Complex model clinical image semantic segmentation method based on frequency phase guidance

Abstract

The invention discloses a complex model clinical image semantic segmentation method based on frequency phase guidance. The method comprises the following steps: preprocessing an original clinical image; the data are input into a Gabor wavelet transformation module to be converted into a complex value domain image; inputting the complex-value-domain image into a coding network, and selectively adjusting perception of a phase component and preference for different frequency components through a complex-value mixed attention module to obtain a complex-value-domain feature; inputting the complex-value-domain features into a frequency adaptive separation module, decomposing the complex-value-domain features into high-frequency features and low-frequency features, combining the low-frequency features and the high-frequency features, and screening out important features through a frequency channel attention module; and inputting the important features into a next-level encoder, and combining the high-frequency features with the features of each level of the decoder to obtain a segmentation result of the clinical image. According to the method, the phase information and the frequency information are effectively utilized, the target area and the similar background in the clinical image can be better distinguished, and the segmentation effect is improved.

Description

Technical Field

[0001] The present invention belongs to the field of image segmentation, relates to computer vision technology, and specifically relates to a clinical image semantic segmentation method based on a complex model guided by frequency and phase. Background Art

[0002] Accurate clinical image segmentation is of vital importance in many clinical tasks. However, the process of manually outlining tumors or organs involves expertise in multiple fields and requires analysis of multiple examination sequences, which is subjective and time-consuming. The introduction of automated segmentation methods can significantly reduce the workload of doctors while improving the consistency and accuracy of analysis. With the help of automated segmentation technology, doctors can not only speed up the diagnosis process, but also formulate treatment plans and monitor disease changes more quickly.

[0003] Recently, research based on data-driven deep learning has significantly advanced the automatic segmentation of clinical images. Deep learning methods, such as convolutional neural networks (CNNs) and Transformers, have been widely used in clinical image segmentation tasks, bringing many advantages to medical image analysis. Deep learning methods can automatically learn features and regularities, eliminating the need to manually design feature extractors, thereby reducing the need for human intervention and improving the degree of automation. However, when faced with the challenge of obvious similarities and distribution proximity between clinical diseases, many current deep learning architectures still have some limitations. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, a method for clinical image semantic segmentation based on a complex model guided by frequency and phase is provided. The designed segmentation model can effectively utilize phase information and frequency information, and can better distinguish the target area in the clinical image from the similar background, so as to assist the clinician in better screening for early lesions.

[0005] Technical solution: To achieve the above purpose, the present invention provides a method for semantic segmentation of clinical images based on a complex model guided by frequency and phase, comprising the following steps:

[0006] S1: Preprocessing of original clinical images;

[0007] S2: input the preprocessed image data obtained in step S1 into a Gabor wavelet transform module to convert it into a complex-valued domain image;

[0008] S3: Input the complex-valued domain image obtained in step S2 into the encoding network, and selectively adjust the perception of the phase component and the preference for different frequency components through the complex-valued mixed attention module to obtain the complex-valued domain feature;

[0009] S4: Input the complex-valued domain features obtained in step S3 into the constructed frequency adaptive separation module, decompose the complex-valued domain features into high-frequency features and low-frequency features, and combine the low-frequency features with the high-frequency features to filter out important features through the frequency channel attention module;

[0010] S5: The important features obtained in step S4 are input into the next level encoder, and the high-frequency features are merged with the features of each level of the decoder. The real and imaginary features are integrated at the last layer of the decoder to obtain the segmentation result of the clinical image.

[0011] Furthermore, the preprocessing method in step S1 is: first normalize to reduce the range of pixel values, and then use the data enhancement strategy to expand the data set, specifically:

[0012] A1: Change the image size to the same size through interpolation so that it can be uniformly input into the network training later;

[0013] A2: Normalize the pixel value range of the image to 0-1 to reduce the interference caused by uneven lighting in clinical images;

[0014] A3: Use three data augmentation methods, namely flipping, rotating and skewing, to expand the data volume.

[0015] Furthermore, the Gabor wavelet transform module in step S2 is composed of a filter group of 4 scales and 5 directions, and a set of Gabor filters is constructed to convert the clinical image into a complex value domain, specifically including:

[0016] B1: For the filter parameters: scale λ and direction θ, take 4 and 5 values ​​uniformly within a certain range to construct a set of Gabor filters. The specific formula is as follows:

[0017]

[0018] Where x′=xcos(θ)+ysin(θ), y′=-xsin(θ)+ycos(θ), θ represents the tilt angle of the Gabor kernel function; λ and ψ represent the filter wavelength and phase offset, respectively; the standard deviation σ and the spatial aspect ratio δ determine the ellipticity of the Gabor kernel function;

[0019] B2: Filter the input image through the constructed Gabor filter set to obtain an image in the complex value domain.

[0020] Furthermore, in step S3, the encoding network adopts a UNet network, and the UNet network is modified into a complex form, including expressing the convolution kernel, activation function and batch normalization in complex form, as follows:

[0021] Complex convolution is represented as the convolution operation between the complex feature map h and the convolution filter W, where h = a + ib, W = A + iB; according to the definition of complex operation, the calculation process of complex convolution is expressed as:

[0022] W*h=(A*aB*b)+i(B*a+A*b);

[0023] The complex activation ReLU satisfies the Cauchy-Riemann equation as follows:

[0024]

[0025] Batch normalization for complex input z is calculated as follows:

[0026]

[0027] in, is the normalization of z, V is the covariance matrix and E[·] is the mean of the input z.

[0028] Furthermore, the complex-valued hybrid attention module in step S3 includes a cross-phase fusion module and an amplitude deviation interference module, which first passes through the cross-phase fusion module and then passes through the amplitude deviation interference module to obtain the complex-valued domain features.

[0029] Furthermore, the operation mode of the cross-phase fusion module in step S3 is:

[0030] C1: Apply 3×3 and 1×1 convolution operations to obtain feature F respectively. 3 and F 1 , and then these features are fused together, a 1×1 convolution and a S-type activation function σ(·) are performed to generate an attention map β representing the importance of local features, as shown below:

[0031] β=σ(Conv 1×1 (F 3 +F 1 ));

[0032] C2: 1-β is expressed as β′ at the same time, and these two soft attention maps are used to automatically select more representative features from each receptive field, as shown below:

[0033] F c =cat(β′×F 3 ,β×F 1 ).

[0034] Furthermore, the operation mode of the amplitude deviation interference module in step S3 is:

[0035] D1: The complex-valued features output from step C2 and the input of the cross-phase fusion module are added pixel by pixel and then passed through 3×3 as the input of the amplitude deviation interference module;

[0036] D2: Apply a 1×1 convolution operation to the output of step D1 to reduce the channel dimension and split it into two independent processing paths;

[0037] D3: The upper branch uses a deep separable volume to extract global context information;

[0038] D4: The lower branch features are separated into amplitude and phase components, and then the amplitude information F is processed by applying convolution with controlled noise. s A slight perturbation is added, and the perturbed amplitude information is recombined with the phase information, as shown below:

[0039] Amplitude=Conv 1×1 ·(Conv 1×1 ·|F s |),

[0040] F complex =Amplitude×cos(arctan(|F s |))+j·Amplitude×sin(arctan(|F s |));

[0041] D5: F complex As a weight multiplied with the upper branch, the interference feature vector is obtained and the original channel dimension is restored through a 3×3 convolution operation.

[0042] The complex-valued network in step S3 is proven to be a data-driven nonlinear wavelet packet. The image after Gabor wavelet transformation can be easily approximated by a complex-valued network. In addition, complex-valued networks can retain and process complex-valued features, so they can retain and restore phase information, making the complex-valued network more sensitive to phase structure. The hybrid attention module first adaptively adjusts the receptive field to select important features through the cross-phase fusion module combined with multi-scale convolution. Secondly, the amplitude deviation interference module is used to solve the problem of the network's over-reliance on specific frequencies.

[0043] Furthermore, the operation process of the frequency adaptive separation module in step S4 is as follows:

[0044] E1: Perform a first-level wavelet decomposition on the feature map along the channel dimension to obtain low-frequency coefficients and high frequency coefficients

[0045] E2: Replace the low-frequency coefficients with zero matrix and perform inverse wavelet transform together with the high-frequency coefficients to obtain the high-frequency feature F H ∈RC×H×W ;

[0046] E3: Use zero values ​​to replace high-frequency coefficients and perform inverse wavelet transform to obtain low-frequency features F L ∈R C×H×W , and the high-frequency feature F H and low frequency features F L Merge into F cat Input frequency channel attention module;

[0047] E4: F cat Divide it into n parts according to the channel, denoted as Then to F i A DCT operation is performed and the stitching is done together to get the overall compressed features as follows:

[0048] F dct =cat([DCT(F 0 ),DCT(F 1 ),…,DCT(F n-1 )]);

[0049] E5: F dct Apply the fully connected operation and activation operation, and then with F cat Multiply to get F hybrid ∈R 2C×H×W .

[0050] The frequency adaptive separation module in step S4 is designed based on the theory that deep learning models tend to learn specific frequencies according to specific tasks. In the segmentation task, the encoder first learns lower frequencies, which contain most of the basic semantic information related to the overall representation. In contrast, the segmentation task requires the decoder to retain the required high-frequency information containing fine-grained details, thereby emphasizing the extraction of subtle changes and contours. The frequency adaptive separation module first utilizes wavelet decomposition to automatically separate high-frequency and low-frequency features, so that specific high-frequency information can be selectively utilized to be transmitted to the decoder through skip connections. In addition, the frequency adaptive separation module is equipped with a channel attention module to refine the most representative frequency features to the next layer of encoder.

[0051] Furthermore, the step S5 is specifically as follows:

[0052] F1: Feature F obtained in step E5 hybrid Change the number of channels to C through 1×1 convolution, which is used as the input of the next encoder;

[0053] F2: The high-frequency feature F obtained in step E2 H Transmitted to the corresponding decoder stage via skip connections to supplement fine-grained information;

[0054] F3: Integrate the real and imaginary features of the network output to obtain the final segmentation feature map. The specific formula is as follows:

[0055]

[0056] Among them, F final , F real and F imag Represent the final segmentation features, real features and imaginary features respectively.

[0057] Furthermore, in step F3, the segmentation result is obtained by activating the Softmax function, and the calculation formula of Softmax is:

[0058]

[0059] The method of the present invention explores the properties of complex-valued models guided by phase and frequency perspectives. Complex-valued neural networks can directly process phase components, providing a potential solution for better perceptual structure for clinical image segmentation. The present invention designs a complex hybrid attention module to selectively adjust the perception of the phase component and the deviation of the model. In addition, the present invention designs a frequency adaptive separation module that uses a module combining wavelet decomposition and frequency channel attention to emphasize the frequency features that are preferentially processed by the encoder and decoder.

[0060] The present invention proposes a method for clinical image semantic segmentation based on a complex model guided by frequency phase, which provides a new idea for clinical image segmentation. This is an end-to-end architecture, and makes full use of phase information to capture the structural features of the image, further improving the effectiveness of the method.

[0061] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0062] 1. The method of the present invention can realize automatic segmentation of clinical images based on deep learning, thereby avoiding time-consuming manual heuristic learning and improving the generalization ability of the model.

[0063] 2. The method of the present invention can make full use of the phase information of clinical images by exploring the properties of complex-valued models under the guidance of phase and frequency perspectives; the complex-valued neural network can directly process the phase component and explore methods suitable for intelligent analysis of clinical images; by utilizing the structural information of the phase and the network's preference for specific frequency components, the method successfully solves the problems of blurred boundaries and difficulty in distinguishing target areas from similar backgrounds in clinical image segmentation, thereby improving the semantic segmentation effect of clinical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flowchart of the method of the present invention;

[0065] Figure 2 It is a structural framework diagram of the method of the present invention applied to automatic segmentation of targets in clinical images;

[0066] Figure 3 Taking breast ultrasound image as an example, the comparison result of preprocessed image and original image is shown;

[0067] Figure 4 Taking a breast ultrasound image as an example, the result diagram of the comparison between the complex-valued domain image obtained after processing in step S2 and the real-valued image obtained in step S1;

[0068] Figure 5 is an architecture diagram of the frequency adaptation module in step S4;

[0069] Figure 6 This is a diagram showing the segmentation result of the complex model guided by the frequency phase of the input image, taking breast ultrasound image as an example. DETAILED DESCRIPTION

[0070] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0071] like Figure 1 and Figure 2 As shown, the present invention provides a complex model clinical image semantic segmentation method based on frequency phase guidance, comprising the following steps:

[0072] S1: Preprocessing of original clinical images;

[0073] The preprocessing method in step S1 is: first normalize to reduce the range of pixel values, and then use the data enhancement strategy to expand the data set, specifically:

[0074] A1: Change the image size to the same size through interpolation so that it can be uniformly input into the network training later;

[0075] A2: Normalize the pixel value range of the image to 0-1 to reduce the interference caused by uneven lighting in clinical images;

[0076] A3: Use three data augmentation methods, namely flipping, rotating and skewing, to expand the data volume.

[0077] In this embodiment, a breast ultrasound image is taken as an example. The comparison result between the pre-processed image and the original image is as follows: Figure 3 shown.

[0078] S2: input the preprocessed image data obtained in step S1 into a Gabor wavelet transform module to convert it into a complex-valued domain image;

[0079] The Gabor wavelet transform module is composed of filter banks with 4 scales and 5 directions, which can cover the family of Gabor wavelet transforms and well describe the local structural information corresponding to scale and direction selectivity;

[0080] By constructing a set of Gabor filters, the clinical image is converted to the complex domain and the phase information of the image is restored, while maintaining good localization characteristics in both the time domain and the frequency domain, including:

[0081] B1: For the filter parameters: scale λ and direction θ, take 4 and 5 values ​​uniformly within a certain range to construct a set of Gabor filters. The specific formula is as follows:

[0082]

[0083] Where x′=xcos(θ)+ysin(θ), y′=-xsin(θ)+ycos(θ), θ represents the tilt angle of the Gabor kernel function; λ and ψ represent the filter wavelength and phase offset, respectively; the standard deviation σ and the spatial aspect ratio δ determine the ellipticity of the Gabor kernel function;

[0084] B2: Filter the input image through the constructed Gabor filter set to obtain an image in the complex value domain.

[0085] This embodiment takes a breast ultrasound image as an example. The comparison result of the complex-valued domain image obtained after processing in step S2 and the real-valued image obtained in step S1 is as follows: Figure 4 shown.

[0086] S3: Input the complex-valued domain image obtained in step S2 into the encoding network, and selectively adjust the perception of the phase component and the preference for different frequency components through the complex-valued mixed attention module to obtain the complex-valued domain feature;

[0087] The encoding network uses the UNet network, which is modified into a complex form, including the convolution kernel, activation function and batch normalization are all expressed in complex form, as follows:

[0088] Complex convolution is represented as the convolution operation between the complex feature map h and the convolution filter W, where h = a + ib, W = A + iB; according to the definition of complex operation, the calculation process of complex convolution is expressed as:

[0089] W*h=(A*aB*b)+i(B*a+A*b);

[0090] The complex activation ReLU satisfies the Cauchy-Riemann equation as follows:

[0091]

[0092] Batch normalization for complex input z is calculated as follows:

[0093]

[0094] in, is the normalization of z, V is the covariance matrix and E[·] is the mean of the input z.

[0095] The complex-valued hybrid attention module includes a cross-phase fusion module and an amplitude deviation interference module. The complex-valued domain features are obtained by first passing through the cross-phase fusion module and then through the amplitude deviation interference module.

[0096] Reference Figure 2 , the operation mode of the cross-phase fusion module is:

[0097] C1: Apply 3×3 and 1×1 convolution operations to obtain feature F respectively. 3 and F 1 , and then these features are fused together, a 1×1 convolution and a S-type activation function σ(·) are performed to generate an attention map β representing the importance of local features, as shown below:

[0098] β=σ(Conv 1×1 (F 3 +F 1 ));

[0099] C2: 1-β is expressed as β′ at the same time, and these two soft attention maps are used to automatically select more representative features from each receptive field, as shown below:

[0100] F c =cat(β′×F 3 ,β×F 1 ).

[0101] Reference Figure 2 , the operation mode of the amplitude deviation interference module is:

[0102] D1: The complex-valued features output from step C2 and the input of the cross-phase fusion module are added pixel by pixel and then passed through 3×3 as the input of the amplitude deviation interference module;

[0103] D2: Apply a 1×1 convolution operation to the output of step D1 to reduce the channel dimension and split it into two independent processing paths;

[0104] D3: The upper branch uses a deep separable volume to extract global context information;

[0105] D4: The lower branch features are separated into amplitude and phase components, and then the amplitude information F is processed by applying convolution with controlled noise. s A slight perturbation is added, and the perturbed amplitude information is recombined with the phase information, as shown below:

[0106] Amplitude=Conv 1×1 ·(Conv 1×1 ·|F s |),

[0107] F complex =Amplitude×cos(arctan(|F s |))+j·Amplitude×sin(arctan(|F s |));

[0108] D5: F complex As a weight multiplied with the upper branch, the interference feature vector is obtained and the original channel dimension is restored through a 3×3 convolution operation.

[0109] S4: Input the complex-valued domain features obtained in step S3 into the constructed frequency adaptive separation module, decompose the complex-valued domain features into high-frequency features and low-frequency features, and combine the low-frequency features with the high-frequency features to filter out important features through the frequency channel attention module;

[0110] Reference Figure 5 , the operation process of the frequency adaptive separation module is:

[0111] E1: Perform a first-level wavelet decomposition on the feature map along the channel dimension to obtain low-frequency coefficients and high frequency coefficients

[0112] E2: Replace the low-frequency coefficients with zero matrix and perform inverse wavelet transform together with the high-frequency coefficients to obtain the high-frequency feature F H ∈R C×H×W ;

[0113] E3: Use zero values ​​to replace high-frequency coefficients and perform inverse wavelet transform to obtain low-frequency features F L ∈R C×H×W , and the high-frequency feature F H and low frequency features F L Merge into F cat Input frequency channel attention module;

[0114] E4: F cat Divide it into n parts according to the channel, denoted as Then to Fi A DCT operation is performed and the stitching is done together to get the overall compressed features as follows:

[0115] F dct =cat([DCT(F 0 ),DCT(F 1 ),…,DCT(F n-1 )]);

[0116] E5: F dct Apply the fully connected operation and activation operation, and then with F cat Multiply to get F hybrid ∈R 2C×H×W .

[0117] S5: input the important features obtained in step S4 into the next level encoder, merge the high-frequency features with the features of each level of the decoder, integrate the real and imaginary features at the last level of the decoder, and obtain the segmentation result of the clinical image;

[0118] Step S5 is specifically as follows:

[0119] F1: Feature F obtained in step E5 hybrid Change the number of channels to C through 1×1 convolution, which is used as the input of the next encoder;

[0120] F2: The high-frequency feature F obtained in step E2 H Transmitted to the corresponding decoder stage via skip connections to supplement fine-grained information;

[0121] F3: Integrate the real and imaginary features of the network output to obtain the final segmentation feature map. The specific formula is as follows:

[0122]

[0123] Among them, F final , F real and F imag Represent the final segmentation features, real features and imaginary features respectively.

[0124] In step F3, the segmentation result is obtained by activating the Softmax function. The calculation formula of Softmax is:

[0125]

[0126] This embodiment takes breast ultrasound images as an example. The segmentation results of the input image after the complex model guided by frequency phase are as follows: Figure 6 As shown in the figure, its Dice and Jaccard values ​​reached 0.914 and 0.842 respectively, which shows that the segmentation effect is very good.

[0127] In summary, the present invention is suitable for automatic segmentation of clinical images. This innovative deep learning method is based on a frequency-phase guided complex model, which provides a new idea for the semantic segmentation of clinical images, makes full use of the structural information in the image, realizes the rapid automatic segmentation of clinical images, and improves the segmentation effect.

[0128] This embodiment also provides a complex model clinical image semantic segmentation system based on frequency phase guidance, which includes a network interface, a memory and a processor; wherein the network interface is used to realize the reception and transmission of signals in the process of sending and receiving information between other external network elements; the memory is used to store computer program instructions that can be run on the processor; the processor is used to execute the steps of the above-mentioned consensus method when running the computer program instructions.

[0129] The present embodiment also provides a computer storage medium, which stores a computer program, and the method described above can be implemented when the processor executes the computer program. The computer readable medium can be considered to be tangible and non-temporary. Non-limiting examples of non-temporary tangible computer-readable media include non-volatile memory circuits (such as flash memory circuits, erasable programmable read-only memory circuits or mask read-only memory circuits), volatile memory circuits (such as static random access memory circuits or dynamic random access memory circuits), magnetic storage media (such as analog or digital tapes or hard drives) and optical storage media (such as CDs, DVDs or Blu-ray discs), etc. The computer program includes processor executable instructions stored on at least one non-temporary tangible computer-readable medium. The computer program may also include or rely on stored data. The computer program may include a basic input / output system (BIOS) that interacts with the hardware of a special-purpose computer, a device driver that interacts with a specific device of a special-purpose computer, one or more operating systems, user applications, background services, background applications, etc.

[0130] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0131] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

Claims

1. A method for semantic segmentation of clinical images based on a complex model guided by frequency phase, characterized in that: The steps include: S1: Preprocessing of original clinical images; S2: input the preprocessed image data obtained in step S1 into a Gabor wavelet transform module to convert it into a complex-valued domain image; S3: Input the complex-valued domain image obtained in step S2 into the encoding network, and selectively adjust the perception of the phase component and the preference for different frequency components through the complex-valued mixed attention module to obtain the complex-valued domain feature; S4: Input the complex-valued domain features obtained in step S3 into the constructed frequency adaptive separation module, decompose the complex-valued domain features into high-frequency features and low-frequency features, and combine the low-frequency features with the high-frequency features to filter out important features through the frequency channel attention module; S5: The important features obtained in step S4 are input into the next level encoder, and the high-frequency features are merged with the features of each level of the decoder. The real and imaginary features are integrated at the last layer of the decoder to obtain the segmentation result of the clinical image.

2. According to the method of claim 1, the method is characterized in that: The preprocessing method in step S1 is: first normalize to reduce the range of pixel values, and then use the data enhancement strategy to expand the data set, specifically: A1: Change the image size to the same size through interpolation so that it can be uniformly input into the network training later; A2: Normalize the pixel value range of the image to 0-1 to reduce the interference caused by uneven lighting in clinical images; A3: Use three data augmentation methods, namely flipping, rotating and skewing, to expand the data volume.

3. The method for clinical image semantic segmentation based on a frequency phase guided complex model according to claim 1, characterized in that: The Gabor wavelet transform module in step S2 is composed of a filter group of 4 scales and 5 directions, and a set of Gabor filters is constructed to convert the clinical image into a complex value domain, specifically including: B1: For the filter parameters: scale λ and direction θ, take 4 values ​​and 5 values ​​respectively to construct a set of Gabor filters. The specific formula is as follows: Where x′=xcos(θ)+ysin(θ), y′=-xsin(θ)+ycos(θ), θ represents the tilt angle of the Gabor kernel function; λ and ψ represent the filter wavelength and phase offset, respectively; the standard deviation σ and the spatial aspect ratio δ determine the ellipticity of the Gabor kernel function; B2: Filter the input image through the constructed Gabor filter set to obtain an image in the complex value domain.

4. The method for clinical image semantic segmentation based on a frequency phase guided complex model according to claim 1, characterized in that: In step S3, the encoding network adopts the UNet network, and the UNet network is modified into a complex form, including expressing the convolution kernel, activation function and batch normalization in complex form, as follows: Complex convolution is represented as the convolution operation between the complex feature map h and the convolution filter W, where h = a + ib, W = A + iB; according to the definition of complex operation, the calculation process of complex convolution is expressed as: W*h=(A*aB*b)+i(B*a+A*b); The complex activation ReLU satisfies the Cauchy-Riemann equation as follows: Batch normalization for complex input z is calculated as follows: in, is the normalization of z, V is the covariance matrix and E[·] is the mean of the input z.

5. The method for clinical image semantic segmentation based on a complex model guided by frequency phase according to claim 1, characterized in that: The complex-valued hybrid attention module in step S3 includes a cross-phase fusion module and an amplitude deviation interference module, which first passes through the cross-phase fusion module and then passes through the amplitude deviation interference module to obtain a complex-valued domain feature.

6. The method for clinical image semantic segmentation based on a frequency phase guided complex model according to claim 5, characterized in that: The operation mode of the cross-phase fusion module in step S3 is: C1: Apply 3×3 and 1×1 convolution operations to obtain feature F respectively. 3 and F 1 , and then these features are fused together, a 1×1 convolution and a S-type activation function σ(·) are performed to generate an attention map β representing the importance of local features, as shown below: β=σ(Conv 1×1 (F 3 +F 1 )); C2: 1-β is expressed as β′ at the same time, and these two soft attention maps are used to automatically select more representative features from each receptive field, as shown below: F c =cat(β′×F 3 ,β×F 1 )。 7. The method for clinical image semantic segmentation based on a frequency-phase-guided complex model according to claim 6, characterized in that: The operation mode of the amplitude deviation interference module in step S3 is: D1: The complex-valued features output from step C2 and the input of the cross-phase fusion module are added pixel by pixel and then passed through 3×3 as the input of the amplitude deviation interference module; D2: Apply a 1×1 convolution operation to the output of step D1 to reduce the channel dimension and split it into two independent processing paths; D3: The upper branch uses a deep separable volume to extract global context information; D4: The lower branch features are separated into amplitude and phase components, and then the amplitude information F is processed by applying convolution with controlled noise. s A slight perturbation is added, and the perturbed amplitude information is recombined with the phase information, as shown below: Amplitude=Conv 1×1 ·(Conv 1×1 ·|F s |), F complex =Amplitude×cos(arctan(|F s |))+j·Amplitude×sin(arctan(|F s |)); D5: F complex As a weight multiplied with the upper branch, the interference feature vector is obtained and the original channel dimension is restored through a 3×3 convolution operation.

8. The method for clinical image semantic segmentation based on a frequency phase guided complex model according to claim 1, characterized in that: The operation process of the frequency adaptive separation module in step S4 is: E1: Perform a first-level wavelet decomposition on the feature map along the channel dimension to obtain low-frequency coefficients and high frequency coefficients E2: Replace the low-frequency coefficients with zero matrix and perform inverse wavelet transform together with the high-frequency coefficients to obtain the high-frequency feature F H ∈R C ×H×W ; E3: Use zero values ​​to replace high-frequency coefficients and perform inverse wavelet transform to obtain low-frequency features F L ∈R C×H×W , and the high-frequency feature F H and low frequency features F L Merge into F cat Input frequency channel attention module; E4: F cat Divide the channel into n parts, denoted as Then to F i A DCT operation is performed and the stitching is done together to get the overall compressed features as follows: F dct =cat([DCT(F0),DCT(F1),…,DCT(F n-1 )]); E5: F dct Apply the fully connected operation and activation operation, and then with F cat Multiply to get F hybrid ∈R 2C×H×W .

9. The method for clinical image semantic segmentation based on a frequency phase guided complex model according to claim 8, characterized in that: The step S5 is specifically as follows: F1: Feature F obtained in step E5 hybrid Change the number of channels to C through 1×1 convolution, which is used as the input of the next encoder; F2: The high-frequency feature F obtained in step E2 H Transmitted to the corresponding decoder stage via skip connections to supplement fine-grained information; F3: Integrate the real and imaginary features of the network output to obtain the final segmentation feature map. The specific formula is as follows: where F final , F real and F imag Represent the final segmentation features, real features and imaginary features respectively.

10. The method for clinical image semantic segmentation based on a frequency-phase-guided complex model according to claim 9, characterized in that: In step F3, the segmentation result is obtained by activating the Softmax function, and the calculation formula of Softmax is:

Images