A method, system and device for three-dimensional modeling and injection path assisted planning of post-thyroid surgery scars

Abstract

The application discloses a kind of thyroid postoperative scar three-dimensional modeling and injection path auxiliary planning method, system and equipment, applied to intelligent auxiliary medical field.It includes: obtaining the color image data and depth point cloud data of patient thyroid postoperative scar, and after feature coding, fusion feature tensor is obtained by fusion;After the fusion feature tensor is fused by multi-scale hollow convolution, spatial detail features are obtained by feature splicing;After the fusion feature tensor is fused by global average pooling, channel features are obtained by channel attention, the channel features are fused with the fusion feature tensor to generate channel context features by channel weighting;The spatial detail features are fused with the channel context features to generate gating map by gating mechanism, and the value of the gating map position node is dynamic fusion gating coefficient;The dynamic fusion gating coefficient is fused with the spatial detail features and the channel context features to obtain enhanced feature tensor.

Description

Technical Field

[0001] This invention relates to the field of intelligent assisted medical care, and more specifically, to a method, system, and device for three-dimensional modeling and injection path planning of post-thyroidectomy scars. Background Technology

[0002] With the increasing prevalence of thyroid surgery, the prevention and treatment of postoperative neck scars has become a key focus and challenge in clinical practice. Scar formation and development is a complex dynamic process involving histological, biomechanical, and immunological factors, exhibiting significant three-dimensional spatial heterogeneity in terms of hardness, thickness, and vascular distribution. In drug injection therapy for scars, doctors typically rely on personal experience to determine injection sites and drug dosages, lacking objective and quantitative planning guidelines. This experience-based treatment model has several limitations: firstly, it is difficult to accurately identify the core proliferative area with the highest hardness and the active marginal area within the scar, potentially leading to undertreatment or overtreatment; secondly, it is impossible to accurately calculate the personalized drug dosage required for different heterogeneous areas, thus affecting efficacy and increasing the risk of side effects such as local tissue atrophy and depigmentation. Therefore, developing an intelligent diagnostic method that can overcome the physical barriers between organs, collaboratively process 3D / 2D data, and output multidimensional prognostic indicators is particularly important. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art, and to provide a method, system and device for three-dimensional modeling and injection path auxiliary planning of thyroid postoperative scars.

[0004] The first aspect of this invention discloses a dual-channel attention feature enhancement method based on post-thyroidectomy scarring, comprising: S101: Obtain color image data and depth point cloud data of the patient's post-thyroidectomy scar, and fuse them after feature encoding to obtain a fused feature tensor; Further, the fusion of the feature encoding and fusion to obtain the fused feature tensor in S101 specifically involves: adaptively normalizing the color image data and obtaining an enhanced color image through vascular enhancement filtering; filtering the depth point cloud data through key points and then performing 3D projection to obtain a normalized depth map; calculating the enhanced color tensor through texture complexity calculation of the enhanced color image; calculating the normalized depth tensor through set transformation calculation of the normalized depth map; calculating attention weights based on the enhanced color tensor and the normalized depth tensor; and calculating the weighted sum of the enhanced color image and the normalized depth map based on the attention weights to obtain the fused feature tensor.

[0005] S102: The fused feature tensor is concatenated with features after multi-scale dilated convolution to obtain spatial detail features; the fused feature tensor is then processed through global average pooling and channel attention to obtain channel features; the channel features are weighted with the fused feature tensor to generate channel context features; the spatial detail features and the channel context features are used to generate a gating graph through a gating mechanism, where the values ​​of the position nodes in the gating graph are dynamic fusion gating coefficients; the dynamic fusion gating coefficients are weighted with the spatial detail features and the channel context features to obtain an enhanced feature tensor.

[0006] Furthermore, the method also includes S103: convolving the enhanced feature tensor to generate a semantic feature vector, and constructing a hardness prediction map by passing the semantic feature vector through a regression sub-network.

[0007] Optionally, the regression subnetwork is composed of two fully connected layers connected in series.

[0008] Optionally, the step of further extracting features from the semantic feature vector through a regression sub-network to obtain a hardness prediction map specifically involves: inputting the semantic feature vector into the first fully connected layer of the regression sub-network for feature extraction, and then inputting it into the second fully connected layer to construct a hardness prediction map.

[0009] Furthermore, the method also includes S104: diffusing and smoothing the hardness prediction map in each direction to obtain a hardness distribution field; calculating the surface curvature and the degree of geometric change based on the normalized depth map to obtain the average surface curvature and the degree of geometric change; calculating a hardness-curvature comprehensive index based on the hardness distribution field, the average surface curvature, and the degree of geometric change; and identifying core pixels and edge pixels based on the hardness-curvature comprehensive index.

[0010] Optionally, the step of diffusing and smoothing the hardness prediction map in each direction to obtain the hardness distribution field specifically involves: constructing a diffusion coefficient function, inputting the hardness prediction map into the diffusion coefficient function, calculating and obtaining the hardness distribution field through gradient operation.

[0011] Preferably, the specific process of constructing the diffusion coefficient function is as follows: calculating the hardness prediction map. The diffusion coefficient function is obtained by dividing the norm by the control edge sensitivity threshold parameter, squaring the result, and taking the reciprocal.

[0012] Optionally, the step of identifying core pixels and edge pixels based on the hardness-curvature composite index specifically involves: when the hardness-curvature composite index is greater than the high threshold for identifying the core area, the current pixel is determined to be a core pixel; when the hardness-curvature composite index is greater than the low threshold for identifying the edge area and less than the high threshold for identifying the core area, the current pixel is determined to be an edge pixel.

[0013] The second aspect of this invention discloses a method for constructing a post-thyroidectomy scar prediction model, comprising: S201: Obtain color image data, depth point cloud data, true hardness label, and partition label of the patient's post-thyroidectomy scar; S202: Input the color image data and the depth point cloud data into the dual-channel attention feature enhancement method based on post-thyroidectomy scar for processing to obtain a hardness prediction map; S203: Construct a total loss function based on the hardness prediction map and the actual hardness label, optimize the total loss function, and obtain a thyroid postoperative scar prediction model.

[0014] Furthermore, the total loss function is obtained by weighting the supervised loss term and the regional consistency regularization term.

[0015] Preferably, the formula for calculating the total loss function is as follows: , in, Represents the total loss function. Indicates monitoring loss items, Represents a region consistency regularization term The weighting coefficients.

[0016] Optionally, the supervised loss term is calculated by taking the mean of the absolute values ​​of the differences between all pixels in the hardness prediction map and the true hardness label.

[0017] Preferably, the formula for calculating the supervision loss term is as follows: , in, A collection of pixels representing the scar area. Indicates the location is The pixel values ​​of the hardness prediction map at that location. Indicates the location at The actual hardness label value.

[0018] Optionally, the calculation process of the region consistency regularization term is as follows: calculate the arithmetic mean of the predicted hardness values ​​of all pixels in the pixel set, and calculate the average of the squares of the differences between all pixels in the hardness prediction map and the arithmetic mean to obtain the region consistency regularization term.

[0019] Optionally, the pixel set specifically refers to the union of core pixels and edge pixels.

[0020] Preferably, the calculation formula for the region consistency regularization term is as follows: , in, Represents the set of core pixels. The set representing the edge pixels. Represents pixels Selected from core pixels or edge pixels, Represents a set of pixels. Indicates the set of pixels The arithmetic mean of the predicted hardness values ​​for all pixels within the range.

[0021] The third aspect of this invention discloses a method for predicting post-thyroidectomy scars and precise injection planning, comprising: S301: Acquire color image data and depth point cloud data of the patient's post-thyroidectomy scar; S302: Input the color image data and depth point cloud data into the thyroid postoperative scar prediction model for processing to obtain the patient's thyroid postoperative scar area; perform three-dimensional reconstruction based on the thyroid postoperative scar area to obtain a three-dimensional scar model; S303: Injection planning is performed based on the three-dimensional scar model to obtain the injection planning result.

[0022] A fourth aspect of the present invention discloses a computer device, the device comprising: a memory and a processor; The memory is used to store program instructions; The processor is used to call program instructions, and when the program instructions are executed, the above method steps are implemented.

[0023] A third aspect of the present invention discloses a computer program product, including a computer program that is implemented by a processor using the above-described method steps.

[0024] A fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method steps.

[0025] Advantages of this invention: Compared with the prior art, the present invention has the following beneficial effects: 1. To address the problems of existing partitioning methods relying on a single hardness threshold, being sensitive to noise, and struggling to accurately distinguish complex heterogeneous structures within scars, this invention proposes a hardness distribution field generation method based on anisotropic diffusion smoothing, combined with an adaptive dual-threshold partitioning method using a hardness-curvature comprehensive index. The generated hardness distribution field can retain important boundary information while suppressing prediction noise. By combining a comprehensive index constructed from geometric features such as curvature with an adaptive threshold, it can robustly and accurately identify the core hardened area and the edge transition area of ​​the scar, solving the problem of inaccurate boundary identification in traditional methods.

[0026] 2. To address the problem that existing hardness prediction models neglect surgical suture priors and struggle to balance local details with global context, leading to prediction biases in key areas, this invention proposes a feature enhancement method based on a dual-pathway gated attention mechanism. This method is further corrected by incorporating clinical priors based on suture tension estimation. On one hand, by processing spatial detail pathways and channel context pathways in parallel and dynamically fusing them using a gating mechanism, the model's ability to capture local features and global semantics is significantly improved. On the other hand, by introducing suture tension prior features, the hardness prediction results better reflect the actual biomechanical state of the scar, with significantly improved prediction accuracy, especially in key areas such as near the suture line. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic flowchart of the method provided in the first aspect of the present invention; Figure 2 This is a system execution architecture diagram provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the fusion feature tensor calculation process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the enhanced feature tensor calculation process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the core area and edge area identification process provided in an embodiment of the present invention. Figure 6 This is a schematic flowchart of a method provided in the second aspect of the present invention; Figure 7 This is a schematic flowchart of a method provided in the third aspect of an embodiment of the present invention.

[0029] Figure 8This is a schematic diagram of a device provided for an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0031] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be performed in the order they appear herein, or may be performed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel.

[0032] 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.

[0033] Figure 1 This is a schematic flowchart of a method provided in the first aspect of the present invention, specifically including the following steps: S101: Obtain color image data and depth point cloud data of the patient's post-thyroidectomy scar, and fuse them after feature encoding to obtain a fused feature tensor; In one specific embodiment, the invention further includes multimodal data acquisition and training dataset construction. Specifically, multimodal data of neck scars from patients after thyroid surgery are systematically acquired. Data acquisition is performed in a standardized clinical environment, using a calibrated RGB-D camera system to simultaneously acquire color images and depth point cloud data of the patient's neck scar area. To ensure data quality and alignment accuracy, the relative position of the camera and the patient's head and neck is fixed during acquisition, and the patient is required to maintain a specific posture to fully expose the scar area, while the ambient lighting is controlled within a stable range. For each patient, in addition to acquiring multimodal data of the current scar state, if available, early postoperative scar imaging data (such as at suture removal) should also be collected for subsequent analysis to estimate the dominant direction of the original suture line. After data collection, the raw data was rigorously annotated. The annotation included: 1) Experienced clinicians or using ultrasound elastography equipment to annotate the scar region pixel-by-pixel or region-by-region on the color images, generating a realistic hardness map. This value served as the supervised target for model prediction; 2) Clinicians precisely delineated the scar contours on the color images, generating a binarized scar region segmentation mask to limit the effective range of loss calculation during training; 3) For early images that could be used to estimate suture priors, clinicians manually labeled several representative suture lines to calculate their average dominant direction. All annotated color images, depth point clouds and their corresponding hardness annotation maps, segmentation masks, and suture line direction information together constituted the training dataset.

[0034] In one specific embodiment, the color texture and depth point cloud data of post-thyroidectomy scars exhibit multimodal and three-dimensional heterogeneity. Color textures contain information about color and vascular distribution, but are easily affected by ambient lighting. Depth point clouds provide three-dimensional geometric information, but are typically subject to data sparsity and noise interference. Conventional techniques process the two modalities independently and then simply stitch them together, easily ignoring semantic relationships and physical scale differences between modalities, resulting in insufficient feature representation in the fusion. Therefore, this invention employs an adaptive normalization and attention-based multimodal fusion method to unify the numerical scale of the data and enhance complementary information between modalities.

[0035] Furthermore, the fused feature tensor obtained after feature encoding in S101 is specifically as follows (see...). Figure 3The process involves: adaptively normalizing the color image data and applying a blood vessel enhancement filter to obtain an enhanced color image; filtering the depth point cloud data through key points and then performing a 3D projection to obtain a normalized depth map; calculating the enhanced color tensor using texture complexity of the enhanced color image; performing set transformation calculation on the normalized depth map to obtain a normalized depth tensor; calculating attention weights based on the enhanced color tensor and the normalized depth tensor; and calculating a weighted sum of the enhanced color image and the normalized depth map based on the attention weights to obtain a fusion feature tensor.

[0036] In a specific embodiment, to address the problem of uneven illumination and color distortion in neck scar color images caused by inconsistent shooting conditions, this invention first converts the neck scar color image from the RGB color space to the LAB color space to decouple luminance and color information. Blood vessel enhancement is then restricted to the luminance channel. Adaptive normalization based on local statistics is performed on the luminance channel, and a blood vessel enhancement filter is superimposed to obtain an enhanced neck scar color image, thereby improving the discriminative power of texture features. This is represented as follows:

[0037] In the formula, Enhanced color images of neck scars In the Channel, pixel position The pixel values ​​at the location are adjusted by adaptive normalization to eliminate uneven lighting and by blood vessel enhancement filtering to highlight the blood vessel texture structure. This represents the image channel index in the LAB color space. Corresponding brightness channel, Corresponding chroma channels, Corresponding chroma channels; Represents the two-dimensional spatial coordinate index of an image pixel. For row index, For column indexes; Enhance color tensors; This indicates that after converting the enhanced color image of the neck scar to the LAB color space, in the... passage, location The pixel value at that location decouples brightness and color information, making it easier to process the brightness channel separately to eliminate the influence of lighting, while retaining color information; Represented in pixels local neighborhood window centered within, no. The arithmetic mean of the channel pixel values ​​is used to estimate the local illumination level at that location; Represented in pixels local neighborhood window centered within, no. The standard deviation of the channel pixel values ​​is used to estimate the local contrast level at that location; Indicated by position A rectangular window centered on a specific location for local statistical calculations, its size being a preset hyperparameter, for example, a size of [missing information]. Pixel; This represents a very small positive integer, used to ensure that the denominator is not zero. Examples of its values ​​are: ; This represents the vessel enhancement coefficient, an adjustable hyperparameter used to control the contribution of the vessel enhancement filter to the final output. An example value is 0.3. Indicates the location of the vascular enhancement filter. The response value at a certain point is used to highlight the tubular texture structure in the image. Its calculation is based on the response of the multi-directional Gabor filter to the brightness channel, and is expressed as... ; The scale set representing the Gabor filter kernel, for example ; The scale parameter representing the Gabor filter kernel is a set of scales. Elements in; Represents the set of directions of the Gabor filter kernel, for example ; The direction angle of the Gabor filter kernel is represented by a set of directions. Elements in; This represents the luminance channel image of a neck scar after conversion from a color image to the LAB color space. Represents the set of traversal directions All directions in the filter are considered, and the maximum absolute value of the filter response is taken. Indicates direction as Scale is Gabor filter kernel function; This represents the two-dimensional convolution operator.

[0038] In a specific embodiment, to address the problem of sparse and noisy deep point cloud data, this invention first filters key points based on the local curvature of the point cloud to retain significant geometric features. Then, the filtered 3D key points are projected onto a 2D image plane aligned with the color image. Dense, normalized deep point cloud data is generated through aggregation and interpolation, as follows: , In the formula, Represents a normalized depth map At pixel position The pixel value at that location provides a numerical range within... The standardized geometric information within the interval is used to describe the three-dimensional shape and surface geometric features of the scar. Represents the normalized depth tensor; Represents deep point cloud data In position The feature value at a given location is the depth point cloud data generated by projecting 3D keypoints at that location. The original depth value at that location is represented as ; Represents deep point cloud data The minimum depth value in the calculation is used as the lower bound for normalization. Represents deep point cloud data The maximum depth value in the range, used as the upper limit for normalization calculations; Represents all pixels projected onto the image. A set of three-dimensional key points at the location; Representing 3D key points The depth coordinates of that point in the camera coordinate system. Axis coordinates; Representing 3D keypoints, it is a set A point in a given location, having three-dimensional coordinates. .

[0039] In practical implementation, if key points are not projected onto certain pixels during the depth point cloud projection process, then during the generation of depth point cloud data... Then, bilinear interpolation is used to fill in the missing pixels to ensure data density.

[0040] In a specific embodiment, to balance the contributions of color texture information and depth geometric information in the fusion process, this invention employs an adaptive attention mechanism. Based on the texture complexity and geometric variation of each pixel's local region, it dynamically calculates and allocates the fusion feature tensor of the color modality and the depth modality, expressed as: , , In the formula, Represents the fusion feature tensor In the passage, location eigenvalues ​​at that location The value can be 1, 2 or 3, and the enhanced color image is weighted by attention weights to dynamically fuse color texture information; Fusion feature tensor In depth channels, location The feature values ​​at the depth channel are fused feature tensors. The fourth channel uses attention weights to weight the normalized depth point cloud data and dynamically fuse depth geometric information. Indicates the location The attention weights assigned to the color modality have a range of values. The calculation method is expressed as A higher weight indicates that the color texture information at that location is more important. Indicates the location The attention weights assigned to the deep modality are calculated as follows: This ensures that the sum of the weights of the two modes at the same location is 1. This is the texture complexity coefficient, used to adjust the strength of the influence of texture complexity on attention weights; an example value is 1.0. Represents the natural exponential function; Represented in pixels local neighborhood window centered The texture complexity is calculated as follows: That is, the sum of gradient magnitudes is calculated based on the brightness channel of the enhanced image of the neck scar; The gradient vector field representing the brightness channel of the enhanced image of the neck scar is specifically calculated using the Sobel operator; Represents the L2 norm; Represented in pixels local neighborhood window centered The degree of geometric change is calculated as follows: ; Indicates that in local window centered Internal, normalized deep point cloud data The characteristic value is the arithmetic mean; This is the geometric change coefficient, used to adjust the strength of the influence of the degree of geometric change on the attention weight. An example value is 0.8.

[0041] S102: The fused feature tensor is concatenated with features after multi-scale dilated convolution to obtain spatial detail features; the fused feature tensor is then processed through global average pooling and channel attention to obtain channel features; the channel features are weighted with the fused feature tensor to generate channel context features; the spatial detail features and the channel context features are used to generate a gating graph through a gating mechanism, where the values ​​of the position nodes in the gating graph are dynamic fusion gating coefficients; the dynamic fusion gating coefficients are weighted with the spatial detail features and the channel context features to obtain an enhanced feature tensor.

[0042] In one embodiment, a spatial attention map is generated using the fused feature tensor. The fused feature tensor is then concatenated using multi-scale dilated convolution to obtain spatial detail features. The spatial attention map and the spatial detail features are fused to obtain a first spatial detail feature. The fused feature tensor is then subjected to global average pooling and channel attention to obtain channel features. The channel features are then weighted with the fused feature tensor to generate channel context features. The first spatial detail features and the channel context features are then used to generate a gating map using a gating mechanism. The values ​​of the position nodes in the gating map are dynamically fused gating coefficients. The dynamically fused gating coefficients are weighted with the spatial detail features and the channel context features to obtain an enhanced feature tensor.

[0043] Furthermore, the method also includes S103: convolving the enhanced feature tensor to generate a semantic feature vector, and constructing a hardness prediction map by passing the semantic feature vector through a regression sub-network.

[0044] Optionally, the regression subnetwork is composed of two fully connected layers connected in series.

[0045] Optionally, the step of further extracting features from the semantic feature vector through a regression sub-network to obtain a hardness prediction map specifically involves: inputting the semantic feature vector into the first fully connected layer of the regression sub-network for feature extraction, and then inputting it into the second fully connected layer to construct a hardness prediction map.

[0046] In one specific embodiment, scar tissue stiffness is a key biomechanical indicator for assessing its maturity and determining drug injection dosage. Direct prediction based on multimodal fusion feature tensors easily overlooks the balance between local details and global context, and struggles to incorporate prior mechanical knowledge from clinical surgical suturing, leading to biased predictions in critical areas. This invention fuses local details and global semantic information through a dual-pathway gated attention mechanism and introduces clinical prior features based on suture tension estimation for correction, thereby generating a high-precision scar stiffness prediction map.

[0047] In a specific embodiment, to balance the extraction of local details and global context information in the fused features, this invention employs a dual-path gating attention module to process the spatial detail path and the channel context path in parallel, and generates an enhanced feature tensor by dynamically fusing the outputs of the two paths through a gating unit, as follows: , In the formula, Represents the augmented feature tensor In the Each feature channel, pixel position The feature values ​​at the location, which integrate spatial details and channel context information, have stronger representation capabilities and can more robustly encode the local texture, geometric structure and global semantic information of scars. This represents the feature channel index, with a value range of [value range missing]. ; This represents the total number of feature channels and is a hyperparameter defined in the network structure. An example value is 128. Indicates spatial location The dynamic fusion gate value at the location has a range of 100%. It is used to adaptively balance the contributions of spatial detail pathways and channel context pathways, and is generated by convolution and activation function operations on the intermediate features of the two pathways. Representation Spatial Detail Tensor In the Each channel, location The spatial detail features are output by the spatial detail pathway, which captures the correlation features between local texture and geometry through multi-scale dilated convolution and uses a spatial attention mechanism to focus on salient regions. Represents the channel context feature tensor In the Each channel, location The channel features are output by the channel context path, which captures the image-level semantic context through global average pooling and uses a channel attention mechanism to recalibrate the importance of each feature channel.

[0048] In a specific embodiment, the spatial detail path processing method is represented as follows: ; Furthermore, the processing method for the channel context path is represented as follows: ; In the formula, The spatial attention map is generated from channel pooling features. Indicates a channel splicing operation; Indicates the use of multi-scale features for fusing Convolution operation; Represents the set of void ratios, for example ; This represents the porosity of dilated convolution; This indicates the corresponding void ratio. The convolutional kernel weights are trainable parameters used to extract spatial features of the fused feature tensor under different receptive fields in order to capture multi-scale local contextual information. This represents the convolution operation; This represents element-wise multiplication. The channel attention vector is generated from the global average pooling features. This indicates a channel weighted multiplication operation.

[0049] In one specific embodiment, the spatial detail pathway focuses on encoding local texture and geometric details; specifically, firstly, it uses pathways with different porosities. Multiple parallel convolutional kernel pairs Perform dilated convolution to generate multiple sets of spatial feature maps. Then, concatenate these multi-scale feature maps along the channel dimension and use... Convolutional layers fuse and reduce the dimensionality of the concatenated features to obtain intermediate spatial features. Simultaneously, from... generate Finally, the integrated intermediate space features are combined with... Perform element-wise multiplication to obtain the output. .

[0050] In one specific embodiment, the channel context path focuses on encoding global semantic context information; specifically, firstly, for Global average pooling is performed along the spatial dimension to obtain the channel description vector. This vector is then processed through a two-layer fully connected network and generated using the Sigmoid activation function. Finally, and Perform channel-weighted multiplication, that is, multiply by channel weight. Each channel is rescaled to obtain the output. .

[0051] In one specific embodiment, the intermediate feature maps of the spatial detail path and the channel context path are concatenated along the channel dimension, and then... Convolutional layers and the Sigmoid activation function generate a range of values. Gating diagram , its first Line 1 The elements of a column are spatial locations. Dynamic fusion gating coefficient at the location .

[0052] In a specific embodiment, by fusing the feature tensor Global max pooling and global average pooling are performed along the channel dimension, and then the two resulting spatial feature maps are concatenated, followed by... Convolutional layers and the sigmoid activation function generate spatial attention maps. .

[0053] In a specific embodiment, the fused feature tensor Global average pooling is performed along the spatial dimension to obtain channel description vectors. Then, channel attention vectors are generated through a two-layer fully connected network and a sigmoid activation function. .

[0054] Furthermore, the method also includes S104: diffusing and smoothing the hardness prediction map in each direction to obtain a hardness distribution field; calculating the surface curvature and the degree of geometric change based on the normalized depth map to obtain the average surface curvature and the degree of geometric change; calculating a hardness-curvature comprehensive index based on the hardness distribution field, the average surface curvature, and the degree of geometric change; and identifying core pixels and edge pixels based on the hardness-curvature comprehensive index.

[0055] Optionally, the step of diffusing and smoothing the hardness prediction map in each direction to obtain the hardness distribution field specifically involves: constructing a diffusion coefficient function, inputting the hardness prediction map into the diffusion coefficient function, calculating and obtaining the hardness distribution field through gradient operation.

[0056] Preferably, the specific process of constructing the diffusion coefficient function is as follows: calculating the hardness prediction map. The diffusion coefficient function is obtained by dividing the norm by the control edge sensitivity threshold parameter, squaring the result, and taking the reciprocal.

[0057] Optionally, the step of identifying core pixels and edge pixels based on the hardness-curvature composite index specifically involves: when the hardness-curvature composite index is greater than the high threshold for identifying the core area, the current pixel is determined to be a core pixel; when the hardness-curvature composite index is greater than the low threshold for identifying the edge area and less than the high threshold for identifying the core area, the current pixel is determined to be an edge pixel.

[0058] In one specific embodiment, see Figure 4 To address the issue that network predictions may overlook local mechanical environment differences caused by surgical sutures, this invention combines prior features of suture tension based on scar orientation and surface curvature estimation, and adaptively weights and fuses them with deep network features. A hardness prediction map is generated through a regression sub-network, as shown below: , In the formula, This indicates the generated hardness prediction map. In position The predicted hardness value at the specified location is normalized to a range using the Sigmoid function. The range indicates that the larger the value, the higher the hardness of the scar tissue at that location; This represents a regression subnetwork, specifically a regression subnetwork consisting of two fully connected layers, used to map the concatenated high-dimensional feature vectors into scalar hardness values; This represents the feature vector concatenation operation; Indicates the location From the enhanced feature tensor Further extraction of deep high-level semantic feature vectors, specifically through the enhancement feature tensor Apply an additional The convolutional layer is used to encode the abstract, high-level characteristics of the scar tissue at that location. Indicates the location The estimated prior eigenvalue of suture tension is a scalar quantity used to reflect the potential increase in local stiffness caused by surgical suture tightening. Its calculation comprehensively considers local geometric information, and the calculation method is expressed as follows: ; This represents the adaptive weights of prior features, which are scalars with a range of . Used based on feature vectors Dynamic adjustment of suture tension prior eigenvalues The strength of the introduction is to avoid excessive interference from prior features when the network can already make accurate predictions; This represents the direction sensitivity coefficient, with an example value of 0.5, which controls the intensity of the influence of the difference between the local main direction of the scar and the predicted dominant direction of the original suture on tension estimation. The principal orientation angle of the scar is obtained by principal component analysis of the gradient field of the brightness and depth images. To estimate the dominant direction of the original suture, the doctor can manually mark several representative sutures on preoperative / postoperative images and take the average direction, or the dominant direction can be automatically estimated from early (such as immediately after surgery) scar images using a linear structure detection algorithm. The average surface curvature is calculated from normalized depth point cloud data. It is considered as a two-dimensional surface, and the average curvature is calculated using the discrete approximation method. This indicates the operation of retrieving the maximum value.

[0059] In one specific embodiment, the local principal direction angle of the scar By pixels Within the neighborhood of the image, calculate the brightness channel of the enhanced image of the neck scar. and normalized deep point cloud data The gradient field will be in the pixel Within the neighborhood of the image, calculate the brightness channel of the enhanced image of the neck scar. and normalized deep point cloud data The gradient fields are calculated separately, their structural tensors are summed, and then principal component analysis is performed on the summed matrix. The direction of the first principal component is taken as the local principal direction angle of the scar. After superimposing the two gradient fields and performing principal component analysis, the direction of the first principal component is taken as the local principal direction angle of the scar. .

[0060] In one specific embodiment, the hardness prediction map exhibits noise and inhomogeneity, and a single hardness threshold is insufficient to accurately distinguish the heterogeneous structure within the scar. Conventional partitioning methods are susceptible to local fluctuations, easily leading to inaccurate boundary identification between the core hardened zone and the edge transition zone. This invention generates a smooth hardness distribution field through anisotropic diffusion to preserve macroscopic trends and boundaries, and constructs a comprehensive index in conjunction with geometric curvature information. An adaptive dual-threshold method is then employed to achieve robust identification of the core and edge zones.

[0061] In one specific embodiment, to preserve the boundary information of the core region while suppressing noise, this invention employs an anisotropic diffusion equation guided by hardness gradient to iteratively smooth the initial hardness prediction map, generating a continuous hardness distribution scalar field. (See...) Figure 5 : is represented as: , In the formula, Indicates the number of diffusion iterations. At time, position The hardness distribution field value at the location, with initial conditions as follows: After being preset The final field is obtained after the second iteration. This field is better able to reflect the macroscopic distribution trend of scar hardness; Indicates the diffusion time The partial derivatives are used to describe the hardness distribution field during the diffusion iteration process. Rate of change over time; The step index representing the diffusion iteration, i.e., the virtual time parameter, is a discrete integer variable, from... The increment is used to control the extent to which the anisotropic diffusion smoothing process proceeds; This represents the divergence operator, specifically the operator for calculating the divergence of a vector field; This represents the gradient operator; The diffusion coefficient function is a key function controlling diffusion behavior, and its input is the initial hardness prediction map. In position gradient magnitude at This function suppresses diffusion in high-gradient regions to preserve edges and promotes diffusion in low-gradient regions to smooth noise. The calculation method is expressed as follows: ; The term, acting as an edge-preserving term, approaches zero when the gradient magnitude is large (at the edge), reducing the diffusion coefficient and thus suppressing diffusion in that region, protecting the edge. The parameter controlling the edge sensitivity threshold... The example value is 0.1; As a smoothing term for a uniform region, when the gradient magnitude is small (in a flat region), the value of this term is close to... This ensures that the diffusion coefficient remains at a certain value, thereby promoting diffusion in the region and smoothing noise. The parameter controlling the smoothing intensity is... The example value is 0.05, which controls the smoothing range. The example value is 0.05.

[0062] In one specific embodiment, to accurately identify the high-hardness core region and the medium-hardness edge transition region, this invention constructs a comprehensive quantitative index by combining the smoothed hardness distribution field and surface curvature information. Based on the global statistical characteristics of this quantitative index, high and low thresholds are automatically calculated to complete the partitioning, thereby identifying the core region pixel set and the edge region pixel set, represented as: , , In the formula, This represents the set of pixels in the identified core area, which typically corresponds to the scar tissue with the most significant proliferation and highest hardness. This represents the set of pixels in the identified edge region, which typically corresponds to transitional scar tissue of medium hardness and in an active state. Indicates the location The hardness-curvature composite index at a location is a scalar quantity used to quantify the probability that the location belongs to the core region of abnormal proliferation. A higher value indicates a higher probability of belonging to the core region. The calculation method is expressed as follows: ; This represents the weighting coefficient of the hardness distribution field in the comprehensive index, with an example value of 0.6. Represents the hardness distribution field The value after max-min normalization within the scar area; This represents the weighting coefficient of the average surface curvature in the comprehensive index, with an example value of 0.3. Indicates the average curvature of the surface The value after max-min normalization within the scar area; This represents the weighting coefficient of the degree of geometric change in the comprehensive index, with an example value of 0.1. Indicates the degree of geometric change The value after max-min normalization within the scar area; To identify the high threshold of the core region, it is represented as ; The low threshold for identifying edge regions is represented as ; Chart showing the combined hardness-curvature index The arithmetic mean of all pixel values ​​within the scar area; This represents the multiplier used to calculate the high threshold of the core area; an example value is 1.5. Chart showing the combined hardness-curvature index Standard deviation of all pixel values ​​within the scar area; This represents the multiplier used to calculate the low threshold of the edge region; an example value is 0.5.

[0063] Figure 6 This invention discloses a method for constructing a post-thyroidectomy scar prediction model, comprising: S201: Obtain color image data, depth point cloud data, true hardness label, and partition label of the patient's post-thyroidectomy scar; S202: Input the color image data and the depth point cloud data into the dual-channel attention feature enhancement method based on post-thyroidectomy scar for processing to obtain a hardness prediction map; S203: Construct a total loss function based on the hardness prediction map and the actual hardness label, optimize the total loss function, and obtain a thyroid postoperative scar prediction model.

[0064] Furthermore, the total loss function is obtained by weighting the supervised loss term and the regional consistency regularization term.

[0065] Preferably, the formula for calculating the total loss function is as follows: , in, Represents the total loss function. Indicates monitoring loss items, Represents a region consistency regularization term The weighting coefficients.

[0066] Optionally, the supervised loss term is calculated by taking the mean of the absolute values ​​of the differences between all pixels in the hardness prediction map and the true hardness label.

[0067] Preferably, the formula for calculating the supervision loss term is as follows: , in, A collection of pixels representing the scar area. Indicates the location is The pixel values ​​of the hardness prediction map at that location. Indicates the location at The actual hardness label value.

[0068] Optionally, the calculation process of the region consistency regularization term is as follows: calculate the arithmetic mean of the predicted hardness values ​​of all pixels in the pixel set, and calculate the average of the squares of the differences between all pixels in the hardness prediction map and the arithmetic mean to obtain the region consistency regularization term.

[0069] Optionally, the pixel set specifically refers to the union of core pixels and edge pixels.

[0070] Preferably, the calculation formula for the region consistency regularization term is as follows: , in, Represents the set of core pixels. The set representing the edge pixels. Represents pixels Selected from core pixels or edge pixels, Represents a set of pixels. Indicates the set of pixels The arithmetic mean of the predicted hardness values ​​for all pixels within the range.

[0071] In one specific embodiment, to effectively train the neural network and make its predictions conform to clinical pathological patterns, this invention guides model optimization by combining supervised prediction error, intra-regional consistency, and physical priors simulating injection treatment. The specific steps are as follows: 1) Composite loss function The composite loss function directly supervises the accuracy of hardness prediction and encourages pixels divided into the same region to have similar predicted hardness values ​​through a regularization term, thereby improving consistency within the partitions. It is expressed as: , In the formula, It is a composite loss function used to simultaneously supervise the accuracy of hardness prediction and promote the consistency of hardness prediction within the same identification zone; To monitor the loss term, mean absolute error is used as a metric, which has good robustness to outliers. Its calculation method is expressed as follows: , Represents a region consistency regularization term The weighting coefficient is used to balance the supervision loss and consistency constraint, with an example value of 0.1. The region consistency regularization term aims to minimize the variance of predicted hardness for pixels within the same region. Its calculation method is expressed as follows: ; The set of pixels representing the scar region is obtained by semantic segmentation of the original input image (which can be done by a pre-trained segmentation network or an interactive annotation tool) and is used to define the effective area for loss calculation. Indicates the location The actual hardness values ​​at the location are determined by ultrasound elastography or clinical evaluation and used to supervise model training. This represents a partition type index, with values... or ,in Representing the core area, Represents the edge area; Represents the set of pixels in the core region or edge region pixel set One of them, by index Decide; Indicates the set of pixels Within, the arithmetic mean of the predicted hardness values ​​for all pixels.

[0072] 2) Injection simulation regularization loss function To incorporate the clinical experience that "the core area is more sensitive to drug injection" into model optimization, this invention constructs a regularized loss based on simplified physical simulation, encouraging the model to predict a stiffness distribution pattern that better matches treatment expectations, expressed as: , In the formula, To simulate the injection regularization loss function, the clinical prior knowledge that "the core area is more sensitive to drug injection" is incorporated into the model training in the form of a physical heuristic regularization term, guiding the model to predict a hardness distribution pattern that is more in line with treatment expectations. This represents the loss term for maximizing the core region response. It is a negative term for the core region, and its optimization objective is to maximize the average value of the "simulated injection response potential" index within the core region. The calculation method is expressed as follows: ; This represents the loss term that minimizes the response in the edge region. It is a positive term for the edge region, and its optimization objective is to minimize the average value of the same index within the edge region. The calculation method is expressed as follows: ; This represents the scaling factor, used to control the order of magnitude of the loss term; an example value is 0.05. This represents the natural logarithm function.

[0073] It should be noted that in actual clinical practice, the core area is usually highly rigid ( Larger and possibly more prominent Small, leading to The project is large, which aligns with the priori assumption that "the core area has great response potential," and... Smaller indicates that the area is closer to the camera, and in a scene with neck scars, it usually corresponds to a raised area.

[0074] 3) Total Loss Function The total loss function is the final optimization objective of model training. It consists of a weighted sum of the prediction loss and the injection simulation regularization loss, expressed as: , In the formula, This represents the weighting coefficient for the injection simulation regularization loss, with an example value of 0.01.

[0075] In one specific embodiment, the preprocessed and fused multimodal fusion feature tensor of the dataset, along with its corresponding ground truth hardness annotations and scar region masks, are used as training inputs. Model parameters are randomly initialized, and training employs an iterative optimization approach. In each training batch, the forward propagation process sequentially performs dual-path gated attention feature enhancement and hardness prediction map generation based on suture tension priors to obtain a predicted hardness map. Then, scar hardness distribution field generation and core / edge region identification are performed to obtain partitioning results. Next, based on the loss function calculation framework, the prediction results, partitioning results, ground truth annotations, and masks of the current batch are combined to calculate the total loss function. This loss integrates prediction error, region consistency constraints, and injection simulation physical priors. During backpropagation, the gradient descent algorithm is used to calculate the gradient of the total loss relative to all trainable parameters of the model, and these parameters are updated to minimize the loss.

[0076] Optionally, an initial learning rate is set during training, and a learning rate decay strategy is employed to improve convergence stability. Training continues for multiple epochs until the model's performance on independent validation sets (e.g., the mean absolute error between predicted and actual hardness) no longer significantly improves or begins to decline. At this point, an early stopping mechanism is triggered to preserve the best-performing model parameters; or training stops after reaching a preset maximum number of epochs. The trained model can automatically generate high-precision hardness prediction maps and accurately identify the core and edge regions from new multimodal post-thyroidectomy scar data.

[0077] Figure 7 A schematic diagram of a method for predicting and precisely planning post-thyroidectomy scars, provided as an embodiment of the present invention, includes: S301: Acquire color image data and depth point cloud data of the patient's post-thyroidectomy scar; S302: Input the color image data and depth point cloud data into the thyroid postoperative scar prediction model for processing to obtain the patient's thyroid postoperative scar area; perform three-dimensional reconstruction based on the thyroid postoperative scar area to obtain a three-dimensional scar model; S303: Injection planning is performed based on the three-dimensional scar model to obtain the injection planning result.

[0078] In one embodiment, the three-dimensional reconstruction process is as follows: first, the hardness attribute and region category attribute of the post-thyroidectomy scar area are mapped to three-dimensional geometric information, and then three-dimensional reconstruction is performed on the mapped three-dimensional geometric information to obtain a three-dimensional model of the scar.

[0079] In one specific embodiment, after the scar region recognition model is trained, multimodal 3D heterogeneity modeling can be performed on new patient data. First, the original color images and depth point cloud data collected from the new patient are processed strictly according to the preprocessing and fusion process to obtain a standardized multimodal fusion feature tensor. Then, this tensor is input into the trained scar region recognition model for forward inference. The model outputs a hardness prediction map of the patient's scar and a macroscopic hardness distribution field after anisotropic diffusion smoothing, automatically identifying the high-hardness core pixel set and the medium-hardness edge pixel set within the scar. Next, these two-dimensional plane analysis results are fused with three-dimensional geometric information. Using the three-dimensional coordinate information provided by the normalized depth point cloud data, each two-dimensional pixel position is mapped back to three-dimensional space, while simultaneously assigning the three-dimensional point a hardness attribute and a region category attribute defined by the partitioning results. Through surface reconstruction techniques (such as Poisson reconstruction or triangulation), these point sets with rich attributes (including original color texture, three-dimensional coordinates, hardness value, and region category) are constructed into a continuous three-dimensional mesh model. The generated multimodal 3D heterogeneity model not only accurately reproduces the surface 3D morphology of the scar, but more importantly, it intuitively displays the continuous variation gradient and spatial distribution heterogeneity of the hardness inside the scar in 3D space, and clearly marks the specific location and range of the core hardened area and the edge transition area on the 3D curved surface, providing a complete and visualized 3D data foundation for subsequent precise injection planning.

[0080] In one embodiment, the planning identifies the region with the highest hardness and the edge region by analyzing the regional and hardness attributes of the scar 3D model. Recommended injection points are generated based on the region with the highest hardness, and auxiliary injection points are generated based on the edge region. The injection dose is calculated for each injection point based on the hardness and regional attributes. A virtual injection path is generated through the injection points and injection doses to obtain the injection planning result.

[0081] In one specific embodiment, based on the results of multimodal 3D heterogeneity modeling, the system performs personalized and precise injection planning. The planning process first, based on the spatial distribution of the core and edge regions marked in the 3D model, prioritizes generating recommended injection points in the core region with the highest hardness, as these areas are key sites where drugs (such as corticosteroids) need to focus their effects to soften scars. For the edge transition areas, auxiliary injection points are strategically placed to control scar activity and prevent outward proliferation. The 3D spatial location of the injection points is determined by both the zoning results and surface curvature, favoring areas with gentle curvature changes that facilitate needle insertion. Secondly, a personalized drug injection dose is calculated for each injection point. Dosage calculation is a decision-making process that integrates multimodal information, primarily based on a weighted adjustment of the 3D hardness distribution field value at the point's location combined with the region category (core or edge region) to ensure that the drug concentration reaches an effective therapeutic level within the target area while avoiding overdose. The system integrates the 3D coordinates of all injection points, the recommended dose, and considers avoiding major subcutaneous vascular areas identified through color texture analysis in the 3D model to automatically plan a safe and efficient virtual injection path. This optimized approach streamlines the needle insertion sequence, aiming to reduce the number of punctures and maximize coverage of the target treatment area. Once the plan is complete, the system generates a visual report containing a 3D distribution map of injection points, a recommended dosage table for each point, and an animation of the injection path. This assists physicians in preoperative protocol evaluation and intraoperative navigation, enabling personalized and precise drug injection treatment tailored to the 3D structure and mechanical heterogeneity of post-thyroidectomy scars.

[0082] The present invention also discloses a computer device, the device comprising: a memory and a processor, such as... Figure 3 As shown: The memory is used to store program instructions; The processor is used to invoke program instructions, and when the program instructions are executed, the method steps are implemented.

[0083] The present invention provides a computer program product, including a computer program, which is implemented by a processor to implement the method steps.

[0084] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method steps.

[0085] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0087] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0088] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The aforementioned integrated modules can be implemented in hardware or as software functional modules.

[0089] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0090] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0091] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dual-channel attention feature enhancement method based on post-thyroidectomy scarring, characterized in that, include: S101: Obtain color image data and depth point cloud data of the patient's post-thyroidectomy scar, and fuse them after feature encoding to obtain a fused feature tensor; S102: The fused feature tensor is concatenated with features after multi-scale dilated convolution to obtain spatial detail features; the fused feature tensor is then processed through global average pooling and channel attention to obtain channel features; the channel features are weighted with the fused feature tensor to generate channel context features; the spatial detail features and the channel context features are used to generate a gating graph through a gating mechanism, where the values ​​of the position nodes in the gating graph are dynamic fusion gating coefficients; the dynamic fusion gating coefficients are weighted with the spatial detail features and the channel context features to obtain an enhanced feature tensor.

2. The dual-channel attention feature enhancement method based on post-thyroidectomy scar according to claim 1, characterized in that, The specific steps of obtaining the fused feature tensor after feature encoding in S101 are as follows: adaptively normalizing the color image data and obtaining an enhanced color image through blood vessel enhancement filtering; filtering the depth point cloud data through key points and performing 3D projection to obtain a normalized depth map; and calculating the enhanced color tensor through texture complexity of the enhanced color image. The normalized depth map is transformed using set transformation to obtain the normalized depth tensor; The attention weights are calculated based on the enhanced color tensor and the normalized depth tensor. The fused feature tensor is obtained by calculating the weighted sum of the enhanced color image and the normalized depth map based on the attention weights.

3. The dual-channel attention feature enhancement method based on post-thyroidectomy scar according to claim 1, characterized in that, The method further includes S103: convolving the enhanced feature tensor to generate a semantic feature vector, and constructing a hardness prediction map through a regression sub-network using the semantic feature vector; Optionally, the regression subnetwork is composed of two fully connected layers connected in series; Optionally, the step of further extracting features from the semantic feature vector through a regression sub-network to obtain a hardness prediction map specifically involves: inputting the semantic feature vector into the first fully connected layer of the regression sub-network for feature extraction, and then inputting it into the second fully connected layer to construct a hardness prediction map.

4. The dual-channel attention feature enhancement method based on post-thyroidectomy scar according to claim 2, characterized in that, The method further includes S104: diffusing and smoothing the hardness prediction map in each direction to obtain a hardness distribution field; Based on the normalized depth map, the surface curvature and geometric change degree are calculated to obtain the average surface curvature and geometric change degree. The hardness-curvature comprehensive index is calculated by weighting the hardness distribution field, the average surface curvature, and the degree of geometric change. The core pixel and the edge pixel are then identified based on the hardness-curvature comprehensive index. Optionally, the step of diffusing and smoothing the hardness prediction map in each direction to obtain the hardness distribution field specifically involves: constructing a diffusion coefficient function, inputting the hardness prediction map into the diffusion coefficient function, calculating and obtaining the hardness distribution field through gradient operation; Preferably, the specific process of constructing the diffusion coefficient function is as follows: calculating the hardness prediction map. The diffusion coefficient function is obtained by dividing the norm by the control edge sensitivity threshold parameter, squaring the result and taking the reciprocal. Optionally, the step of identifying core pixels and edge pixels based on the hardness-curvature composite index specifically involves: when the hardness-curvature composite index is greater than the high threshold for identifying the core area, the current pixel is determined to be a core pixel; when the hardness-curvature composite index is greater than the low threshold for identifying the edge area and less than the high threshold for identifying the core area, the current pixel is determined to be an edge pixel.

5. A method for constructing a predictive model for post-thyroidectomy scarring, characterized in that, include: S201: Obtain color image data, depth point cloud data, true hardness label, and partition label of the patient's post-thyroidectomy scar; S202: Input the color image data and the depth point cloud data into the dual-channel attention feature enhancement method based on post-thyroidectomy scars according to any one of claims 1-4 for processing to obtain a hardness prediction map; S203: Construct a total loss function based on the hardness prediction map and the actual hardness label, optimize the total loss function, and obtain a thyroid postoperative scar prediction model.

6. The method for constructing a thyroid postoperative scar prediction model according to claim 5, characterized in that, The total loss function is obtained by weighting the supervised loss term and the region consistency regularization term. Preferably, the formula for calculating the total loss function is as follows: ， in, Represents the total loss function. Indicates monitoring loss items, Represents a region consistency regularization term Weighting coefficients; Optionally, the supervised loss term is obtained by calculating the mean of the absolute values ​​of the differences between all pixels of the hardness prediction map and the true hardness label; Preferably, the formula for calculating the supervision loss term is as follows: ， in, A collection of pixels representing the scar area. Indicates the location is The pixel values ​​of the hardness prediction map at that location. Indicates the location at The actual hardness label value is displayed. Optionally, the calculation process of the region consistency regularization term is as follows: calculate the arithmetic mean of the predicted hardness values ​​of all pixels in the pixel set, and calculate the average of the squares of the differences between all pixels in the hardness prediction map and the arithmetic mean to obtain the region consistency regularization term. Optionally, the pixel set specifically refers to the union of core pixels and edge pixels; Preferably, the calculation formula for the region consistency regularization term is as follows: , in, Represents the set of core pixels. The set representing the edge pixels. Represents pixels Selected from core pixels or edge pixels, Represents a set of pixels. Indicates the set of pixels The arithmetic mean of the predicted hardness values ​​for all pixels within the range.

7. A method for predicting post-thyroidectomy scarring and precise injection planning, characterized in that, include: S301: Acquire color image data and depth point cloud data of the patient's post-thyroidectomy scar; S302: Input the color image data and depth point cloud data into the thyroid postoperative scar prediction model described in claim 5 or 6 for processing to obtain the patient's thyroid postoperative scar area; A three-dimensional model of the scar was obtained by three-dimensional reconstruction of the scar area after thyroid surgery. S303: Injection planning is performed based on the three-dimensional scar model to obtain the injection planning result.

8. A computer device, comprising: Memory and processor; The memory is used to store program instructions; The processor is used to invoke program instructions, which, when executed, implement the method according to any one of claims 1-7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1-7.

Images