Three-dimensional image data segmentation method and system

Abstract

The invention provides a three-dimensional image data segmentation method and system, and belongs to the technical field of image processing, and the method comprises the steps: inputting three-dimensional image data into a first segmentation network, and obtaining a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result and the three-dimensional image data in a channel dimension to obtain fusion data; inputting the fused data into the second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used for processing the fused data so as to correct segmentation errors in the preliminary three-dimensional segmentation result. According to the method, the two-stage segmentation network is constructed, the preliminary segmentation result and the original image data are fused, and the second segmentation network is used for optimizing segmentation errors such as false positive, poor continuity and inaccurate boundary, so that the segmentation precision and robustness of the three-dimensional image data are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a three-dimensional image data segmentation method and system. Background Art

[0002] Three-dimensional image data segmentation technology is an important research direction in image processing and is widely used in target region extraction and analysis based on three-dimensional image data. In the prior art, three-dimensional segmentation algorithms are usually developed based on traditional image processing methods or single deep learning models. Although traditional methods such as threshold segmentation, region growing or graph-based segmentation methods have a certain theoretical basis, they often find it difficult to achieve satisfactory results when faced with complex target structures in three-dimensional images. These methods lack robustness to the boundaries and regional information of specific targets and are easily affected by noise interference or other high-similarity regions, resulting in insufficient accuracy of the segmentation results.

[0003] With the development of deep learning technology, 3D segmentation has gradually shifted to models based on convolutional neural networks (CNNs). These methods use multi-layer network structures to extract features of 3D images and can better capture detailed information of local areas. However, a single network structure faces certain challenges in complex 3D image data, such as insufficient understanding of global contextual information, resulting in defects in the continuity and integrity of segmentation results. In addition, due to the interference of non-target areas in image data, existing technologies are difficult to effectively suppress the generation of false positive errors, especially in scenes with complex shapes or similar signals, the segmentation accuracy and boundary accuracy are further reduced.

[0004] Therefore, how to improve the accuracy of three-dimensional image data segmentation has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The present invention provides a three-dimensional image data segmentation method, system, electronic device and storage medium, which are used to solve the defects in the prior art and improve the accuracy of three-dimensional image data segmentation.

[0006] The present invention provides a three-dimensional image data segmentation method, comprising the following steps: Inputting the three-dimensional image data into the first segmentation network to obtain a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result with the three-dimensional image data in the channel dimension to obtain fused data; The fused data is input into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

[0007] According to a three-dimensional image data segmentation method provided by the present invention, the second segmentation network is trained by the following method: Determining a preset segmentation label of the three-dimensional image data sample; Performing a false positive simulation operation on the three-dimensional image data sample to obtain a first damaged label of the three-dimensional image data sample; Performing a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample; Performing a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample; The three-dimensional image data samples and the actual damage labels are spliced ​​in the channel dimension, input into the initial model for training, and the preset segmentation labels are used as training targets to train and obtain the second segmentation network.

[0008] According to a 3D image data segmentation method provided by the present invention, performing a false positive simulation operation on the 3D image data sample to obtain a first damaged label of the 3D image data sample specifically includes: Determining a maximum voxel value of the three-dimensional image data sample; calculating a screening threshold based on the maximum voxel value; Based on the screening threshold, all voxel values ​​of the three-dimensional image data sample are binarized to obtain the first damage label; wherein the first damage label includes multiple voxel labels, when the voxel value is less than the screening threshold, the corresponding voxel label is set to 0, and when the voxel value is not less than the screening threshold, the corresponding voxel label is set to 1.

[0009] According to a three-dimensional image data segmentation method provided by the present invention, performing a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample specifically includes: Identify various morphological operations; Use a 3D structure element of a preset size as the structure element for morphological operations; A morphological operation is randomly selected, and the randomly selected morphological operation is performed on the first damaged label according to the structural element to obtain the second damaged label.

[0010] According to a three-dimensional image data segmentation method provided by the present invention, performing a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample specifically includes: Set the random masking ratio k∈[0,1]; According to the random masking ratio, a random mask having the same shape as the second damaged label is generated, wherein: the probability of setting the value of each position in the random mask to 1 is k, and the probability of setting the value to 0 is 1-k; The random mask is multiplied element-by-element by the second damage label to obtain the actual damage label.

[0011] According to a 3D image data segmentation method provided by the present invention, the first segmentation network adopts an encoder-decoder structure; the 3D image data is input into the first segmentation network to obtain a preliminary 3D segmentation result, specifically comprising: In the encoder, a dual-branch network is used to extract features from the three-dimensional image data, and the dual-branch network includes a convolutional neural network branch and a Transformer branch; At different scales of the encoder, the three-dimensional image data is encoded and feature extracted by a convolutional neural network branch and a Transformer branch respectively, to obtain a first feature map at different scales and a second feature map at different scales; The first feature map and the second feature map at the same scale are fused to obtain multiple fused features; In the decoder, all the fused features are decoded by convolution and upsampling operations, and a skip connection is established between the decoder and the corresponding scales of the encoder; The preliminary three-dimensional segmentation result is obtained through the output of the decoder.

[0012] According to a three-dimensional image data segmentation method provided by the present invention, the fusion data is input into a second segmentation network to obtain a final three-dimensional segmentation result, which specifically includes: Inputting the fused data into a second segmentation network, correcting segmentation errors in the preliminary three-dimensional segmentation result in the second segmentation network, including eliminating false positive areas in the preliminary three-dimensional segmentation result, repairing broken parts in the preliminary three-dimensional segmentation result caused by discontinuity, and refining the boundaries of the preliminary three-dimensional segmentation result; Output the corrected final 3D segmentation result.

[0013] The present invention also provides a three-dimensional image data segmentation system, comprising the following modules: A processing module, used for inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; A splicing module, used for splicing the preliminary 3D segmentation result with the 3D image data in a channel dimension to obtain fused data; The processing module is also used to input the fused data into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the three-dimensional image data segmentation methods described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the three-dimensional image data segmentation method as described above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any one of the three-dimensional image data segmentation methods described above.

[0017] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: By inputting the 3D image data into the first segmentation network, the dual-branch structure of the first segmentation network is used to extract local features and global context information respectively, thereby generating a preliminary 3D segmentation result. The preliminary segmentation result provides the basic outline and semantic information of the target area, laying the foundation for subsequent optimization. However, due to the complexity of 3D image data, such as background noise and subtle features of the target structure, only preliminary segmentation may lead to many false positives, discontinuous target areas, blurred boundaries and other problems, which require further optimization. By splicing the preliminary 3D segmentation results with the original 3D image data in the channel dimension to generate fused data, the semantic information of the segmentation results is organically combined with the detailed features of the image data, enriching the feature expression of the network input. The fused data can not only provide preliminary positioning information of the target area, but also retain the texture details and contrast information in the image, providing more comprehensive context support for subsequent processing. This design effectively makes up for the problem of insufficient accuracy caused by the single feature in the preliminary segmentation. By inputting the fused data into the second segmentation network, the encoder-decoder structure is used to extract and optimize the fused data in depth, thereby correcting the segmentation errors in the preliminary 3D segmentation results. The second segmentation network can gradually correct the errors in the initial segmentation results by specifically processing false positive areas, discontinuous target structures and fuzzy boundary problems, significantly improving the overall quality of segmentation. False positive areas are effectively eliminated, broken parts of target structures are repaired, and the clarity and fit of boundaries are significantly enhanced. Through the above design scheme, the accuracy of the final generated 3D segmentation results is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 This is one of the flow charts of the three-dimensional image data segmentation method provided by the present invention.

[0020] Figure 2 This is the second flow chart of the three-dimensional image data segmentation method provided by the present invention.

[0021] Figure 3 This is the third flow chart of the three-dimensional image data segmentation method provided by the present invention.

[0022] Figure 4 This is the fourth flow chart of the three-dimensional image data segmentation method provided by the present invention.

[0023] Figure 5 This is the fifth flow chart of the three-dimensional image data segmentation method provided by the present invention.

[0024] Figure 6 It is a structural schematic diagram of the three-dimensional image data segmentation system provided by the present invention.

[0025] Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] It should be noted that, in the description of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the system or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0028] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0029] Combine the following Figure 1-Figure 7 The present invention describes a three-dimensional image data segmentation method, system, electronic device and storage medium.

[0030] Figure 1 FIG. 1 is one of the flow charts of the three-dimensional image data segmentation method provided by the present invention. Figure 1 As shown, including but not limited to the following steps: Step 11: Input the 3D image data into the first segmentation network to obtain a preliminary 3D segmentation result.

[0031] Specifically, in this solution, 3D image data covers a wide range of application scenarios and data types, including medical image data, industrial inspection image data, autonomous driving perception image data, and other types of data. Due to the complexity of imaging targets, limitations of imaging conditions, and deficiencies in existing segmentation methods, 3D image data often face common problems such as a large number of false positive segmentation errors, poor continuity, and inaccurate boundaries during the segmentation process, as shown below: In medical imaging data, 3D imaging data is used for 3D structure segmentation of cells or tissues. However, due to the high similarity of signal intensity between cells or tissues, it is easy to misjudge non-target areas as target areas, resulting in false positive errors. At the same time, the continuity of cell segmentation is poor, especially in areas with complex morphology or weak signals. In addition, the fuzzy boundary makes it difficult for the segmentation results to match the actual morphology.

[0032] In industrial inspection image data, industrial CT or non-destructive testing is widely used to detect defects inside materials (such as pores, cracks or inclusions). Due to the complexity of the material structure and noise interference, non-defective areas are easily misjudged as defects during the segmentation process, resulting in a large number of false positive errors. In addition, for targets with slender shapes such as cracks, existing methods are difficult to ensure the continuity of their segmentation, resulting in target breakage or incompleteness. The boundary of the defective area is often difficult to accurately fit the actual contour due to the weak signal gradient.

[0033] In autonomous driving perception image data, object segmentation based on LiDAR point cloud or RGB-D image is a key task. Due to the sparsity of point cloud and environmental interference (such as rainy and foggy weather or dynamic objects), non-target areas (such as ground noise or reflected signals) are easily misclassified as target objects, resulting in false positive errors. At the same time, the continuity of segmentation performs poorly when dealing with complex three-dimensional structures (such as pedestrians, vehicles or trees), and some targets are prone to fragmentation. In addition, due to the limited resolution of point cloud or depth map, the boundaries of target objects are difficult to clearly segment.

[0034] Therefore, based on the problems of many false positive segmentation errors, poor continuity, and imprecise boundaries in the above-mentioned three-dimensional image data, when processing three-dimensional image data, the task of the first segmentation network in this solution is to perform preliminary segmentation of the input data to extract the main features of the target area, providing a basis for subsequent optimized segmentation. However, three-dimensional image data usually has a complex spatial structure and diverse intensity distribution. It is difficult to effectively capture local details and global relationships in the image by relying solely on traditional single methods, which may lead to problems such as blurred segmentation boundaries and discontinuous structures. To solve these problems, the first segmentation network is designed as a dual-branch structure that integrates convolutional neural networks (CNNs) and Transformers, combining the advantages of both to perform comprehensive feature extraction and processing of three-dimensional image data.

[0035] In a possible implementation, step 11 specifically includes the following steps: In the encoder, a dual-branch network is used to extract features from 3D image data. The dual-branch network includes a convolutional neural network branch and a Transformer branch. At different scales of the encoder, the three-dimensional image data is encoded and feature extracted by the convolutional neural network branch and the Transformer branch, respectively, to obtain the first feature map at different scales and the second feature map at different scales; The first feature map and the second feature map at the same scale are fused to obtain multiple fused features; In the decoder, all fused features are decoded through convolution and upsampling operations, and skip connections are established between the corresponding scales of the decoder and the encoder; Through the output of the decoder, a preliminary three-dimensional segmentation result is obtained.

[0036] In the segmentation and processing method of three-dimensional image data provided in the above embodiment, the core task of the first segmentation network is to encode and decode the input three-dimensional image data, extract the key features of the target area and generate a preliminary three-dimensional segmentation result. In order to make full use of the complementarity of local and global information in three-dimensional image data, the encoder adopts a dual-branch network structure, which includes a convolutional neural network branch and a Transformer branch. The convolutional neural network branch can efficiently extract detail features such as edges and textures in the image through the local receptive field, while the Transformer branch captures long-distance spatial dependencies through the self-attention mechanism and models the global context information of the image. Through this design, the dual-branch network can realize the joint modeling of local details and global features, making up for the deficiency that traditional methods are difficult to take into account these two types of information.

[0037] At different scales of the encoder, the 3D image data is first input into the convolutional neural network branch and the Transformer branch for encoding, generating the first feature map and the second feature map respectively. The multi-scale encoding method can gradually extract the low-level to high-level features of the image and capture the significant features of the target area at different resolutions. At the same time, this multi-scale feature extraction method helps to improve the network's ability to express complex target shapes, so that it can maintain the integrity of the global structure while capturing boundary details.

[0038] To further enhance the collaborative expression of local and global information, the first feature map and the second feature map at the same scale are integrated through the feature fusion module. The fusion method can adopt the strategy of element-by-element addition or channel splicing to combine the two features into a unified fusion feature. The fused features contain both the boundary details provided by the convolution branch and the contextual relationship provided by the Transformer branch, so that the network can use richer semantic information for reconstruction in the subsequent decoding stage.

[0039] After entering the decoder, the fused features are gradually restored to their original resolution through convolution and upsampling operations. Skip connections are introduced in the decoding stage to directly pass high-resolution features of the corresponding scale in the encoder to the decoder. The design of skip connections can effectively alleviate the problem of resolution loss caused by multi-layer downsampling and enhance the network's ability to restore image details. At the same time, this design can also improve the accuracy of the segmentation network for boundary areas and reduce the blur or discontinuity of boundaries.

[0040] The decoder finally outputs a feature map that has been decoded and reconstructed layer by layer to generate a preliminary 3D segmentation result. This segmentation result can show high accuracy in terms of boundary clarity, overall continuity, and detail integrity of the target area through multi-scale feature extraction and fusion in the encoding stage, as well as resolution recovery and detail enhancement in the decoding stage. This design not only improves the quality of the preliminary segmentation results, but also provides reliable input for the subsequent optimization of false positive errors, imprecise boundaries, and other issues through the second segmentation network.

[0041] Step 12: Combine the preliminary 3D segmentation result and the 3D image data in the channel dimension to obtain fused data.

[0042] In the task of 3D image data segmentation, although the preliminary 3D segmentation results provide basic contour information of the target area, there are still many false positive segmentation errors, blurred boundaries and poor continuity. These problems are often caused by complex image backgrounds and subtle features of the target structure, which are difficult to solve by preliminary segmentation alone. Therefore, in order to improve the segmentation effect, the preliminary 3D segmentation results are fused with the original 3D image data, which can simultaneously utilize the semantic information in the segmentation results and the detailed features in the original data, further enhancing the network's ability to express the target area.

[0043] Specifically, after aligning the preliminary 3D segmentation results with the 3D image data, they are spliced ​​according to the channel dimension to generate fused data with multiple channels. During the fusion process, the preliminary segmentation results serve as additional semantic information to guide the network to focus on the location and basic outline of the target area, while the original 3D image data provides high-resolution details and texture information of the target area. Through this splicing operation, the fused data can provide comprehensive input for the second segmentation network, allowing the network to rely on both semantic information and image details for further optimization.

[0044] Step 13: Input the fused data into the second segmentation network to obtain the final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct the segmentation errors in the preliminary three-dimensional segmentation result.

[0045] In the process of segmenting 3D image data, although the preliminary 3D segmentation results contain the general outline information of the target area, they still cannot directly meet the requirements of accurate segmentation due to many false positive segmentation errors, discontinuous structures, and imprecise boundaries. In order to further optimize the segmentation results, inputting the fused data into the second segmentation network is a key step. The second segmentation network is designed to correct the segmentation errors in the preliminary 3D segmentation results. Through deep feature extraction and detail optimization, it eliminates the wrongly segmented areas, enhances the continuity and boundary accuracy of the segmentation results, and thus generates a more accurate final 3D segmentation result.

[0046] The following embodiment is a specific training method of the second segmentation network.

[0047] In one possible implementation, refer to Figure 2 , Figure 2 FIG. 2 is a flow chart of the three-dimensional image data segmentation method provided by the present invention. Figure 2 As shown, the method for training the second segmentation network specifically includes steps 21-25: Step 21: Determine a preset segmentation label of the 3D image data sample.

[0048] When training the second segmentation network, determining the preset segmentation labels of the 3D image data samples is an essential basic step. The preset segmentation labels are the actual segmentation results of the 3D image data samples, which are used to guide the network to learn how to identify the target area and correct segmentation errors during training.

[0049] Specifically, the preset segmentation labels are usually derived from manual annotation or high-precision reference methods. To ensure the reliability of the labels, each voxel of the 3D image data is assigned a category label to indicate whether the voxel belongs to the target area. For complex 3D image data samples, the preset segmentation labels need to fully reflect the boundary information, spatial structure, and distinguishing characteristics of the target area from the background. These labels are not only the basis for network training, but also the benchmark for subsequent damage operations and network optimization.

[0050] By establishing a correspondence between the preset segmentation labels and the 3D image data samples, the voxel category information of each data sample is clearly visible. During the training process, the output of the network is compared with the preset segmentation labels, the error is calculated, and the model parameters are adjusted through back propagation.

[0051] Step 22: Perform a false positive simulation operation on the three-dimensional image data sample to obtain a first damaged label of the three-dimensional image data sample.

[0052] When training the second segmentation network, in order to enhance its ability to correct false positive errors, it is necessary to generate a first damaged label with false positive errors by performing false positive simulation operations on the 3D image data samples. The purpose of this process is to simulate the common false positive segmentation errors in the real segmentation process, so that the network can be exposed to these problems during the training phase and master the ability to correct false positive errors through learning.

[0053] In one possible implementation, refer to Figure 3 , Figure 3 FIG. 3 is a flow chart of the three-dimensional image data segmentation method provided by the present invention. Figure 3 As shown, step 22 specifically includes steps 221-223: Step 221: Determine the maximum voxel value of the 3D image data sample.

[0054] In the process of false positive simulation operation, determining the maximum voxel value of the 3D image data sample is the key first step. The main purpose of this process is to provide a basis for the subsequent calculation of the screening threshold, so that the generated first damaged label can accurately simulate the distribution characteristics of false positive segmentation errors. 3D image data usually contains voxel values ​​of different intensities. The maximum voxel value represents the highest brightness area or signal intensity in the data sample, reflecting the extreme value of the target area or highlighted background. By obtaining this value, the threshold calculation can be dynamically adjusted to make the false positive simulation operation closer to the error characteristics in the actual segmentation task.

[0055] Specifically, for a given 3D image data sample , it is necessary to traverse the positions (i, j, k) of all voxels in the data sample and find the maximum value of the image data This calculation method can efficiently determine the highest intensity value in the image data without being interfered by background noise or local features. At the same time, the determination of the maximum voxel value m is global, ensuring that the subsequent threshold calculation is based on the characteristics of the entire data sample rather than being limited to the local area.

[0056] Step 222: Calculate a screening threshold based on the maximum voxel value.

[0057] In the false positive simulation operation, the purpose of this operation based on the maximum voxel value of the 3D image data sample is to dynamically adjust the threshold so that the false positive simulation can adapt to the brightness distribution characteristics of different types of 3D image data, thereby generating a first damaged label that is closer to the actual segmentation error. The setting of the screening threshold not only affects the distribution of false positive areas, but also determines the pertinence and effectiveness of the simulation results for the training network.

[0058] Specifically, the calculation of the screening threshold is based on the determined maximum voxel value m. By setting a fixed ratio r ∈ [0, 1], the screening threshold t = r × m is calculated. Here, the ratio r is an adjustable parameter used to control the range and density of the generated false positive regions. When r is small, the screening threshold t is low, and more voxels will be marked as target regions, with a denser false positive distribution in the simulation results; while when r is large, the screening threshold t is high, and only voxels in high-intensity regions will be marked as target, with a relatively sparse false positive distribution in the simulation results.

[0059] Step 223: Perform binarization on all voxel values of the three-dimensional image data sample based on the screening threshold to obtain a first damage label; wherein, the first damage label contains multiple voxel labels. When the voxel value is less than the screening threshold, the corresponding voxel label is set to 0, and when the voxel value is not less than the screening threshold, the corresponding voxel label is set to 1.

[0060] In the false positive simulation operation, the core purpose of performing binarization on the three-dimensional image data sample based on the screening threshold is to distinguish high-intensity regions from low-intensity regions by setting a classification criterion for voxel intensity, thereby simulating the distribution characteristics of false positive segmentation errors. The design of binarization can efficiently separate the target regions from the background regions in the three-dimensional image data, while introducing regions with incorrect markings to provide training data for the network against false positive errors.

[0061] Specifically, the calculated screening threshold is used as the criterion for voxel classification, and each voxel value V(i, j, k) in the three-dimensional image data sample is compared. When the voxel value V(i, j, k) < t, the voxel is marked as the background (label value is 0); when the voxel value V(i, j, k) ≥ t, the voxel is marked as the target (label value is 1). This operation generates a first damage label containing the classification results of all voxels , which has the same shape as the original image data, but the label values are only 0 or 1.

[0062] The logic of this binarization can accurately simulate the common false positive error phenomenon in the segmentation task. By adjusting the screening threshold t, the number and distribution of false positive regions can be flexibly controlled. For example, when the threshold is low, high-intensity background regions are more likely to be mislabeled as targets, thus expanding the range of false positive distribution; while when the threshold is high, only some high-intensity regions are mislabeled, and the false positive distribution is sparser. This dynamically adaptable characteristic ensures that the simulation results can cover a variety of false positive scenarios, enhancing the diversity of network training.

[0063] The first damaged label generated by binarization successfully introduces false positive errors into the damaged label, providing a real simulated segmentation error sample for the second segmentation network. This design effectively improves the network's ability to identify and correct false positive areas during training, laying a solid foundation for the high accuracy and low error rate of the final segmentation results. At the same time, the generation method of damaged labels is simple and efficient, and is suitable for different types of 3D image data.

[0064] Step 23: Perform a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample.

[0065] After the false positive simulation operation generates the first damaged label, in order to further enhance the learning ability of the second segmentation network for complex segmentation errors, it is necessary to perform a random morphological operation on the first damaged label to generate the second damaged label. The purpose of this operation is to simulate the common regional morphological irregularities in segmentation tasks, such as morphological changes in the target area, boundary errors, or local structural damage, by randomly introducing morphological changes, thereby increasing the diversity of training data and improving the network's robustness to complex morphological segmentation errors.

[0066] In one possible implementation, refer to Figure 4 , Figure 4 FIG. 4 is a flow chart of the three-dimensional image data segmentation method provided by the present invention. Figure 4 As shown, step 23 specifically includes steps 231-233: Step 231: Determine multiple morphological operations.

[0067] In the process of performing random morphological operations on the first damaged label, determining multiple morphological operations is a key step in achieving morphological diversity simulation. Common forms of errors in segmentation tasks include enlargement, reduction, breakage or irregular morphology of the target area, which are difficult to fully simulate with a single morphological operation. Therefore, by pre-defining multiple morphological operations such as dilation, erosion, opening and closing operations, we can lay the foundation for generating richer second damaged labels, allowing the network to learn more segmentation error types and their correction strategies during training.

[0068] Specifically, the determination of multiple morphological operations includes but is not limited to the following operations: no morphological operation is performed; the dilation operation is used to expand the boundary of the target area to simulate the expansion of the false positive area; the erosion operation is used to shrink the boundary of the target area to simulate the error of shrinking the target area or incomplete boundary; the opening operation can simulate the removal of small false positive areas or the elimination of background noise by first eroding and then dilating; the closing operation can simulate the smoothness of the boundary of the target area or the local broken connection by first dilating and then eroding. By selecting these classic morphological operations, it is ensured that the simulated segmentation errors can cover a variety of common scenarios.

[0069] Step 232: Use a three-dimensional structure element of a preset size as a structure element for morphological operations.

[0070] Specifically, in random morphological operations, choosing a suitable structural element is the key to the morphological processing being able to accurately simulate segmentation errors. The shape and size of the structural element directly determine the scope and effect of the morphological operation on the first damaged label. In order to ensure that the morphological operation can reflect the local characteristics of the segmentation error and avoid excessive modification of the target area, a three-dimensional cross structural element is selected as the three-dimensional structural element. The three-dimensional cross structural element consists of a central voxel and its neighboring voxels in six main directions. This design can perform detailed processing on the local features of the target area, such as extending the boundary, shrinking the area, or smoothing the edge, without causing too much impact on the non-local area in three dimensions. The preset size of the structural element is preferably 3×3×3, which not only ensures the high efficiency of the morphological operation, but also can flexibly adapt to the features of different scales in the three-dimensional image data.

[0071] This structural element is chosen because it can accurately simulate common morphological problems in segmentation, such as insufficient boundary extension or the connection of false positive regions. By limiting the scope of the morphological operation, it ensures that the generated second damaged labels contain representative segmentation errors while maintaining a high degree of consistency with the characteristics of the real data. This design avoids the problem of segmentation error distortion that may be caused by too large or too small structural elements, while improving the authenticity and diversity of damaged labels.

[0072] Step 233: randomly select a morphological operation, and perform the randomly selected morphological operation on the first damaged label according to the structural element to obtain a second damaged label.

[0073] Specifically, when performing random selection, a morphological operation is uniformly sampled from no operation, dilation, erosion, opening operation, and closing operation. The dilation operation simulates segmentation errors of target region enlargement or false positive region enlargement by expanding the target region boundary; the erosion operation simulates segmentation errors of target region reduction or missing boundary by shrinking the target region boundary; the opening operation simulates the removal of small false positive regions by first eroding and then dilating; the closing operation simulates boundary smoothing or local broken connection by first dilating and then eroding. Random selection ensures that the generation of different segmentation error types is uncertain, making the training data more diverse.

[0074] Once the randomly selected operation is determined, the first damaged label is morphologically transformed using the preset 3×3×3 three-dimensional cross structure element to obtain the label damaged by the random morphological operation. The structural element limits the operation range, so that the changes are mainly concentrated in the local area, thus avoiding excessive impact on the overall morphology. The transformed second damaged label can more realistically simulate the morphological errors in the target area, including insufficient expansion of the boundary, connection of false positive areas, or residual small noise.

[0075] By randomly selecting morphological operations, the generation of the second damaged label is no longer limited to fixed error forms, but can dynamically present different segmentation error characteristics. This design significantly enhances the diversity of training data and improves the ability of the second segmentation network to cope with different segmentation errors during training, enabling it to better identify and correct various types of segmentation errors in actual tasks, providing an important guarantee for the high accuracy and robustness of the final segmentation results.

[0076] Step 24: Perform a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample.

[0077] After generating the second damaged label, in order to further enrich the types and manifestations of segmentation errors, it is necessary to perform a random masking operation on the second damaged label to simulate the common situation of target area information loss or partial occlusion in segmentation tasks. The core purpose of this operation is to introduce uncertainty into the second damaged label through a randomized masking strategy, generate actual damaged labels with higher complexity and diversity, and thus enhance the adaptability of the second segmentation network to segmentation errors.

[0078] In one possible implementation, refer to Figure 5 , Figure 5 FIG. 5 is a flow chart of the three-dimensional image data segmentation method provided by the present invention. Figure 5 As shown, step 24 specifically includes the following steps: Step 241: Set a random masking ratio k∈[0,1].

[0079] In the random masking operation, setting the random masking ratio k∈[0,1] is a key step in generating random masks. The setting of the masking ratio directly affects the ratio of voxels that are retained and masked in the actual damaged label, thereby controlling the scope and degree of information loss in the target area. The flexible adjustment of this ratio can simulate partial loss or occlusion of the target area in different scenarios, making the generated actual damaged labels more diverse and representative, thereby providing more targeted training samples for the second segmentation network.

[0080] Specifically, the masking ratio k represents the probability that each voxel value in the generated random mask is 1 (retained), while 1−k represents the probability that the voxel value is 0 (masked). When the masking ratio k is small, most voxels will be masked, and the target area information loss in the actual damaged label is high; when the masking ratio k is large, the target area information is highly retained, and only a small number of voxels are masked. This flexible ratio control enables the random masking operation to adapt to segmentation scenarios of different complexity.

[0081] Step 242: Generate a random mask with the same shape as the second damaged label according to the random masking ratio, wherein: the probability of setting the value of each position in the random mask to 1 is k, and the probability of setting it to 0 is 1-k.

[0082] In the random masking operation, generating a random mask for the second damaged label according to the set random masking ratio k is a key step. The generated random mask determines which voxels in the second damaged label will be retained and which will be masked, thereby simulating the information loss of the target area. The design of the random mask needs to fully consider the diversity of segmentation errors, and enhance the complexity of the damaged label by introducing randomness, providing more challenging samples for the training of the second segmentation network.

[0083] Specifically, the random mask has the same shape D×H×W as the second damaged label. The mask value at each position is determined by random sampling, where the probability of the value being 1 is k (indicating that the corresponding voxel is retained) and the probability of the value being 0 is 1−k (indicating that the corresponding voxel is masked). This randomized design can introduce uncertainty at different locations of the label, so that the target area presents a variety of information loss distribution characteristics in different samples.

[0084] The process of generating random masks ensures the diversity of masking effects. For example, at a lower masking ratio k, most voxels are masked, and the generated corrupted labels can simulate the scene of severe target area loss; while at a higher masking ratio k, only a small number of voxels are masked, simulating the scene of slight loss or noise occlusion of the target area. This flexibility allows the generated corrupted labels to cover different degrees of segmentation errors, providing more comprehensive error samples for the network's learning process.

[0085] Step 243: Multiply the random mask and the second damaged label element by element to obtain the actual damaged label.

[0086] In the random masking operation, the generated random mask is element-wise multiplied with the second damaged label. The purpose of this operation is to control the voxel retention and masking of the target area through the mask, and to generate an actual damaged label with random information missing characteristics, thereby simulating the scene in which the target area is partially occluded or lost in the segmentation task, and providing diverse input samples for the training of the second segmentation network.

[0087] Specifically, the second damaged label L2 and the random mask M have the same shape, both of which are D×H×W. The voxels with a value of 1 in the random mask M indicate that the label at the corresponding position is retained, and the voxels with a value of 0 indicate that the label at the corresponding position is masked. By element-by-element multiplication, the actual damaged label L actual is generated, and its calculation formula is: ; Among them, (i, j, k) is the position of the voxel. When remain unchanged; when When Forced to 0, indicating that the voxel is masked.

[0088] This process can randomly introduce information loss in the second damaged label, and the generated actual damaged label contains randomly distributed target area missing characteristics. By adjusting the masking ratio k in mask generation, the degree of target area information loss of the actual damaged label can be flexibly controlled, ranging from slight loss to large-scale occlusion.

[0089] The operation of element-wise multiplication of the random mask and the second damaged label makes the actual damaged label more random and complex. Through this design, the second segmentation network can be exposed to various types of segmentation error scenarios during training, including partial loss of target area, loss of boundary information and other problems, thereby significantly improving the network's ability to deal with complex segmentation errors in real tasks, providing an important guarantee for the high accuracy and robustness of the final segmentation results.

[0090] Step 25: The three-dimensional image data samples and the actual damage labels are concatenated in the channel dimension, input into the initial model for training, and the preset segmentation labels are used as training targets to train a second segmentation network.

[0091] Specifically, the 3D image data sample V and the actual damage label L actual First, the concatenation is performed in the channel dimension to generate fused data F, each channel of which contains the intensity information of the original image data and the category information of the damaged label. The fused data not only retains the detailed features in the image data, but also introduces the simulated segmentation error features through the actual damaged label, providing rich input information for network learning.

[0092] The fused data F is input into the initial model for training. During the training process, the network extracts the features of the input data through the encoder, gradually identifies the characteristic patterns of the target area and the distribution characteristics of segmentation errors, and generates optimized segmentation results in the decoding stage. The network output results are consistent with the preset segmentation labels L true The loss function is generated by comparing the two and calculating the difference between them. The loss function is often in the form of cross entropy loss or Dice coefficient loss, which is used to quantify the deviation between the network prediction result and the true label.

[0093] The loss value updates the network parameters through the back-propagation mechanism, allowing the network to gradually learn how to extract effective features from the fused data and correct segmentation errors. Specifically, the network can eliminate false positive areas, repair discontinuities in the target structure, and optimize the accuracy of the target boundary by learning the error characteristics in the actual damaged labels.

[0094] Through this training process, the second segmentation network gradually acquires the ability to identify and correct various types of segmentation errors. The design of fused data ensures the richness of network input information. The generation of actual damaged labels simulates typical error scenarios in segmentation tasks, and the preset segmentation labels provide clear learning goals. Ultimately, the trained second segmentation network can significantly improve segmentation accuracy and robustness, providing a reliable solution for high-quality segmentation of 3D imaging data.

[0095] In a possible implementation manner, step 13 specifically includes the following steps: Inputting the fused data into the second segmentation network, correcting the segmentation errors in the preliminary three-dimensional segmentation result in the second segmentation network, including eliminating the false positive area in the preliminary three-dimensional segmentation result, repairing the broken part caused by discontinuity in the preliminary three-dimensional segmentation result, and refining the boundary of the preliminary three-dimensional segmentation result; Output the corrected final 3D segmentation result.

[0096] Specifically, the fused data is generated by splicing the preliminary 3D segmentation results and the original 3D image data in the channel dimension, which contains both the preliminary semantic information of the target area and the detailed features of the original image. The second segmentation network adopts an encoder-decoder structure to perform deep feature extraction on the fused data through multi-layer convolution and feature aggregation. In the encoding stage, the network identifies possible false positive areas in the fused data, the breakpoints of the target structure, and the location and characteristics of the blurred boundary areas.

[0097] In the decoding stage, the network uses jump connections to directly transfer the high-resolution features of the encoding stage to the decoding stage, and restores the spatial resolution of the fused data through gradual upsampling. In this process, the network eliminates false positive areas, repairs broken target structures, and refines boundary areas based on the contextual information in the feature map, making the target area more complete and with clear boundaries. By combining the encoding and decoding processes, the second segmentation network can comprehensively consider global and local information and accurately correct errors in the initial segmentation results.

[0098] After being processed by the second segmentation network, the generated optimized output is the final 3D segmentation result. The optimized segmentation result eliminates the main errors in the initial segmentation result, including redundant parts of false positive areas, discontinuous points of the target structure, and blurred areas of the boundary. The generation of the final segmentation result marks the completion of the segmentation optimization process, and its accuracy and robustness are significantly better than the initial segmentation result.

[0099] By outputting the corrected final 3D segmentation results, the second segmentation network not only solves the shortcomings of the initial segmentation stage, but also achieves high-precision processing of complex segmentation tasks by learning the characteristics of fused data. This result significantly improves the practicality of the segmentation task and provides a reliable foundation for subsequent analysis and application based on the segmentation results.

[0100] Reference Figure 6 , Figure 6 : is a schematic diagram of the structure of the three-dimensional image data segmentation system provided by the present invention, the system comprises: A processing module, used for inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; A splicing module is used to splice the preliminary 3D segmentation result with the 3D image data in the channel dimension to obtain fused data; The processing module is also used to input the fused data into the second segmentation network to obtain the final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct the segmentation errors in the preliminary three-dimensional segmentation result.

[0101] In a possible implementation, the system further includes a model training module, which is used to: Determining a preset segmentation label for a three-dimensional image data sample; Performing a false positive simulation operation on the three-dimensional image data sample to obtain a first damaged label of the three-dimensional image data sample; Performing a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample; Performing a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample; The 3D image data samples and actual damage labels are concatenated in the channel dimension and input into the initial model for training. The preset segmentation labels are used as training targets to train the second segmentation network.

[0102] In a possible implementation, the model training module is further used to: determining a maximum voxel value of a sample of three-dimensional image data; Based on the maximum voxel value, the screening threshold was calculated; Based on the screening threshold, all voxel values ​​of the three-dimensional image data sample are binarized to obtain a first damage label; wherein the first damage label includes multiple voxel labels, when the voxel value is less than the screening threshold, the corresponding voxel label is set to 0, and when the voxel value is not less than the screening threshold, the corresponding voxel label is set to 1.

[0103] In a possible implementation, the model training module is further used to: Identify various morphological operations; Use a 3D structure element of a preset size as the structure element for morphological operations; A morphological operation is randomly selected, and the randomly selected morphological operation is performed on the first damaged label according to the structural element to obtain a second damaged label.

[0104] In a possible implementation, the model training module is further used to: Set the random masking ratio k∈[0,1]; According to the random masking ratio, a random mask with the same shape as the second damaged label is generated, where: the probability of setting the value of each position in the random mask to 1 is k, and the probability of setting it to 0 is 1-k; The random mask is element-wise multiplied with the second damaged label to obtain the actual damaged label.

[0105] In a possible implementation manner, the processing module is further configured to: In the encoder, a dual-branch network is used to extract features from 3D image data. The dual-branch network includes a convolutional neural network branch and a Transformer branch. At different scales of the encoder, the three-dimensional image data is encoded and feature extracted by the convolutional neural network branch and the Transformer branch, respectively, to obtain the first feature map at different scales and the second feature map at different scales; The first feature map and the second feature map at the same scale are fused to obtain multiple fused features; In the decoder, all fused features are decoded through convolution and upsampling operations, and skip connections are established between the corresponding scales of the decoder and the encoder; Through the output of the decoder, a preliminary three-dimensional segmentation result is obtained.

[0106] In a possible implementation manner, the processing module is further configured to: Inputting the fused data into the second segmentation network, correcting the segmentation errors in the preliminary three-dimensional segmentation result in the second segmentation network, including eliminating the false positive area in the preliminary three-dimensional segmentation result, repairing the broken part caused by discontinuity in the preliminary three-dimensional segmentation result, and refining the boundary of the preliminary three-dimensional segmentation result; Output the corrected final 3D segmentation result.

[0107] It should be noted that the 3D image data segmentation system provided by the present invention can execute the 3D image data segmentation method of any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0108] Figure 7 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710 (processor), a communication interface 720 (CommunicationsInterface), a memory 730 (memory) and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the three-dimensional image data segmentation method, which includes: inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result with the three-dimensional image data in the channel dimension to obtain fused data; inputting the fused data into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct the segmentation errors in the preliminary three-dimensional segmentation result.

[0109] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0110] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the three-dimensional image data segmentation method provided by the above-mentioned embodiments, and the method includes: inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result with the three-dimensional image data in the channel dimension to obtain fused data; inputting the fused data into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

[0111] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by the processor 710, it is implemented to execute the three-dimensional image data segmentation method provided by the above-mentioned embodiments, the method comprising: inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result with the three-dimensional image data in the channel dimension to obtain fused data; inputting the fused data into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

[0112] The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0113] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional image data segmentation method, characterized in that: include: Inputting the three-dimensional image data into the first segmentation network to obtain a preliminary three-dimensional segmentation result; splicing the preliminary three-dimensional segmentation result with the three-dimensional image data in the channel dimension to obtain fused data; The fused data is input into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

2. The three-dimensional image data segmentation method according to claim 1, characterized in that: The second segmentation network is trained by the following method: Determining a preset segmentation label of the three-dimensional image data sample; Performing a false positive simulation operation on the three-dimensional image data sample to obtain a first damaged label of the three-dimensional image data sample; Performing a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample; Performing a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample; The three-dimensional image data samples and the actual damage labels are spliced ​​in the channel dimension, input into the initial model for training, and the preset segmentation labels are used as training targets to train and obtain the second segmentation network.

3. The three-dimensional image data segmentation method according to claim 2, characterized in that: The performing a false positive simulation operation on the three-dimensional image data sample to obtain a first damaged label of the three-dimensional image data sample specifically includes: Determining a maximum voxel value of the three-dimensional image data sample; calculating a screening threshold based on the maximum voxel value; Based on the screening threshold, all voxel values ​​of the three-dimensional image data sample are binarized to obtain the first damage label; wherein the first damage label includes multiple voxel labels, when the voxel value is less than the screening threshold, the corresponding voxel label is set to 0, and when the voxel value is not less than the screening threshold, the corresponding voxel label is set to 1.

4. The three-dimensional image data segmentation method according to claim 2, characterized in that: The performing a random morphological operation on the first damaged label to obtain a second damaged label of the three-dimensional image data sample specifically includes: Identify various morphological operations; Use a 3D structure element of a preset size as the structure element for morphological operations; A morphological operation is randomly selected, and the randomly selected morphological operation is performed on the first damaged label according to the structural element to obtain the second damaged label.

5. The three-dimensional image data segmentation method according to claim 2, characterized in that: The performing a random masking operation on the second damaged label to obtain an actual damaged label of the three-dimensional image data sample specifically includes: Set the random masking ratio k∈[0,1]; According to the random masking ratio, a random mask having the same shape as the second damaged label is generated, wherein: the probability of setting the value of each position in the random mask to 1 is k, and the probability of setting the value to 0 is 1-k; The random mask is multiplied element-by-element by the second damage label to obtain the actual damage label.

6. The three-dimensional image data segmentation method according to claim 1, characterized in that: The first segmentation network adopts an encoder-decoder structure; the three-dimensional image data is input into the first segmentation network to obtain a preliminary three-dimensional segmentation result, specifically including: In the encoder, a dual-branch network is used to extract features from the three-dimensional image data, and the dual-branch network includes a convolutional neural network branch and a Transformer branch; At different scales of the encoder, the three-dimensional image data is encoded and feature extracted by a convolutional neural network branch and a Transformer branch respectively, to obtain a first feature map at different scales and a second feature map at different scales; The first feature map and the second feature map at the same scale are fused to obtain multiple fused features; In the decoder, all the fused features are decoded by convolution and upsampling operations, and a skip connection is established between the decoder and the corresponding scales of the encoder; The preliminary three-dimensional segmentation result is obtained through the output of the decoder.

7. The three-dimensional image data segmentation method according to claim 1, characterized in that: The step of inputting the fused data into a second segmentation network to obtain a final three-dimensional segmentation result specifically includes: Inputting the fused data into a second segmentation network, correcting segmentation errors in the preliminary three-dimensional segmentation result in the second segmentation network, including eliminating false positive areas in the preliminary three-dimensional segmentation result, repairing broken parts in the preliminary three-dimensional segmentation result caused by discontinuity, and refining the boundaries of the preliminary three-dimensional segmentation result; Output the corrected final 3D segmentation result.

8. A three-dimensional image data segmentation system, characterized in that: include: A processing module, used for inputting the three-dimensional image data into a first segmentation network to obtain a preliminary three-dimensional segmentation result; A splicing module, used for splicing the preliminary 3D segmentation result with the 3D image data in a channel dimension to obtain fused data; The processing module is also used to input the fused data into a second segmentation network to obtain a final three-dimensional segmentation result; wherein the second segmentation network is used to process the fused data to correct segmentation errors in the preliminary three-dimensional segmentation result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the three-dimensional image data segmentation method as described in any one of claims 1-7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the three-dimensional image data segmentation method as described in any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the three-dimensional image data segmentation method as described in any one of claims 1 to 7 is implemented.

Images