Three-dimensional medical image segmentation method without resampling

Abstract

The application provides a three-dimensional medical image segmentation method without resampling, relates to the field of computer vision, and particularly relates to a three-dimensional medical image segmentation method without resampling. The three-dimensional medical image segmentation framework of the application does not need resampling in the training and reasoning stages, can retain the information of the original medical image to a greater extent, avoids the loss of details caused by resampling, and thus improves the segmentation accuracy. By introducing convolutional low-rank decomposition into the attention module, the application can approximate the global attention calculation of high-dimensional matrix multiplication to low-rank matrix operation, thereby effectively reducing the calculation complexity. At the same time, the corresponding adaptive parameterized Euclidean matrix is also compressed to a small scale, under the condition of ensuring the global dependence modeling capability, the memory occupation and the calculation resource consumption are greatly reduced, so that the application can be efficiently run on large-scale three-dimensional medical image data.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more specifically to a three-dimensional medical image segmentation method that does not require resampling. Background Technology

[0002] With the development of deep learning technology, computer vision has been widely used in the field of medical image recognition. In 3D medical image segmentation, different imaging devices and scanning protocols often lead to inconsistencies in voxel spacing and resolution.

[0003] To enable 3D images to be used as input for deep learning models, related techniques typically require resampling to standardize them to a fixed voxel spacing or size for subsequent processing. In clinical settings, such as preoperative tumor delineation in neurosurgery, vascular segmentation in hepatobiliary surgery, and kidney and ureter identification in urology, segmentation accuracy directly impacts surgical planning and intraoperative navigation.

[0004] However, resampling introduces interpolation errors and information loss, leading to a lack of spatial detail and unnecessary noise. This weakens the model's sensitivity to real lesion features and minute anatomical structures, thus limiting further improvements in overall segmentation performance. Especially in scenarios involving intraoperative navigation, the loss of detail can lead to misjudgments of lesion boundaries or blood vessel orientation, posing potential risks. Furthermore, resampling increases computational and storage overhead, hindering efficient real-time applications in clinical practice. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a three-dimensional medical image segmentation method that does not require resampling, solving the technical problem of achieving high-precision three-dimensional medical image segmentation without the need for resampling.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A resampling-free 3D medical image segmentation method, based on a 3D medical image segmentation framework, wherein the 3D medical image segmentation framework includes a hierarchical encoder and a hierarchical decoder; the 3D medical image segmentation method includes:

[0010] Acquire and preprocess raw medical image data acquired in slice form; wherein the preprocessing includes normalization and slicing.

[0011] The hierarchical encoder is used to extract features from the preprocessed data, retaining the output features of each level and the final features extracted by the hierarchical encoder.

[0012] The final feature is used as the input to the hierarchical decoder, and each output feature is used as the input to the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer.

[0013] The reconstructed 3D medical image is post-processed to obtain segmentation results of the target region; wherein the post-processing includes at least block fusion.

[0014] Preferably, after obtaining the segmentation results of the target region, the segmentation results are visualized and converted in format.

[0015] Preferably, each level of the hierarchical encoder includes several Transformer modules, and each Transformer module includes a first attention submodule and a convolution submodule.

[0016] Preferably, the first attention submodule includes a convolutional low-rank decomposition unit, which is used for:

[0017] Project the current input features separately to obtain the query, key, and value;

[0018] The dimensions of the keys and values ​​are reduced using two different convolutional layers to obtain the projected keys and projected values.

[0019] Preferably, the first attention submodule includes an adaptive parameterized Euclidean matrix weighting unit, which is used for:

[0020] Calculate the maximum mean difference (MMD) between each pair of adjacent elements of the projected value, and obtain the length scale through linear or nonlinear mapping.

[0021] Based on the coordinate matrix of the stated value and the coordinate matrix of the projected value, a distance matrix is ​​obtained; and combined with the length scale, a scaled distance matrix is ​​obtained.

[0022] The scaled distance matrix is ​​used as the input to the linear mapping layer, and the output of the linear mapping layer is residually concatenated with the distance matrix to obtain the corrected parameters.

[0023] The corrected distance matrix is ​​normalized using Softmax, and the Hadamard product of the normalized result and the original attention score matrix is ​​calculated to obtain the double-weighted attention score matrix; wherein the original attention score matrix is ​​the outer product of the query and the projected key;

[0024] Calculate the outer product of the projected value and the second-weighted attention score matrix, and use it as the current attention feature.

[0025] Preferably, each level of the hierarchical decoder includes several Transformer modules, wherein the first Transformer module of each level of the hierarchical decoder is a cross Transformer module, and the remaining Transformer modules have the same structure as the Transformer modules used in the hierarchical encoder.

[0026] Preferably, the cross-Transformer module includes a second attention submodule and a convolution submodule. Each second attention submodule is hop-connected to the hierarchical encoder at the corresponding level. The second attention submodule includes a convolutional low-rank decomposition unit and an adaptive parameterized Euclidean matrix weighting unit.

[0027] A resampling-free 3D medical image segmentation system, based on a 3D medical image segmentation framework, wherein the 3D medical image segmentation framework includes a hierarchical encoder and a hierarchical decoder; the 3D medical image segmentation system includes:

[0028] The data acquisition and preprocessing module is used to acquire and preprocess raw medical image data collected in slice form; wherein the preprocessing includes normalization and slicing.

[0029] The feature encoding module is used to extract features from the preprocessed data using the hierarchical encoder, retaining the output features of each level and the final features extracted by the hierarchical encoder.

[0030] The feature decoding module is used to take the final feature as the input of the hierarchical decoder and take each output feature as the input of the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer.

[0031] The result acquisition module post-processes the reconstructed 3D medical image to obtain the segmentation result of the target region; wherein the post-processing includes at least block fusion.

[0032] A storage medium storing a computer program for three-dimensional medical image segmentation without resampling, wherein the computer program causes a computer to perform the three-dimensional medical image segmentation method as described above.

[0033] An electronic device, characterized in that it comprises:

[0034] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the three-dimensional medical image segmentation method as described above.

[0035] (III) Beneficial Effects

[0036] This invention provides a three-dimensional medical image segmentation method that does not require resampling. Compared with existing technologies, it has the following advantages:

[0037] In this invention, firstly, raw medical image data acquired in slice form is obtained and preprocessed; secondly, based on a hierarchical encoder, features are extracted from the preprocessed data, and then based on a hierarchical decoder, three-dimensional medical images are reconstructed layer by layer; finally, the reconstructed three-dimensional medical images are post-processed to obtain the segmentation results of the target region. This invention employs normalization and block processing as preprocessing methods, as well as block fusion and other post-processing methods, to replace the traditional resampling step. This approach preserves more information from the original medical image data, avoids detail loss caused by resampling, and thus improves the segmentation accuracy of target regions such as organs, tissues, or lesion areas. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0039] Figure 1 A flowchart illustrating a three-dimensional medical image segmentation method without resampling, provided in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of a three-dimensional medical image segmentation framework provided in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of a Transformer module provided in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the structure of an attention module provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of another attention module provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0045] This application provides a three-dimensional medical image segmentation method that does not require resampling, thus solving the technical problem of achieving high-precision three-dimensional medical image segmentation without resampling.

[0046] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0047] In addition to considering the drawbacks of resampling operations in related technologies, this invention also addresses the lack of adaptive feature extraction mechanisms when processing samples from different acquisition devices and at different resolutions, resulting in insufficient model generalization ability. Furthermore, while Transformer-based global attention methods can model long-range dependencies, they suffer from excessive computational demands and memory consumption on 3D data, making it difficult to meet the efficiency and real-time requirements of segmentation models in clinical surgical scenarios. Therefore, to achieve high-precision 3D medical image segmentation without resampling, this invention provides a deep learning-based 3D medical image segmentation method and framework related to data preprocessing and post-processing, which can be widely applied to preoperative planning and intraoperative navigation in surgical scenarios such as neurosurgery, hepatobiliary surgery, urology, and thoracic surgery.

[0048] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0049] Example 1:

[0050] like Figure 1 As shown, this embodiment of the invention provides a three-dimensional medical image segmentation method without resampling, based on a three-dimensional medical image segmentation framework, which includes a hierarchical encoder and a hierarchical decoder; the three-dimensional medical image segmentation method includes:

[0051] S1. Acquire and preprocess the raw medical image data acquired in slice form; wherein the preprocessing includes normalization and slicing.

[0052] S2. Use the hierarchical encoder to extract features from the preprocessed data, retaining the output features of each level and the final features extracted by the hierarchical encoder;

[0053] S3. The final feature is used as the input of the hierarchical decoder, and each output feature is used as the input of the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer.

[0054] S4. The reconstructed three-dimensional medical image after post-processing is used to obtain the segmentation result of the target region; wherein the post-processing includes at least block fusion.

[0055] The embodiments of the present invention employ preprocessing methods such as normalization and block processing, as well as postprocessing methods such as block fusion, to replace the traditional resampling steps. This can preserve more information from the original medical image data, avoid detail loss caused by resampling, and thus improve the segmentation accuracy of target areas such as target organs, tissues, or lesion areas.

[0056] In an optional implementation, S4 further includes:

[0057] After obtaining the segmentation results of the target region, the segmentation results are visualized and converted in format.

[0058] This invention improves the way information is transmitted by visualizing and converting the segmentation results of the target region, making it easier for users to quickly and intuitively identify complex information, and also facilitating deeper intelligent diagnostic analysis.

[0059] like Figure 2 As shown, Figure 2 A schematic diagram of a three-dimensional medical image segmentation framework is disclosed, which includes a hierarchical encoder and a hierarchical decoder. Wherein:

[0060] The hierarchical encoder is divided into several levels, preferably 3 to 4 levels. Figure 2 The demonstrated hierarchical encoder has three levels to extract feature information from shallow to deep layers. Furthermore, a dedicated downsampling module is included before the input to each level. Figure 2 (Not shown in the image) is used to reduce the size of the feature map and enhance its semantics, thereby improving the model's ability to perceive structures at different scales while ensuring computational efficiency.

[0061] The hierarchical decoder is also divided into several levels, consistent with the number of levels in the hierarchical encoder. Figure 2 The demonstrated hierarchical decoder also has three levels. Furthermore, a dedicated upsampling module is provided before the input of each level. Figure 2 (Not shown in the image), used to enlarge the feature map and restore spatial details.

[0062] The following section will explain each step of the above solution from a reasoning perspective:

[0063] In step S1, the raw medical image data acquired in slice form is obtained and preprocessed; wherein the preprocessing is normalization and slicing.

[0064] This embodiment limits the acquisition of raw medical image data in the form of slices, such as computed tomography (CT) images, magnetic resonance imaging (MRI) images, etc.

[0065] It is important to note that these types of images often exhibit strong heterogeneity: the slice thickness settings vary between different image samples, resulting in a lack of uniformity in spatial resolution across different cases; and within the same image sample, while slices typically have high resolution in the planar direction (length and width), the resolution in the slice thickness direction is often significantly reduced, thus creating anisotropic voxel structures. Furthermore, gaps between slices may exist in some MRI images, causing insufficient continuity in 3D reconstruction. In summary, these characteristics necessitate the reliance on resampling in related techniques to achieve spatial scale uniformity, but this also introduces interpolation errors and artifacts, leading to blurred boundaries and loss of detail.

[0066] Unlike resampling preprocessing, in order to maintain the original resolution information of the data, this embodiment of the invention performs normalization processing on the original medical image data during the inference stage, and performs block operation according to the hardware memory conditions and task requirements. The block size can be flexibly set.

[0067] It is worth noting that during the training phase, existing general data augmentation methods, such as random rotation, flipping, scaling, brightness / contrast adjustment, and noise simulation, can be further combined to improve the robustness of the model.

[0068] In step S2, the hierarchical encoder is used to extract features from the preprocessed data, retaining the output features of each level and the final features extracted by the hierarchical encoder.

[0069] Each level of the hierarchical encoder in this embodiment of the invention includes several Transformer modules. Each Transformer module includes a first attention submodule and a convolutional (CNN) submodule. The basic structure of the Transformer module can be found in [reference needed]. Figure 3 .

[0070] Correspondingly, the output feature of each level refers to the final output after all Transformer modules in that level have been processed in sequence, that is, the output feature of the last improved Transformer module in that level. This output feature can integrate multiple feature expressions within that level and represent the final feature representation of that level.

[0071] Furthermore, the first attention submodule of this embodiment improves upon the traditional structure by introducing a convolutional low-rank decomposition unit and an adaptive parameterized Euclidean matrix weighting unit. That is, the first attention submodule includes a convolutional low-rank decomposition unit and an adaptive parameterized Euclidean matrix weighting unit, wherein:

[0072] like Figure 4 As shown, the convolutional low-rank decomposition unit is used to approximate global attention while reducing computational complexity, and its implementation is defined as follows:

[0073] S10. Project the current input features respectively to obtain the query Q, key K and value V.

[0074] S20. Dimensionally reduce the size of the key K and value V using two different convolutional layers to obtain the projected key k and projected value v.

[0075] For example, the dimensions of the key and value can be reduced from D×H×W to d×h×w, where D, H, and W represent the depth / size, length, and width of the key and value before dimensionality reduction, respectively, and d, h, and w represent the depth / size, length, and width of the key and value after dimensionality reduction, respectively.

[0076] like Figure 4 As shown, the adaptive parameterized Euclidean matrix quadratic weighting unit is used to reweight the original attention score matrix to enhance the response of the target region and suppress irrelevant regions. Its implementation is defined as follows:

[0077] S100. Calculate the maximum mean discrepancy (MMD) between each pair of adjacent elements of the projected value v of size d×h×w, obtain d×h×w-1 MMD values, and obtain the length scale, denoted as Θ, through linear or nonlinear mapping; where the θ parameter for calculating MMD is a learnable parameter.

[0078] For example, embodiments of the present invention employ Fast Fourier Transform (FFT) to approximate the kernel function of MMD, thereby significantly improving computational efficiency while maintaining effectiveness.

[0079] S200. Based on the coordinate matrix G (D×H×W) of the value V and the coordinate matrix g (d×h×w) of the projected value v, obtain a distance matrix D of size D×H×W×d×h×w, and divide it by the length scale Θ to obtain the scaled distance matrix α.

[0080] S300. The scaled distance matrix α is used as the input of the linear mapping layer, and the output of the linear mapping layer is residually connected with the distance matrix α to obtain the corrected distance matrix β.

[0081] For example, the linear mapping layer can be a multilayer perceptron (MLP).

[0082] It is important to note that the weights and biases of the linear mapping layer are initialized to zero to ensure that β and α remain consistent in the initial stage, thereby avoiding instability in the early stages of training. As training progresses, the linear layer gradually learns adaptive correction capabilities, enabling β to maintain consistency with α while possessing greater expressive flexibility.

[0083] S400. Normalize the corrected distance matrix β using Softmax, and calculate the Hadamard product of the normalized result and the original attention score matrix S to obtain the double-weighted attention score matrix; wherein the original attention score matrix S is the outer product of the query Q and the projected key k.

[0084] S500. Calculate the outer product of the projected value v and the second-weighted attention score matrix, and use it as the current attention feature Y.

[0085] Based on the above, except for step S500, the adaptive parameterized Euclidean matrix quadratic weighted unit formula can be defined as:

[0086]

[0087] Furthermore, it should be noted that the above-mentioned convolutional sub-modules can adopt conventional convolutional structures in related technologies, such as combinations of three-dimensional convolutional layers, convolutional normalization layers, or convolutional activation layers, and the embodiments of the present invention do not limit this.

[0088] In step S3, the final feature is used as the input of the hierarchical decoder, and each output feature is used as the input of the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer.

[0089] Each level of the hierarchical decoder in this embodiment of the invention includes several Transformer modules. The first Transformer module of each level of the hierarchical decoder is a cross-transformer module, and the remaining Transformer modules have the same structure as the Transformer modules used in the hierarchical encoder. That is, the basic structure of the remaining Transformer modules can also be found in [reference needed]. Figure 3 .

[0090] Furthermore, the aforementioned cross-Transformer module includes a second attention submodule and a convolutional submodule. Each of the second attention submodules is hop-connected to the hierarchical encoder at the corresponding level. The second attention submodule includes a convolutional low-rank decomposition unit and an adaptive parameterized Euclidean matrix weighted unit. Specifically:

[0091] For the cross-Transformer module of the hierarchical decoder, its input includes both the output features of the encoder's corresponding level and the input features of the decoder's current level. The features of the encoder's corresponding level serve as the query Q, and the features of the decoder's current level are projected as the key K and value V, respectively. During the calculation of the cross-Transformer module, convolutional low-rank decomposition units and adaptive parameterized Euclidean matrix weighted units are also introduced to enhance the response of the target region while reducing computational complexity. This establishes an effective interaction relationship between the encoder and decoder features and achieves information fusion. Since the specific calculation process can be fully referred to in step S2, it will not be repeated here.

[0092] The embodiments of the present invention achieve cross-level feature interaction and supplementation through the above-described skip connection design, thereby gradually restoring spatial information and improving segmentation accuracy without relying on resampling.

[0093] In step S4, the reconstructed three-dimensional medical image is post-processed to obtain the segmentation result of the target region; wherein the post-processing includes at least block fusion.

[0094] In this embodiment of the invention, the post-processing in both the training and inference stages does not include a resampling step, which is consistent with the aforementioned pre-processing, thereby avoiding the loss of spatial details caused by resampling.

[0095] For the inference stage, the above post-processing includes at least block fusion (such as 0.5 overlap rate weighted fusion), and other general post-processing can also be selected simultaneously as needed (such as maximum connected component preservation, small connected component removal, thresholding, probability smoothing, etc.) to ensure that target regions such as target organs, tissues or lesion areas are identified and segmented.

[0096] Furthermore, to facilitate users' quick and intuitive identification of complex information and to achieve deeper intelligent diagnostic analysis, after obtaining the segmentation results of the target region, this embodiment of the invention also performs visualization and format conversion on the segmentation results, such as pseudo-color mapping, 3D surface modeling, NIfTI / DICOM output, etc.

[0097] It is worth noting that, for the training phase, the post-processing described above may include result normalization (such as Softmax or Sigmoid), calculating the loss function in units of chunks, and optional result cleaning (such as removing small connected components, filling holes, unifying label format, etc.).

[0098] Example 2:

[0099] This invention provides a three-dimensional medical image segmentation system that does not require resampling. Based on a three-dimensional medical image segmentation framework, the framework includes a hierarchical encoder and a hierarchical decoder. The three-dimensional medical image segmentation system includes:

[0100] The data acquisition and preprocessing module is used to acquire and preprocess raw medical image data collected in slice form; wherein the preprocessing includes normalization and slicing.

[0101] The feature encoding module is used to extract features from the preprocessed data using the hierarchical encoder, retaining the output features of each level and the final features extracted by the hierarchical encoder.

[0102] The feature decoding module is used to take the final feature as the input of the hierarchical decoder and take each output feature as the input of the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer.

[0103] The result acquisition module post-processes the reconstructed 3D medical image to obtain the segmentation result of the target region; wherein the post-processing includes at least block fusion.

[0104] In an optional implementation, the result acquisition module is further configured to:

[0105] After obtaining the segmentation results of the target region, the segmentation results are visualized and converted in format.

[0106] Example 3:

[0107] This invention provides a storage medium storing a computer program for three-dimensional medical image segmentation without resampling, wherein the computer program causes a computer to execute the three-dimensional medical image segmentation method as described in Embodiment 1.

[0108] Example 4:

[0109] This invention provides an electronic device, comprising:

[0110] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the three-dimensional medical image segmentation method as described in Example 1.

[0111] It is understood that the resampling-free three-dimensional medical image segmentation system, storage medium, and electronic device provided in the embodiments of the present invention correspond to the resampling-free three-dimensional medical image segmentation method provided in the embodiments of the present invention. The explanations, examples, and beneficial effects of the relevant contents can be referred to the corresponding parts of the method, and will not be repeated here.

[0112] In summary, compared with existing technologies, it has the following beneficial effects:

[0113] 1. The three-dimensional medical image segmentation framework of this invention does not require resampling during the training and inference stages, which can preserve the information of the original medical image to a greater extent, avoid the loss of details caused by resampling, and thus improve the segmentation accuracy.

[0114] 2. By introducing convolutional low-rank decomposition into the attention module, this embodiment of the invention can approximate the global attention calculation, which originally required high-dimensional matrix multiplication, into low-rank matrix operations, thereby effectively reducing computational complexity. Simultaneously, the corresponding adaptive parameterized Euclidean matrix is ​​also compressed to a smaller scale, significantly reducing memory usage and computational resource consumption while maintaining global dependency modeling capabilities, enabling efficient operation on large-scale 3D medical image data. Compared with existing low-rank decomposition methods, the convolutional low-rank decomposition of this embodiment further significantly reduces the number of parameters and improves computational and training efficiency, exhibiting better practicality and scalability.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional medical image segmentation method that does not require resampling, characterized in that, Based on a three-dimensional medical image segmentation framework, the three-dimensional medical image segmentation framework includes a hierarchical encoder and a hierarchical decoder; The three-dimensional medical image segmentation method includes: Acquire and preprocess raw medical image data acquired in slice form; wherein the preprocessing includes normalization and slicing. The hierarchical encoder is used to extract features from the preprocessed data, retaining the output features of each level and the final features extracted by the hierarchical encoder. The final feature is used as the input to the hierarchical decoder, and each output feature is used as the input to the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer. The post-processed reconstructed 3D medical image is used to obtain segmentation results of the target region; wherein the post-processing includes at least block fusion. Each level of the hierarchical encoder includes several Transformer modules, and each Transformer module includes a first attention submodule and a convolution submodule. The first attention submodule includes a convolutional low-rank decomposition unit, which is used for: Project the current input features separately to obtain the query, key, and value; The dimensions of the keys and values ​​are reduced using two different convolutional layers to obtain the projected keys and projected values. The first attention submodule includes an adaptive parameterized Euclidean matrix weighting unit, which is used for: Calculate the maximum mean difference (MMD) between each pair of adjacent elements of the projected value, and obtain the length scale through linear or nonlinear mapping. Based on the coordinate matrix of the stated value and the coordinate matrix of the projected value, a distance matrix is ​​obtained; and combined with the length scale, a scaled distance matrix is ​​obtained. The scaled distance matrix is ​​used as the input to the linear mapping layer, and the output of the linear mapping layer is residually concatenated with the distance matrix to obtain the corrected distance matrix. The corrected distance matrix is ​​normalized using Softmax, and the Hadamard product of the normalized result and the original attention score matrix is ​​calculated to obtain the double-weighted attention score matrix; wherein the original attention score matrix is ​​the outer product of the query and the projected key; Calculate the outer product of the projected value and the second-weighted attention score matrix, and use it as the current attention feature.

2. The three-dimensional medical image segmentation method as described in claim 1, characterized in that, After obtaining the segmentation results of the target region, the segmentation results are visualized and converted in format.

3. The three-dimensional medical image segmentation method as described in claim 1, characterized in that, Each level of the hierarchical decoder includes several Transformer modules, wherein the first Transformer module of each level of the hierarchical decoder is a cross Transformer module, and the remaining Transformer modules have the same structure as the Transformer modules used in the hierarchical encoder.

4. The three-dimensional medical image segmentation method as described in claim 3, characterized in that, The cross-Transformer module includes a second attention submodule and a convolution submodule. Each second attention submodule is hop-connected to the hierarchical encoder at the corresponding level. The second attention submodule includes a convolutional low-rank decomposition unit and an adaptive parameterized Euclidean matrix weighting unit.

5. A three-dimensional medical image segmentation system that does not require resampling, characterized in that, Based on a three-dimensional medical image segmentation framework, the three-dimensional medical image segmentation framework includes a hierarchical encoder and a hierarchical decoder; The three-dimensional medical image segmentation system is used to perform the three-dimensional medical image segmentation method as described in any one of claims 1 to 4, including: The data acquisition and preprocessing module is used to acquire and preprocess raw medical image data collected in slice form; wherein the preprocessing includes normalization and slicing. The feature encoding module is used to extract features from the preprocessed data using the hierarchical encoder, retaining the output features of each level and the final features extracted by the hierarchical encoder. The feature decoding module is used to take the final feature as the input of the hierarchical decoder and take each output feature as the input of the corresponding level of the hierarchical decoder to reconstruct the three-dimensional medical image layer by layer. The result acquisition module post-processes the reconstructed 3D medical image to obtain the segmentation result of the target region; wherein the post-processing includes at least block fusion.

6. A storage medium, characterized in that, It stores a computer program for three-dimensional medical image segmentation without resampling, wherein the computer program causes a computer to perform the three-dimensional medical image segmentation method as described in any one of claims 1 to 4.

7. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the three-dimensional medical image segmentation method as described in any one of claims 1 to 4.

Images