Mri image segmentation method based on wavelet transform high frequency feature guidance
Patent Information
- Application Number
- CN202611396752.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-10-30
AI Technical Summary
然而，现有频域方法在高频结构信息利用方面仍较为有限，通常未对不同高频子带进行区分利用，难以充分挖掘各高频子带所蕴含的结构特征，从而限制了模型对细粒度结构信息的表达能力
[0015]本发明的有益效果在于：提取待分割的MRI图像中的高频信息，并对高频信息进行调制获得高频特征，抑制不稳定高频响应对特征表达的干扰，使所获得的高频特征能够更加准确地表征病灶相关结构信息，为后续特征学习提供具有结构意义的引导信息；再将高频特征引入编码阶段，使其中蕴含的结构信息在多层特征学习过程中持续发挥引导作用，缓解关键结构信息在特征压缩过程中的衰减，从而提高模型对细粒度结构及复杂边界区域的表达能力。
Smart Images
Figure CN123015578A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical device technology, and in particular to an MRI image segmentation method guided by wavelet transform high-frequency features. Background Technology
[0002] Breast cancer is one of the most common malignant tumors among women worldwide, posing a serious threat to women's health. Numerous studies have shown that early detection and accurate assessment of breast lesions can significantly improve patient survival rates and treatment outcomes. In the diagnosis and treatment of breast cancer, the size, shape, boundary characteristics, and relationship with surrounding tissues of the lesion are all important bases for clinical diagnosis and treatment planning. Therefore, accurate segmentation of breast lesions has significant clinical application value.
[0003] Magnetic resonance imaging (MRI) has become an important imaging tool for breast cancer screening, diagnosis, and treatment evaluation due to its excellent soft tissue resolution and high lesion sensitivity. However, breast MRI images often suffer from problems such as complex tissue structures, uneven gray-scale distribution, blurred lesion boundaries, and imaging noise. In addition, some lesions are small and irregular in shape, and their key structural information is easily weakened or lost during network downsampling and feature aggregation, thus affecting the accurate localization and fine segmentation of the lesion area.
[0004] In recent years, deep learning-based medical image segmentation methods have made significant progress. U-Net and its variants, in particular, have effectively improved segmentation performance by achieving multi-level feature fusion through encoder-decoder structures and skip connections. To further enhance the model's ability to represent target structures, researchers have proposed methods such as multi-scale feature fusion, attention mechanisms, and feature pyramids, which achieve feature aggregation by strengthening the interaction between semantic information and spatial detail information at different levels. However, these methods typically implicitly model structural information and are prone to introducing background responses unrelated to the target, thus weakening their ability to represent fine-grained structures.
[0005] Furthermore, to enhance the model's ability to perceive structural information, some studies have introduced edge modeling methods, guiding the network to focus on target contour information through edge prediction or boundary supervision mechanisms. However, these methods are highly dependent on spatial domain boundary information, and their structural modeling effectiveness is easily limited when lesion boundaries are blurred or have low contrast.
[0006] To further enhance the model's ability to perceive structural details, some studies have introduced frequency domain representation methods. Existing methods typically utilize Fourier Transform or Wavelet Transform to obtain frequency information from images. Fourier Transform focuses on global frequency representation but lacks the ability to explicitly represent local spatial information; Wavelet Transform, on the other hand, can represent both frequency and spatial structural information, preserving the spatial location information of the image while performing frequency analysis. However, existing frequency domain methods are still relatively limited in utilizing high-frequency structural information. They often fail to differentiate between different high-frequency subbands, making it difficult to fully explore the structural features contained in each high-frequency subband, thus limiting the model's ability to express fine-grained structural information. Summary of the Invention
[0007] The purpose of this invention is to design an MRI image segmentation method based on wavelet transform high-frequency features to solve the above problems.
[0008] The present invention achieves the above objectives through the following technical solutions:
[0009] MRI image segmentation methods guided by high-frequency features of wavelet transform include:
[0010] S1. Acquire the MRI image to be segmented and preprocess the MRI image to be segmented;
[0011] S2. Frequency domain decomposition is performed on the preprocessed MRI image to extract high-frequency information, and the high-frequency information is classified and modulated. Then, the modulated high-frequency information is reconstructed back into the spatial domain to obtain high-frequency features.
[0012] S3. The image size is adjusted by performing two downsampling operations on the preprocessed MRI image through two consecutively connected first downsampling modules to obtain the adjusted MRI image.
[0013] S4. The encoder fuses the adjusted MRI image and high-frequency features, and performs structural information modeling to obtain coded features;
[0014] S5. Upsample the encoded features and output the final lesion segmentation result through the segmentation head.
[0015] The beneficial effects of this invention are as follows: high-frequency information is extracted from the MRI image to be segmented, and the high-frequency information is modulated to obtain high-frequency features, suppressing the interference of unstable high-frequency responses on feature expression, so that the obtained high-frequency features can more accurately characterize the lesion-related structural information, providing structurally meaningful guiding information for subsequent feature learning; then the high-frequency features are introduced into the encoding stage, so that the structural information contained therein continues to play a guiding role in the multi-layer feature learning process, alleviating the attenuation of key structural information in the feature compression process, thereby improving the model's ability to express fine-grained structures and complex boundary regions. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the MRI image segmentation method based on wavelet transform high-frequency features guided by the present invention.
[0017] Figure 2 This is a schematic diagram illustrating the segmentation results of this method using a breast MRI dataset;
[0018] Figure 3 This diagram illustrates the segmentation results of this method using the publicly available dataset BraTS2019. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0022] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0023] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.
[0024] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0026] like Figure 1 As shown, the MRI image segmentation method based on wavelet transform high-frequency features includes:
[0027] S1. Acquire the MRI images to be segmented and preprocess them. Preprocessing includes resampling, data normalization, and cropping to unify the spatial resolution and grayscale distribution of images from different patients and improve the stability of subsequent feature learning.
[0028] S2. Frequency domain decomposition is performed on the preprocessed MRI image to extract high-frequency information, and the high-frequency information is classified and modulated. The modulated high-frequency information is then reconstructed back into the spatial domain to obtain high-frequency features.
[0029] S21. The preprocessed MRI image is decomposed in the frequency domain using discrete wavelet transform, resulting in a low-frequency sub-band and multiple high-frequency sub-bands, as shown below: ,in, The images are preprocessed MRI images. For low-frequency sub-band, , , , , , and This represents a sub-band component containing high-frequency responses along different spatial axes, where H and L indicate high or low frequencies along each spatial axis, respectively.
[0030] S22. Select a first sub-band and a second sub-band from the high-frequency sub-bands. The first sub-band and the second sub-band contain different numbers of spatial axes corresponding to their high-frequency responses. Select the high-frequency sub-band containing only a single spatial axis high-frequency response as the first sub-band; select the high-frequency sub-band containing only two spatial axis high-frequency responses as the second sub-band. Since the high-frequency sub-bands... It contains high-frequency responses along three axes, with strong frequency fluctuations and relatively little stable structural information. Therefore, this method mainly uses the first and second subbands for subsequent structural modeling.
[0031] S23. Calculate the structural adaptive modulation coefficient using the first sub-band. , is represented as: , , where k is a hyperparameter used to control the modulation intensity; the modulation coefficients act on the first and second subbands to suppress the interference of excessive high-frequency response on feature expression and retain structural information with spatial consistency; S represents the structural response intensity, used to measure the degree of local structural change; the larger the value of S, the stronger the high-frequency structural response at the current location.
[0032] S24. Adaptive modulation is performed on the first and second subbands using the structure-adaptive modulation coefficients to obtain the modulated first and second subbands, as follows: , , where represents the first or second subband after modulation, and ⊙ represents element-level multiplication. When the local high-frequency response is too strong, the modulation coefficient is reduced accordingly to suppress the interference of noise and abnormally strong responses on the structural expression. When the local high-frequency response is weak, the modulation coefficient is increased accordingly to preserve effective structural information and fine-grained structural features with spatial consistency.
[0033] S25. Through inverse discrete wavelet transform The modulated first subband and modulated second subband are reconstructed back into the spatial domain to obtain the first feature and the second feature as high-frequency features. , is represented as: .
[0034] This method is based on the axial correlation structure modulation mechanism of wavelet transform. It uses three-dimensional discrete wavelet transform to obtain high-frequency structural information in the image, and constructs high-frequency feature representations with different axial response characteristics according to the spatial axial relationship corresponding to the high-frequency response. At the same time, it uses modulation to suppress the interference of unstable high-frequency response on feature expression, so that the obtained high-frequency features can more effectively retain the structural information related to the lesion, and provide structurally meaningful guiding information for subsequent feature learning.
[0035] S2': The high-frequency features are gradually adjusted to the same spatial size as the input features input to the first, second, third, and fourth encoding stages, so as to facilitate subsequent feature fusion.
[0036] S3. The preprocessed MRI image is processed through two consecutively connected first downsampling modules. The image size was adjusted by performing two downsampling operations to obtain the adjusted MRI image;
[0037] S4. The encoder models the structural information of the adjusted MRI image and high-frequency features to obtain encoded features;
[0038] The encoder consists of four encoding stages from input to output: a first encoding stage, a second encoding stage, a third encoding stage, and a fourth encoding stage. High-frequency features are input into each of these stages. The first encoding stage includes two cascaded local structure enhancement modules and a second downsampling module. The second encoding stage includes four local structure enhancement modules, two global feature aggregation modules, and a second downsampling module. The third encoding stage includes two local structure enhancement modules, four global feature aggregation modules, and a second downsampling module. The fourth encoding stage includes three cascaded global feature aggregation modules and a second downsampling module. High-frequency features are input into each of the local structure enhancement modules and each of the global feature aggregation modules in the first, second, third, and fourth encoding stages.
[0039] The local structure enhancement module fuses high-frequency features with input features and models local structural information through axial convolution to obtain local structure enhancement features. The global feature aggregation module fuses high-frequency features with input features and models global structural information through global pooling to obtain global aggregated features. The concatenation layer is used to concatenate the local structure enhancement features output by the last local structure enhancement module of the second or third encoding stage and the global aggregated features of the last global feature aggregation module in a channel-wise concatenation manner to obtain concatenated features. The second downsampling module is used to downsample the local result enhancement features output by the last local structure enhancement module of the first encoding stage, the concatenated features of the second encoding stage, the concatenated features of the third encoding stage, or the global aggregated features output by the last global feature aggregation module of the fourth encoding stage to obtain first downsampled features, second downsampled features, third downsampled features, or fourth downsampled features. The first downsampled features, second downsampled features, third downsampled features, and fourth downsampled features are all used as encoded features output by the encoder.
[0040] The local structure enhancement module specifically processes the input features, including:
[0041] The local structure enhancement module utilizes the axially correlated structural information contained in high-frequency features to guide convolutional branches along different spatial planes to learn the local structural information of the input features, thereby obtaining planar local structural features corresponding to different spatial planes. , is represented as: The planar local structural features corresponding to all different spatial planes are concatenated along the channel dimension, and feature mapping is performed through pointwise convolution to obtain local structural enhancement features. , is represented as: , ; To provide local information on different axes in the channel dimension The splicing process yields the splicing features; The input for this module is the image features; There are three orthogonal spatial planes; This is a convolution operation along the i-th spatial plane; The first feature contains only the high-frequency response of the orthogonal direction of the i-th spatial plane; This is the second feature corresponding to the i-th spatial plane; This is a channel-level splicing operation. This is a pointwise convolution operation;
[0042] The global feature aggregation module processes the input features by using high-frequency features to guide the model in learning the global structural information of the input features in different planes, thereby obtaining three axial global features along different spatial directions. , is represented as: The three axial global features along different spatial directions are stacked, and the Sigmoid function is used to generate gating coefficients to weight the input features, thus obtaining the global aggregated features. , is represented as: , ; The input image features are used as input to this module; STACK indicates stacking the features along a new dimension. This represents the Sigmoid activation function;
[0043] By embedding the local structure enhancement module and the global feature aggregation module into different encoding stages of the encoder, the key structural information is more fully preserved during the downsampling process through the synergistic effect of these two modules, providing richer structural feature representations for the subsequent decoding stage.
[0044] Step S4 specifically includes:
[0045] S41. Use the adjusted MRI image as the input feature for the first coding stage. The high-frequency features are input together into the first encoding stage, which processes the fused input features. The high-frequency features are processed to obtain the first downsampled features;
[0046] S42. Use the first downsampled features as the input features for the second encoding stage. The input features are fed into the second encoding stage along with the high-frequency features. The high-frequency features are divided into two parts along the channel. The first part, after adding the input features to the corresponding high-frequency features, passes through four consecutive local structure enhancement modules. Except for the last module, the output features of each module are added to the high-frequency features before being input to the next module. The second part, after adding the input features to the corresponding high-frequency features, passes through two consecutive global feature aggregation modules. Again, except for the last module, the output features of each module are added to the high-frequency features before being input to the next module. Then, the local structure enhancement features output from the fourth local structure enhancement module and the global aggregation features output from the second global feature aggregation module are input together into the concatenation layer along the channel dimension for concatenation. Finally, a second downsampling module, consisting of pooling and convolution, performs a downsampling operation to reduce the dimensionality, resulting in the second downsampled features.
[0047] S43. Use the second downsampled feature as the input feature for the third encoding stage. The input features are fed into the third encoding stage along with the high-frequency features. The high-frequency features are divided into two parts along the channel. The first part, after adding the input features to the corresponding high-frequency features, passes through two consecutive local structure enhancement modules. Except for the last module, the output features of each module are added to the high-frequency features before being input to the next module. The second part, after adding the input features to the corresponding high-frequency features, passes through four consecutive global feature aggregation modules. Again, except for the last module, the output features of each module are added to the high-frequency features before being input to the next module. Then, the local structure enhancement features output from the second local structure enhancement module and the global aggregation features output from the fourth global feature aggregation module are input together into the concatenation layer along the channel dimension for concatenation. Finally, a second downsampling module, consisting of pooling and convolution, performs downsampling to reduce the dimensionality, resulting in the third downsampled feature.
[0048] S44. Use the third downsampled feature from the third encoding stage as the input feature for the fourth encoding stage. The high-frequency features are input together into the fourth encoding stage, which then processes the fused input features. The high-frequency features are processed to obtain the fourth downsampled feature; the first downsampled feature, the second downsampled feature, the third downsampled feature and the fourth downsampled feature are all used as the encoded features output by the encoder.
[0049] This method is based on an axially correlated high-frequency feature-guided encoding mechanism. It introduces high-frequency features into the encoding stage and uses two methods, local structure enhancement and global feature aggregation, to ensure that structure-related high-frequency information continues to participate in feature extraction during the step-by-step downsampling process. This alleviates the attenuation of key structural information during feature compression and improves the model's ability to express fine-grained structures and complex boundary regions.
[0050] S5. Upsample the encoded features and output the final lesion segmentation result through the segmentation head. Specifically, all encoded features output by the encoder are mapped to a unified channel dimension through convolution, and adjusted to a unified spatial resolution through interpolation. Then, they are fused to obtain fused features. The fused features are then upsampled step by step, and after each upsampling, they are spliced and fused with the encoded features of the corresponding scale generated in S3 to gradually restore the spatial resolution until it is restored to the same spatial size as the MRI image to be segmented. Finally, the final lesion segmentation result is output through the segmentation head.
[0051] By constructing an axially correlated high-frequency structural representation and continuously guiding the encoder's feature learning using this representation, an effective combination of frequency domain structural information and deep feature learning is achieved. Compared with existing technologies, this invention can more fully utilize structurally significant high-frequency information, enhancing the model's ability to represent the fine-grained structure and boundary information of MRI lesions, providing rich structural information for small lesion segmentation, and possessing significant practical value.
[0052] Example 1
[0053] The method of this invention was validated using a breast MRI dataset. First, the MRI images underwent preprocessing operations such as resampling, normalization, and resizing. Then, a medical image segmentation network guided by wavelet high-frequency features was constructed according to the steps described in this invention. During training, Dice loss and Cross Entropy loss were jointly optimized to enhance the ability to distinguish between lesion regions and background regions.
[0054] After training, the segmentation performance was evaluated on the test set using the Dice coefficient and Intersection over Union (IoU). As shown in Table 1, the experimental results demonstrate that introducing high-frequency axially correlated features enhances the model's ability to represent lesion details, thereby obtaining more accurate segmentation results. Figure 2 As shown, the effectiveness of the technical solution of the present invention has been verified.
[0055] Example 2
[0056] The method of this invention was validated using the publicly available BraTS2019 dataset. A segmentation model was constructed following the same training procedure as in Example 1, and the Dice coefficient and IoU metric were used for evaluation. As shown in Table 2, the results demonstrate that this invention can still achieve good segmentation results on publicly available medical image datasets. Figure 3 As shown, this demonstrates that the technical solution of the present invention has good applicability and a certain degree of generalization ability, further verifying the effectiveness of the technical solution of the present invention.
[0057] Table 1 shows the evaluation of the segmentation performance of this method using a breast MRI dataset.
[0058]
[0059] Table 2 is an evaluation table of the segmentation performance of this method using the publicly available dataset BraTS2019.
[0060]
[0061] This method enhances the model's ability to express and perceive high-frequency structural information by explicitly modeling local structural information. It fully mines the structural information contained in high-frequency features, improves the model's ability to represent fine-grained structural changes, and addresses the problem of insufficient response to small-scale lesions in existing methods. It also alleviates the problem of gradual decay of key structural information during encoding and progressive downsampling, enabling structure-related high-frequency features to continue to play a guiding role in the multi-layer feature transmission process, thereby improving the model's localization accuracy, boundary characterization ability, and segmentation stability of lesion regions. Furthermore, it enhances the model's ability to model fine-grained structural information and improves its feature expression ability for small-scale lesions and complex boundary regions, thus improving the segmentation effect of breast MRI lesions.
[0062] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.
Claims
1. A wavelet transform-guided MRI image segmentation method, characterized in that, include: S1. Acquire the MRI image to be segmented and preprocess the MRI image to be segmented; S2. Frequency domain decomposition is performed on the preprocessed MRI image to extract high-frequency information, and the high-frequency information is classified and modulated. The modulated high-frequency information is then reconstructed back into the spatial domain to obtain high-frequency features. S3. The image size is adjusted by performing two downsampling operations on the preprocessed MRI image through two consecutively connected first downsampling modules to obtain the adjusted MRI image. S4. The encoder fuses the adjusted MRI image and high-frequency features, and performs structural information modeling to obtain coded features; S5. Upsample the encoded features and output the final lesion segmentation result through the segmentation head.
2. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 1, characterized in that, In S1, preprocessing includes resampling, data normalization, and pruning.
3. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 1, characterized in that, S2 includes: S21. Use discrete wavelet transform to perform frequency domain decomposition on the preprocessed MRI image, decomposing the preprocessed MRI image into a low-frequency sub-band and multiple high-frequency sub-bands. S22. Select a first sub-band and a second sub-band from the high-frequency sub-bands. The first sub-band and the second sub-band contain different numbers of spatial axes corresponding to the high-frequency responses. S23. Calculate the adaptive modulation coefficient of the structure using the first sub-band; S24. Adaptive modulation of the first sub-band and the second sub-band is performed using the structural adaptive modulation coefficients to obtain the modulated first sub-band and the modulated second sub-band. S25. The modulated first sub-band and the modulated second sub-band are reconstructed back into the spatial domain by inverse discrete wavelet transform, and the first feature and the second feature are obtained as high-frequency features.
4. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 3, characterized in that, Structural Adaptive Modulation Coefficient Represented as: , The first subband after modulation or the second subband after modulation Represented as: Where k is a hyperparameter used to control the modulation intensity, and ⊙ represents element-wise multiplication. It can be either the first sub-band or the second sub-band; This represents the structural response intensity.
5. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 3, characterized in that, The encoder consists of four encoding stages from input to output: a first encoding stage, a second encoding stage, a third encoding stage, and a fourth encoding stage. High-frequency features are input into each of these stages. The first encoding stage includes multiple cascaded local structure enhancement modules and a second downsampling module. The second and third encoding stages each include multiple cascaded local structure enhancement modules, multiple cascaded global feature aggregation modules, a splicing layer, and a second downsampling module. The fourth encoding stage includes multiple cascaded global feature aggregation modules and a second downsampling module. High-frequency features are input into each local structure enhancement module and each global feature aggregation module in each of the first, second, third, and fourth encoding stages. The local structure enhancement module fuses high-frequency features with input features and models local structure information through axial convolution to obtain local structure enhancement features; The global feature aggregation module fuses high-frequency features with input features and obtains global aggregated features by modeling global structural information through global pooling. The splicing layer is used to splice the local structure enhancement features output by the last local structure enhancement module of the second or third encoding stage and the global aggregation features of the last global feature aggregation module in a channel splicing manner to obtain spliced features; the downsampling module is used to perform downsampling operations on the local structure enhancement features output by the last local structure enhancement module of the first encoding stage, the spliced features of the second encoding stage, the spliced features of the third encoding stage, or the global aggregation features output by the last global feature aggregation module of the fourth encoding stage to obtain first downsampling features, second downsampling features, third downsampling features, or fourth downsampling features; the first downsampling features, second downsampling features, third downsampling features, and fourth downsampling features are all used as encoded features output by the encoder.
6. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 5, characterized in that, The local structure enhancement module specifically processes the input features, including: The local structure enhancement module utilizes the axially related structural information contained in high-frequency features to guide the convolutional branches along different spatial planes to learn the local structural information of the input features, thereby obtaining planar local structural features corresponding to different spatial planes. The planar local structural features corresponding to all different spatial planes are concatenated in the channel dimension, and feature mapping is performed through pointwise convolution to obtain local structure enhancement features. The global feature aggregation module processes the input features in the following ways: it uses high-frequency features to guide the model to learn the global structural information of the input features in different planes, and obtains three axial global features along different spatial directions; it stacks the three axial global features along different spatial directions, and uses the Sigmoid function to generate gating coefficients to weight the input features, thereby obtaining the global aggregated features.
7. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 5, characterized in that, Between S2 and S3, there is also S2', which adjusts the high-frequency features step by step to the same spatial size as the input features input to the first coding stage, the second coding stage, the third coding stage, and the fourth coding stage.
8. The MRI image segmentation method based on wavelet transform high-frequency features guided according to claim 1, characterized in that, In S4, all encoded features are mapped to a uniform channel dimension through convolution, and then adjusted to a uniform spatial resolution through interpolation before being fused to obtain fused features.