field is shown. Figure 4 (c) shows the GVF field after smoothing filtering. Figure 4 (d) shows the traditional Snake segmentation result. Figure 4 (e) is the optical flow vector diagram. Figure 4 f is the wavelet multi-scale confidence matrix. Figure 4 In the middle (g), the external force field is the Wavelet-GVF field. Figure 4 (h) represents the final segmentation result after the fusion of the composite energy fields. Detailed Implementation

[0049] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The embodiments are given for illustrative purposes only and should not be construed as limiting the present invention. The accompanying drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0050] The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model provided in this embodiment of the invention, such as... Figure 1 The flowchart shown includes the following steps:

[0051] S1. Input the ultrasound sequence for dynamic detection to obtain the damage candidate region. Each damage candidate region includes multiple response points that are spatially adjacent.

[0052] S2. Construct an adaptive ROI region based on the spatial distribution range of response points within the damage candidate region;

[0053] S3. Generate an initial contour with a reasonable structure and continuous closure within the adaptive ROI region;

[0054] S4. Under the joint drive of the composite energy field, the initial contour is iteratively updated using the Snake active contour segmentation model to obtain the segmentation result; the composite energy field is composed of external energy terms based on regional texture features, boundary structure and dynamic information of the Snake contour's internal region.

[0055] This method utilizes the spatial prior information (damage candidate region) provided during the detection phase to effectively limit the segmentation search range. Furthermore, by introducing various image and dynamic information, a composite energy field is constructed to guide the Snake active contour segmentation model to evolve stably in complex ultrasound backgrounds.

[0056] The Snake active contour model is highly sensitive to the initial contour position and search space range. Directly initializing and evolving the contour within the entire ultrasound image is susceptible to interference from background noise, artifacts, and non-target structures, especially when the damage boundary is weak, leading to instability or false convergence in the segmentation process. Therefore, effectively constraining the search space and generating a reasonable initial contour during the segmentation stage is crucial for improving segmentation stability. In the algorithm implementation, the damage dynamic detection stage often generates multiple detection response points within the same damage region. These response points exhibit certain spatial clustering characteristics, but their positions are not entirely consistent. Directly using a single detection point as the segmentation center can easily lead to ROI shift, thus affecting the integrity of the segmentation result. To address this, this invention introduces a spatiotemporally adaptive ROI construction and initial contour automatic generation method based on damage detection results during the segmentation stage. This method utilizes detection priors to jointly constrain the segmentation search space and initialization process, thereby improving the stability and efficiency of the segmentation process.

[0057] In step S1, based on the detection response points obtained from the damage dynamic detection algorithm, spatial aggregation processing is performed on the response points to construct damage candidate regions. Let the set of response points detected in the current frame of the ultrasound image be denoted as . Where N represents the number of detection response points, This represents the coordinates of the k-th response point in the image. Next, connectivity analysis is performed on the set of response points, merging spatially adjacent response points into the same damage candidate region. For the r-th damage candidate region, its corresponding response set can be represented as... ,in This represents the number of response points within the region.

[0058] In step S2, an initial ROI region is first constructed based on the spatial distribution range of response points within the damage candidate region. The maximum and minimum coordinates of the response points in the horizontal and vertical directions are used as the boundary coordinates of the initial ROI region, denoted as follows: To ensure complete coverage of the actual lesion area, an adaptive expansion mechanism is introduced based on the initial ROI: the ROI is expanded horizontally and vertically, with the expanded boundary not exceeding the width and height of the entire ultrasound image; the expansion range is determined by multiplying the scale of the initial ROI by a preset expansion coefficient.

[0059] Let the initial ROI have the following scales in the horizontal and vertical directions: ( These are the maximum and minimum coordinates of the initial ROI in the horizontal direction, respectively. (where are the maximum and minimum coordinates of the initial ROI in the vertical direction, respectively). The expanded ROI boundary is shown in the following equation:

[0060] ,

[0061] Where W and H represent the width and height of the ultrasound image, respectively. This is the ROI expansion coefficient, used to control the degree of expansion. In this method, it is taken as... =0.75, The coordinates of the center point of the initial ROI. , ; , The minimum and maximum coordinates of the adaptive ROI region in the horizontal direction; , The minimum and maximum coordinates of the adaptive ROI region in the vertical direction.

[0062] For continuous ultrasound sequences, the above ROI construction process is repeated in each frame of the image, where the ROI corresponding to the t-th frame can be represented as:

[0063] ,

[0064] The location parameters of the ROI are recalculated from the detection results of the current frame.

[0065] In this way, the ROI can be dynamically updated as the location of the damaged area changes in the ultrasound sequence, thus ensuring the continuity and consistency of the segmentation process in the time dimension. By introducing spatial constraints based on detection priors, this method effectively narrows the search range of Snake segmentation, reduces the interference of background noise on the segmentation results, and improves the stability and computational efficiency of the overall segmentation process.

[0066] In step S3, after completing the spatiotemporal adaptive ROI construction, this method further generates the initial contour required for the Snake model within the adaptive ROI region. If the initial contour is generated directly using edge detection methods within the ROI, it is easily affected by local noise and artifacts, leading to contour position shifts and affecting the stability of subsequent Snake evolution. Therefore, this method introduces the automatic micro-damage detection results as a spatial prior to constrain and correct the initial contour generation process.

[0067] Specifically, firstly, Canny edge detection is performed on the ultrasound image within the constructed adaptive ROI region to extract a set of candidate edge points. Then, based on the response center output from the lesion detection stage, the candidate edge point set is spatially consistent, retaining only the main edge contours that match the detection localization. Further, a positional offset correction strategy is used to adjust the overall position of the selected edge contours, ensuring their center position is consistent with the lesion localization result given in the detection stage, thereby avoiding initial contour offset problems caused by noise interference or incomplete edges. The initial contour generation and center correction process is as follows: Figure 2 As shown, Figure 2 Image (a) shows the region of interest in the damage. Figure 2 (b) shows the traditional Snake segmentation result; Figure 2 (c) shows the main edge contours after spatial consistency screening; Figure 2 (d) is the initial contour after center correction.

[0068] The contour obtained after the above processing is used as the initial contour input for the Snake model. This method fully utilizes the spatial prior information provided by the automatic micro-damage detection results to achieve automatic generation and position correction of the initial contour. Compared with directly using the edge detection results as the initial contour, this strategy can significantly improve the spatial consistency between the initial contour and the actual damage area, providing a reliable starting point for the stable evolution of the Snake contour under multi-source external force driving.

[0069] For step S4, after completing the spatiotemporal adaptive ROI construction and initial contour automatic generation based on the detection results, this method further proposes a Snake active contour segmentation model that utilizes multiple external forces to achieve fine segmentation of HIFU damage regions in complex ultrasound imaging environments. Based on the traditional Snake framework, this model introduces texture statistical features, boundary structure information, and dynamic optical flow information, and combines these with prior detection constraints to guide the contour to evolve stably and accurately towards the true damage boundary.

[0070] The Snake model achieves the iterative evolution of the Snake profile by minimizing the energy functional. Essentially, it's an energy optimization process under the combined influence of internal constraints and external driving forces. By continuously adjusting the positions of the Snake profile points, the overall energy function is minimized, thus gradually converging the Snake profile to the target boundary. (Energy function) Usually composed of internal energy and external energy The composition can be represented as:

[0071] ,

[0072] in, Used to constrain the smoothness and continuity of the snake profile, preventing violent oscillations or irregular deformations during its evolution. The ability to attract contours toward the target boundary is the key driving force behind Snake model's target segmentation.

[0073] The internal energy is mainly composed of first and second derivative terms, and its mathematical expression is:

[0074] ,

[0075] in, The coordinates of the Snake contour points (s refers to the Snake contour points) are represented by the first term, which is the first derivative term, used to describe the distance change between adjacent Snake contour points. Its physical meaning is equivalent to an elastic constraint, which can limit the stretching between Snake contour points and prevent the contour from breaking or being excessively stretched. The second term is the second derivative term, used to describe the change in contour curvature. Its physical meaning is equivalent to a rigid constraint, which can suppress excessive bending of the Snake contour and keep the Snake contour smooth overall. Control the elasticity and rigidity of the Snake profile separately.

[0076] From an energy optimization perspective, the evolution of the Snake model can be viewed as a process of gradually reaching a state of force equilibrium under the combined influence of internal and external energy. When the internal constraint force and the external driving force reach equilibrium on the contour, the energy function reaches its minimum value, and the Snake contour stops evolving and stabilizes near the target boundary. Therefore, the reasonable construction of the Snake energy function and the setting of the weight parameters of each energy term directly determine the stability of the Snake contour evolution process and the accuracy of the segmentation results.

[0077] External energy term It is a key factor driving the evolution of the Snake contour toward the target boundary, and its design directly determines the attraction of the segmentation result to the target region and the suppression effect on complex backgrounds. Considering the characteristics of HIFU damage in ultrasound images, such as blurred boundaries, indistinct texture changes, and dynamic evolution, a single external energy form is often difficult to balance segmentation accuracy and stability.

[0078] In traditional Snake models, the external energy primarily functions in regions with significant image gradients. When gradients are weak, boundaries are discontinuous, or blurred at target edges, the Snake contour struggles to acquire effective boundary guidance information, leading to limited capture range and unstable segmentation results. Furthermore, when dealing with targets with complex shapes or concave structures, while the smoothing constraint in the internal energy term helps maintain the continuity of the Snake contour, it also limits its local flexibility, making it difficult to fully conform to complex boundary structures, resulting in phenomena such as "crossing" or "straightening." Simultaneously, speckle noise and artifacts prevalent in ultrasound images can easily create false gradients, introducing false edge information. These noise interferences mislead the evolution direction of the Snake contour, reducing the model's ability to respond to real target boundaries and further decreasing the accuracy and robustness of the segmentation results.

[0079] Based on the above analysis, relying solely on single gradient information to construct external energy is insufficient to meet the practical requirements of HIFU injury ultrasound image segmentation. To overcome the shortcomings of traditional Snake models in terms of capture range, complex boundary handling, and noise resistance, this method introduces a composite energy field joint driving mechanism under prior detection constraints. By fusing regional statistical features, boundary structure information, and dynamic optical flow features, it provides complementary and stable driving forces for the Snake contour from multiple perspectives, laying the foundation for subsequent fine segmentation.

[0080] In one implementation, the combined drive of external energy can be equivalently represented as:

[0081] ,

[0082] in, This represents the external energy term based on region texture features. Represents the external energy term based on the boundary structure. This represents the external energy term based on dynamic information. Indicates the current contour point The relative contribution coefficients of each external force term adaptively change with the position of the contour point and local image evidence. Furthermore, the regional texture response and dynamic information response can first construct the target confidence through preset fusion coefficients, and then the boundary structure response can be gated to adjust the boundary traction term and the outward expansion driving term, thereby achieving adaptive joint driving of each external force term during contour evolution. HIFU-damaged areas in ultrasound images are often accompanied by changes in local tissue structure, statistically manifested as changes in texture features. To enhance the model's ability to perceive differences in regional attributes, this method introduces a regional external energy term based on local texture feature statistics within the ROI. By characterizing the difference in texture distribution between the target region and the background region, it guides the Snake contour to evolve towards the true boundary.

[0083] First, the normalized input image (Initial in-contour image) Construct a texture intensity field. For each pixel location in the image... Select a local window of size R*L centered on it:

[0084] ,

[0085] Within this local neighborhood, a texture statistics operator is used to analyze the image. Specifically, for pixel locations... In its corresponding local window Extracting texture feature vectors from within ,in This represents the texture statistical feature extraction process. This paper selects the sixth dimension of the feature vector as the stability intensity response to characterize the stability complexity of a local region, defined as: To eliminate differences in texture response across different regions at numerical scales and to facilitate subsequent texture field analysis, the texture field intensity is normalized, thereby obtaining a texture intensity field with a consistent scale globally. This provides a foundation for subsequent texture gradient calculations and the construction of external energy in the region.

[0086] Then the spatial gradient of the texture gradient field is calculated. and gradient magnitude Spatial gradient reflects the rate of change of texture in space, and its magnitude reflects the degree of abrupt change in texture statistical features in space. When the Snake curve is located near the boundary of the real target, if the texture distribution difference between the areas inside and outside the curve is significant, the gradient magnitude will be... The gradient magnitude is set to a larger value, while it is smaller within regions with uniform texture. Therefore, this method measures the texture gradient magnitude. As an important basis for the external energy of the region, it guides the contour to converge towards the boundary position with the greatest texture difference.

[0087] Based on the gradient magnitude of the texture gradient field mentioned above The texture-driven external energy term can be constructed by integrating the gradient magnitude of the texture intensity field over the entire Snake contour:

[0088] ,

[0089] Where C represents the current Snake curve, Define the image domain. This is the Dirac function corresponding to the contour curve. This energy term constrains contour evolution through texture statistical features, guiding the contour to cluster towards regions with significant texture changes, thus enabling the model to maintain stable and accurate segmentation performance even under weak boundary and complex texture background conditions.

[0090] After obtaining a texture field based on texture statistical features, although it can characterize the region boundary location to some extent, its range of influence is usually limited to the vicinity of the boundary. Its ability to pull the Snake curve at long distances is limited, and it easily leads to insufficient curve convergence in concave boundaries or complex structural regions. To enhance the capture range of the external force field and improve its adaptability to concave boundaries, this method introduces a Gradient Vector Flow (GVF) model to extend the texture gradient field. The classic GVF is obtained by minimizing the following energy functional:

[0091] ,

[0092] in, It's a GVF field. and These represent the components of the GVF vector field in the horizontal (x-axis) and vertical (y-axis) directions of the image, respectively. Let be the potential energy function, representing the texture potential energy of the image; The first term is the smoothing coefficient; the second term is the smoothing diffusion term, and the third term is the data preservation term used to approach the true boundary. The GVF field diffuses the initial gradient vector field, allowing external forces to maintain a tendency to point towards the boundary over a large range, thereby effectively improving the convergence range of the Snake model and its ability to track complex boundary structures.

[0093] However, GVF fields still have certain limitations in practical applications: in noisy or weak gradient regions, due to the unreliability of gradient information, the GVF diffusion process is prone to generating vector hedging phenomena with opposite directions or small amplitudes, thus forming a large number of unstable critical points in the vector field.

[0094] To alleviate the aforementioned problems, existing studies typically introduce filtering strategies to smooth the GVF field, thereby suppressing local vector oscillations. Filtering operations can reduce the number of vector confluence points to some extent. However, relying solely on spatial filtering is insufficient to completely eliminate pseudo-critical points caused by noise or weak gradient regions. These critical points are often not true target boundaries, but rather caused by random perturbations or low-contrast pseudo-edges in texture-flat regions, and may still interfere with the stable evolution of the Snake curve.

[0095] Based on this, in order to further suppress pseudo-critical points caused by noise or weak gradient regions, this method introduces a confidence weighting mechanism based on wavelet transform during the construction of the GVF field to adaptively modulate the GVF field.

[0096] First, the texture intensity field is decomposed into high-frequency subband coefficients at different scales to obtain the coefficients of the corresponding subbands, including the horizontal subband (LH), vertical subband (HL), and diagonal subband (HH). Since real target edges and bright spot structures in the image have significant responses in the high-frequency components, while flat regions and random noise have weaker responses in the high-frequency subbands, the high-frequency wavelet coefficients can effectively reflect the reliability of local structures. Based on these characteristics, this method fuses the energy of the multi-scale high-frequency subbands to construct a wavelet energy map.

[0097] ,

[0098] in, Indicates the wavelet decomposition scale. They represent the first The high-frequency wavelet coefficients in the horizontal, vertical, and diagonal directions are obtained from the layer decomposition. Subsequently, the wavelet energy map is normalized (e.g., linear normalization) and mapped to a confidence weight function. The weighting function describes the reliability of texture and structural information at the current pixel location; a larger value indicates that the location is more likely to be near the boundary of the real target. Based on this, by utilizing... We perform weighted smoothing on the classical GVF field to form a weighted GVF energy functional based on wavelet confidence:

[0099] .

[0100] By applying the Euler-Lagrange equations to the weighted GVF energy functional, we can obtain the evolution equations of the weighted GVF field:

[0101] ,

[0102] As can be seen from the above formula, the confidence weight function By modulating the diffusion term of the GVF field, the diffusion of the vector field is reduced in the high-confidence region, thus effectively maintaining the consistency of the vector direction near the true boundary; while in the low-confidence region, the diffusion effect is enhanced, thereby suppressing vector hedging phenomena and the generation of pseudo-critical points caused by noise or weak gradient regions. In the Snake curve evolution process, this method introduces the aforementioned weighted GVF field as an external force field into the model, defined as:

[0103] .

[0104] Based on the construction of external potential energy based on texture statistical features, this method introduces a wavelet confidence-weighted GVF field and effectively suppresses the pseudo-critical point problem caused by noise and weak gradient regions through an adaptive modulation vector diffusion process. This provides more reliable regional external energy support for the stable evolution of Snake contour curves under complex background and weak boundary conditions.

[0105] HIFU lesions are often accompanied by minute tissue movements and gradual evolution of lesion morphology during treatment. Relying solely on grayscale, texture, and boundary features extracted from a single static image can easily overlook the temporal changes in the lesion region, thus affecting the consistency and stability of the segmentation results across the image sequence. To enhance the Snake model's adaptability to the dynamic evolution of HIFU lesions, this method introduces optical flow information as a dynamic external energy term to temporally constrain the curve evolution.

[0106] By performing optical flow calculations on adjacent frames of ultrasound images, pixel-level optical flow vector fields can be obtained. The magnitude of the optical flow vector reflects the intensity of motion in a local region over time; based on this, an optical flow amplitude map is constructed:

[0107] ,

[0108] in, , They represent Motion components in the horizontal and vertical directions.

[0109] Regions with large optical flow amplitudes typically correspond to locations where significant structural changes occur and are often closely related to the expansion behavior of HIFU-damaged regions. Based on the above analysis, this method introduces the optical flow amplitude as a dynamic modulation factor into the external energy model. Specifically, the optical flow amplitude is normalized to construct a dynamic external energy term:

[0110] ,

[0111] in, This represents the normalized optical flow amplitude, used to characterize the intensity of local change at the current contour point in the time dimension. The actual contribution of the dynamic external energy term to the total external energy is given by the previously mentioned... Adjustments were made.

[0112] By introducing this dynamic external energy term, the Snake curve can gain a stronger driving force in regions with significant motion changes, thus responding more sensitively to the morphological evolution of the damaged area in the time series.

[0113] In practical applications, the dynamic external energy term works synergistically with the region external energy based on texture statistical features and the wavelet confidence-weighted GVF field. The region and structure energy terms primarily constrain the spatial localization of the contour, while the dynamic energy term supplements the constraint on curve evolution from a temporal perspective. Through the fusion of multi-source external information, the constructed model maintains good segmentation stability and temporal consistency even under complex noise backgrounds and weak boundary conditions.

[0114] Driven by multiple external energies, the Snake profile gradually converges to the damage boundary through iterative updates. To improve the stability of profile evolution in a high-noise environment, this method introduces directional constraints, step size control, and an adaptive stopping criterion during the iteration process.

[0115] First, to suppress the overall drift caused by tangential perturbation, this method only retains the component of the external force in the contour normal direction (normal projection external force). Let the i-th contour point in the k-th iteration be... The unit outward normal is External force Then the normal projected external force is: .

[0116] Secondly, to avoid ineffective outward expansion of the contour near the boundary, this method constructs a directional suppression strategy based on boundary distance transformation. Let... Represents contour points The distance to the nearest boundary, when satisfying ( When the update direction shows an outward expansion trend (with a preset distance threshold), the outward expansion normal component is removed, and only the portion that allows for contraction or tangential adjustment is retained, thereby enhancing the reliability of boundary convergence. Simultaneously, to prevent excessive external force from causing abrupt contour changes, this method imposes amplitude constraints on the single update displacement and introduces a centroid locking strategy, returning the centroid of the contour after each update to the initial centroid position to suppress overall translation.

[0117] Finally, this method uses the average normal displacement as the profile stability criterion. Let the profile point update amount be... Then when it satisfies When the contour has converged, the iteration terminates, where N represents the total number of contour points. This indicates a preset convergence threshold. On the other hand, to avoid contour collapse, this method introduces an early stopping mechanism with a lower area limit: when the contour area... Smaller than the initial area Multiply by the proportional threshold When the time comes, roll back to the previous contour and terminate the iteration to ensure the integrity of the segmented region.

[0118] This method incorporates corresponding constraints and stopping strategies during the contour iterative evolution process to improve the stability and robustness of the segmentation results.

[0119] In summary, the focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model provided by this invention addresses the problems of contour drift, pseudo-critical point interference, and convergence instability that easily occur in traditional Snake methods under weak boundary and strong noise environments. This method systematically improves upon these problems from three aspects: spatial constraints, structural guidance, and temporal consistency.

[0120] First, during the segmentation initialization stage, damage dynamic detection results are introduced to construct a spatiotemporally adaptive ROI, effectively compressing the search space. At the same time, through the automatic generation of the initial contour and the offset correction strategy, the spatial consistency between the contour and the actual damage area is significantly improved.

[0121] Secondly, regarding external energy construction, a multi-source external force joint driving mechanism is proposed, integrating regional external forces based on texture statistical features, wavelet confidence-weighted GVF structural external forces, and dynamically modulated external forces based on optical flow intensity. This imposes constraints on contour evolution from two dimensions: spatial structural information and temporal evolution information. Among these, the wavelet confidence mechanism can effectively suppress the pseudo-critical point problem caused by noise regions, while optical flow intensity modulation enhances the model's response to dynamically expanding damage regions, thereby improving the consistency and stability of temporal segmentation.

[0122] During the contour iterative evolution process, this method uses normal projection external force to suppress tangential disturbance, constructs directional suppression based on boundary distance to prevent invalid expansion, limits the amplitude of single update displacement and introduces centroid locking to suppress overall translation; at the same time, it uses the average normal displacement as the convergence criterion and sets an early stopping mechanism with lower area limit to avoid contour collapse.

[0123] To verify the performance advantage of the proposed method in the HIFU damage segmentation task, several typical segmentation methods were selected for comparative experiments, including: (1) the traditional Snake method (Snake); (2) the Snake method based on standard GVF (GVF-Snake); (3) the iGVF-Snake method with added spherical external force and filtering constraints (iGVF-Snake); and (4) the proposed method. All methods were tested under the same ROI range and the same initialization conditions to ensure the fairness of the comparison. The average performance index results of each method on 56 sets of experimental data are shown in Table 1. Dice (dice coefficient) is used to measure the similarity between two sample sets. Essentially, it is the ratio of twice the overlapping area of ​​the two sets to the sum of their total sizes. IoU (Intersection over Union) is the ratio of the intersection to the union of two sets. Compared with Dice, IoU has stricter requirements on the overlap rate and is a more stringent evaluation standard. The larger the Dice and IoU values, the better the segmentation performance. MBD (Mean Boundary Distance) represents the degree to which the segmentation boundary closely matches the ground truth boundary. The smaller the value, the closer the segmentation result is to the ground truth boundary. The best score for each metric is highlighted in bold.

[0124] Table 1

[0125]

[0126] As shown in Table 1, the traditional Snake method is prone to contour overflow in weak boundary regions, resulting in a low Dice coefficient of only 0.68 and a large boundary error, with a MBD value reaching 2.81. The GVF-Snake method, by introducing gradient vector flow, expands the capture range to some extent, but still performs poorly in terms of Dice (0.508) and MBD (5.043), especially in noisy regions where pseudo-critical point interference remains, leading to instability in the segmentation results. The iGVF-Snake method, by introducing wavelet confidence-weighted GVF, improves the Dice coefficient (0.677) and IoU (0.663) compared to the GVF-Snake method, but during the dynamic expansion phase of the damaged region, the segmentation results still fluctuate over time, failing to stably capture all damaged regions. In comparison, the proposed multi-source external force driven Snake method achieves the best results in both Dice coefficient (0.803) and IoU (0.679), and is significantly lower than other methods in MBD (2.435), indicating that this method can more accurately fit the real damage boundary. Specifically, the proposed method optimizes the segmentation accuracy of the damage region from multiple dimensions by fusing external forces in the texture region, wavelet confidence-weighted GVF external forces, and optical flow intensity dynamically modulated external forces. This strategy further enhances the fitting and conformation ability of the active contour to the real tissue boundary while ensuring high stability and high accuracy of the segmentation results.

[0127] To further verify the above conclusions visually and analyze the segmentation stability of each method at different irradiation stages, this paper selects representative frames from the same set of HIFU ultrasound sequences at three irradiation stages for visual comparison. The results are as follows: Figure 3 As shown, Figure 3 In the image, (a1), (b1), (c1), (d1), (e1), and (f1) represent the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the early stages of the proposed method, respectively. Figure 3 In the image, (a2), (b2), (c2), (d2), (e2), and (f2) are representative frames in the intermediate stage of the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the proposed method, respectively. Figure 3 In the image, (a3), (b3), (c3), (d3), (e3), and (f3) are representative frames of the original image, the ground truth image, the Snake method, the GVF-Snake method, the iGVF-Snake method, and the proposed method in the late stage, respectively.

[0128] from Figure 3 It can be seen that the different methods show significant differences in their performance during the lesion formation and propagation process:

[0129] In the early stages, the damage echo is enhanced, and the weak boundary is discontinuous. Traditional snakes are prone to undersegmentation, such as... Figure 3 In the middle (c1); although GVF-Snake expands the capture range, it produces abnormal inward traction under the influence of noise and weak gradient artifacts, causing the contour to "collapse" and deviate from the true damage, such as Figure 3 In the middle (d1), iGVF-Snake effectively alleviates collapse and improves stability through filtering and spherical equilibrium external forces, but the traction of weak boundary regions is still insufficient, and the segmentation results are still slightly undersegmented, such as... Figure 3 In the middle (e1); by adding optical flow intensity, it can capture the minute movements and expansion trends of adjacent frames, providing temporal guidance for the contour, making it less prone to stalling along the expansion direction, and resulting in more complete boundary coverage, such as Figure 3 (f1)

[0130] The overall pattern in the mid-term stage is similar to that in the early stage: Snake is still undersegmented, GVF-Snake is unstable, iGVF-Snake has improved but still has insufficient local coverage, and methods with optical flow are more continuous in tracking the expansion direction.

[0131] In the later stages, the lesion morphology is larger and the artifacts are stronger, making it impossible for traditional Snake to segment the lesion based on the lesion detection points, such as... Figure 3 (c3); GVF-Snake collapse further aggravated, such as Figure 3 In the middle (d3); iGVF-Snake exhibits a certain degree of oversegmentation, manifested as the contour expanding into the non-damaged highlighted area, such as Figure 3 (e3)

[0132] In contrast, this method can balance noise suppression and boundary traction in all three stages, avoiding collapse and overexpansion, and achieving a more stable and accurate damage segmentation effect.

[0133] To further verify the effectiveness of the proposed composite energy field in HIFU micro-damage segmentation, the synergistic effect of external forces in the texture region, wavelet confidence screening of GVF boundaries, and optical flow dynamic constraints was analyzed. The superiority of this approach is illustrated through visual comparison of the results, as shown in the figure. Figure 4 As shown, Figure 4 (a) is the original ultrasound image. Figure 4 In the middle (b), the original GVF vector field is shown. Figure 4 (c) shows the GVF field after smoothing filtering. Figure 4 (d) shows the traditional Snake segmentation result. Figure 4 (e) is the optical flow vector diagram. Figure 4 f is the wavelet multi-scale confidence matrix. Figure 4 In the middle (g), the external force field is the Wavelet-GVF field. Figure 4 (h) represents the final segmentation result after the fusion of the composite energy fields.

[0134] In ultrasound images, the standard GVF field force is easily affected by background noise due to the characteristics of weak gradients at the damage boundary, structural fractures, and significant noise artifacts. Figure 4 Figure (b) shows the unfiltered GVF vector field. It can be observed that there are numerous directional irregularities and local contralateral phenomena in the weak gradient and speckle noise regions. In some areas, even "vortex-like" locally stable structures are formed. These structures do not correspond to the true boundary but are diffusion results caused by noise or weak gradient artifacts, often leading to inward collapse of the segmented contour. Even after conventional smoothing processing, such as... Figure 4 In (c), the vector continuity is improved, but some hedging tendency still remains near the true boundary, thus forming a pseudo-critical point in the Snake evolution process, such as... Figure 4 (d) In this type of critical point, there is no obvious edge support in the original image, and the gradient direction distribution around it is highly disordered, which can easily cause incorrect attraction to the contour, resulting in the segmentation result deviating from the true damage boundary.

[0135] To address the aforementioned issues, this paper introduces wavelet multi-scale energy to construct a pixel-level confidence matrix, such as... Figure 4 As shown in (f), since the target edge and bright spot structure have strong amplitude responses in the high-frequency subbands (LH, HL, HH), while the high-frequency energy in the flat noise region is weak, the normalized multi-scale wavelet energy is used as the structural reliability weight to modulate the GVF diffusion process. The weighted Wavelet-GVF results are shown in Figure (f). Figure 4 As shown in (g), the vector directions are more consistent near the real structure, while the pseudo-hedging and local convergence phenomena are significantly reduced in regions without obvious boundary support, thus effectively suppressing the formation of pseudo-critical points.

[0136] Meanwhile, considering the dynamic evolution of the damaged area expanding frame by frame over time, this paper introduces optical flow constraint to modulate the temporal consistency of the contour evolution process. However, due to the presence of a large amount of speckle noise and local brightness fluctuations in ultrasound images, the optical flow field is prone to directional convergence or vortex-like flow field structures in weakly textured or noisy regions. Figure 4 In section (e), the phenomenon manifests as vector convergence and divergence driven by non-real motion. Therefore, this paper uses optical flow intensity to describe the motion of the damage. The final segmentation result is as follows: Figure 4 As shown in (h). By utilizing the composite energy field mechanism, the noise-induced spurious attraction structure is weakened while maintaining the true boundary traction capability, making the external force field distribution more consistent with the actual damage morphology. This provides a reliable guarantee for the subsequent stable convergence of the Snake, verifying the superiority of the multi-source external force construction strategy in the weak boundary ultrasonic environment.

[0137] The above-mentioned ex vivo tissue experiments verified the superior performance of the method proposed in this invention in terms of Dice coefficient, IoU, and boundary error, and proved that the method can achieve more stable and accurate damage boundary fitting under complex ultrasound background and weak boundary conditions, and has good robustness and clinical application potential.

[0138] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A focused ultrasound micro-damage image segmentation method based on a composite energy field-guided Snake model, characterized in that, Including the following steps: Input an ultrasound sequence for dynamic detection to obtain damage candidate regions. Each damage candidate region includes multiple spatially adjacent response points. An adaptive ROI region is constructed based on the spatial distribution range of response points within the damage candidate region. Generate a structurally sound and continuously closed initial contour within the adaptive ROI region; Driven by the composite energy field, the initial contour is iteratively updated using the Snake active contour segmentation model to obtain the segmentation result; the composite energy field is composed of external energy terms based on regional texture features, boundary structure and dynamic information of the Snake contour's internal region.

2. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 1, characterized in that: The composite energy field is constructed as a weighted sum of three external energy terms based on regional texture features, boundary structure, and dynamic information within the Snake contour region.

3. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 2, characterized in that, The external energy term based on region texture features is calculated in the following way: After normalizing the internal region of the Snake contour, a local window is selected with each pixel position as the center to extract the texture statistical feature vector. The sixth dimension of the statistical vector is taken as the local texture intensity response and then normalized to obtain the texture intensity field. Calculate the spatial gradient magnitude of the texture intensity field for each pixel; By performing a line integral along the Snake contour with respect to the gradient magnitude of the texture intensity field within the Snake contour region, an external energy term based on the region's texture features is obtained.

4. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 3, characterized in that, The external energy term based on the boundary structure is calculated in the following way: Multi-scale wavelet decomposition of the texture intensity field is performed to extract the high-frequency sub-band coefficients in the horizontal, vertical and diagonal directions at each scale; By fusing the high-frequency subband energies at various scales and in various directions, a wavelet energy map is obtained. The wavelet energy map is normalized and mapped to a confidence weight function; The diffusion term of the gradient vector flow field is weighted and smoothed using the confidence weight function to obtain the weighted gradient vector flow field. The weighted gradient vector flow field is used as the external energy term based on the boundary structure.

5. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 4, characterized in that, The external energy term based on dynamic information is calculated in the following way: Optical flow calculations are performed on adjacent ultrasound images to obtain pixel-level optical flow vector fields; The optical flow amplitude is calculated based on the horizontal and vertical components of the optical flow vector field. The optical flow amplitude is used as a dynamic modulation factor to construct an external energy term based on dynamic information.

6. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to any one of claims 1 to 5, characterized in that, Generating a structurally sound and continuously closed initial contour within the adaptive ROI region specifically includes: Canny edge detection is performed on the ultrasound image within the adaptive ROI region to extract a set of candidate edge points; Based on the response center output during the damage detection phase, the candidate edge point set is spatially consistent and the edge contour that matches the detection location is retained. A position offset correction strategy is used to adjust the overall position of the screened edge contours so that their center position is consistent with the damage localization result given in the detection stage, thus obtaining the initial contour.

7. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 6, characterized in that, An adaptive Region of Interest (ROI) is constructed based on the spatial distribution range of response points within the candidate damage region, specifically including: The initial ROI region is constructed by using the maximum and minimum coordinates of the response points in the horizontal and vertical directions within the damage candidate region as the boundary coordinates of the initial ROI. An adaptive expansion mechanism is introduced based on the initial ROI to obtain an adaptive ROI region; The adaptive expansion mechanism is as follows: taking the center point of the initial ROI region as the center, it expands in both the horizontal and vertical directions, and the expanded boundary does not exceed the width and height range of the entire ultrasound image; the expansion range is determined by multiplying the scale of the initial ROI region by a preset expansion coefficient.

8. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 1, characterized in that, The initial contour is iteratively updated under the joint drive of the composite energy field, including: In each iteration, the external force generated by the composite energy field is projected onto the normal direction of the contour point, retaining only the normal component; The update direction is determined based on the distance from the contour point to the nearest boundary. When the distance from the contour point to the nearest boundary is less than the preset distance threshold and the update direction shows an outward expansion trend, the outward expansion normal component is removed, and only the part that allows shrinkage or tangential adjustment is retained.

9. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 8, characterized in that, The iterative update of the initial contour under the joint drive of the composite energy field also includes: Apply amplitude limits to the displacement of a single update; The centroid of the outline is restored to its initial position after each update. When the average normal displacement of all contour points is less than the preset convergence threshold, the contour is determined to have converged and the iteration stops.

10. The focused ultrasound micro-damage image segmentation method based on the composite energy field-guided Snake model according to claim 8, characterized in that, The initial contour is iteratively updated under the joint drive of the composite energy field, and the process further includes: if the contour area is less than the initial area multiplied by a preset ratio threshold during the iteration process, the process reverts to the previous contour and terminates the iteration.

Images