Domain generalization three-dimensional medical image segmentation method based on expansion amplitude transformation

Abstract

The invention discloses a field generalization three-dimensional medical image segmentation method based on expansion amplitude transformation, which mainly comprises four processes of small-batch disturbance, three-dimensional image decomposition, three-dimensional style adjustment and three-dimensional image synthesis, can effectively solve the field difference problem in a clinical medical scene, and can effectively improve the segmentation efficiency especially under the condition that a field label is unknown. And the generalization performance of the model is improved. Firstly, an input three-dimensional medical image is preprocessed through a small-batch disturbance technology so as to improve the randomness of data. Secondly, decomposing the image into content features and style features by using a three-dimensional image decomposition technology; then, three-dimensional style adjustment is adopted, style features of a specific field are disturbed through a cuboid mask, and meanwhile, invariable content features of the field are reserved so as to simulate potential field differences; and finally, recombining the processed features through a three-dimensional image synthesis technology, and generating a medical image suitable for an unknown target field. The method can improve the performance of three-dimensional medical image segmentation, and has good universality.

Description

Technical Field

[0001] The present invention relates to a domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transformation, and belongs to the technical field of three-dimensional medical image segmentation. Background Art

[0002] In the field of medical image segmentation, accurate segmentation plays a crucial role in key processes such as disease diagnosis, surgical planning, and treatment outcome evaluation. Medical image segmentation technology can extract key structures, such as organs, blood vessels, and tumors, from complex image backgrounds, accurately locating lesions or anatomical structures. This technology provides crucial support for disease diagnosis, surgical planning, and treatment outcome evaluation, and provides doctors with intuitive and accurate visual information, enabling more precise treatment plans, improving treatment outcomes, and effectively reducing patient suffering.

[0003] Many current studies are focusing on domain generalization (DG) technology to address the ubiquitous domain differences in medical images. Through innovative methods such as adversarial training, federated learning, and causal enhancement, DG technology is breaking through the bottlenecks of device dependency and data fragmentation in medical image analysis. DG technology attempts to enhance the generalization capabilities of the model during training so that it can maintain good performance in new domains such as different hospitals, devices, and patient populations. Many existing DG methods aim to improve the performance of models in multi-source domains through techniques such as contrastive learning, invariant content collaborative learning, and meta-learning. These methods have, to a certain extent, improved the accuracy and robustness of medical image segmentation.

[0004] However, existing domain generalization technologies still face a significant problem. Most existing domain generalization calculations rely on domain labels to achieve domain generalization. However, in actual clinical environments, due to differences in factors such as hospitals, imaging equipment, and operators, obtaining accurate domain labels becomes extremely difficult. This reliance on domain labels limits the application of existing technologies in complex and changing clinical scenarios, causing domain generalization models to perform poorly when faced with unknown domain data. Therefore, current domain generalization models are difficult to apply to clinical medical scenarios. Therefore, there is an urgent need for a method that can achieve 3D medical image segmentation without the need for domain labels. Summary of the Invention

[0005] The content of this application is used to briefly introduce concepts that will be described in detail in the detailed description section below. The content of this application is not intended to identify key features or essential features of the technical solution for which protection is sought, nor is it intended to limit the scope of the technical solution for which protection is sought.

[0006] In view of the problems and shortcomings in the prior art, the present invention aims to provide a domain-generalized 3D medical image segmentation method based on dilation amplitude transformation, which improves the generalization ability of the model by simulating potential domain differences, and uses a 3D style adjuster to interfere with the style features of the domain in a specific way through the cuboid perturbation technology while keeping the content features unchanged, so that the model

[0007] The model can more effectively learn 3D medical image features, so that it can achieve 3D medical image segmentation without the need for domain labels, thereby solving the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention discloses a domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transformation, comprising the following steps:

[0010] Step 1, in response to obtaining a three-dimensional medical image dataset with unknown domain labels;

[0011] Step 2: dividing the three-dimensional medical image dataset into source domain and target domain image data by leaving one domain out;

[0012] Step 3: inputting the source domain image data into a dilation amplitude transform framework to train a three-dimensional medical segmentation model;

[0013] Step 4, using the source domain image data as a training set and optimizing the parameters of the 3D medical segmentation model in combination with a loss function;

[0014] Step 5: Using the target domain image data as a test set, verify the generalization performance of the 3D medical segmentation model and then output it;

[0015] The dilated amplitude transform framework includes small-batch perturbation, 3D image decomposition, 3D style adjustment and 3D image synthesis parts.

[0016] Preferably, the small batch perturbation part in step 3 is to perform random small batch perturbation on the input source domain image data to obtain small batch perturbation source domain image data, which is used to improve the randomness of the data.

[0017] Preferably, the three-dimensional image decomposition step in step 3 further includes:

[0018] Step 3.1.1, performing a three-dimensional offline Fourier transform on the source domain and the small batch perturbed source domain image data to obtain a frequency domain representation;

[0019] Step 3.1.2, decomposing the Fourier transform based on the frequency domain representation to obtain the real part and the imaginary part;

[0020] Step 3.1.3, calculate the amplitude and phase of the source domain and the source domain image after small batch perturbation according to the real part and the imaginary part according to the complex operation.

[0021] Preferably, the three-dimensional style adjustment step in step 3 further includes:

[0022] Step 3.2.1, set the mask of the cuboid mask perturbation;

[0023] Step 3.2.2, applying a rectangular mask to the source domain image amplitude to obtain the amplitudes of its low-frequency and high-frequency components;

[0024] In step 3.2.3, the amplitude of the source domain image after small batch perturbation is then masked with a cuboid to obtain the amplitude of its low-frequency and high-frequency components;

[0025] In step 3.2.4, the amplitude of the source domain image and the low-frequency component of the amplitude of the source domain image after small-batch perturbation are used to obtain the perturbed amplitude through expansion fusion.

[0026] Preferably, the three-dimensional image synthesis step in step 3 further includes:

[0027] Step 3.3.1, synthesize the amplitude and phase after the disturbance into the corresponding frequency domain representation;

[0028] In step 3.3.2, the frequency domain representation is synthesized by three-dimensional offline inverse Fourier transform to obtain a new perturbed image.

[0029] Preferably, in step 3.2.4, the source domain image amplitude and the low-frequency component of the source domain image amplitude after small-batch perturbation are used to obtain the perturbed amplitude through expansion fusion, which is expressed as:

[0030]

[0031] in, represents the amplitude of the low-frequency component of the source domain image, represents the amplitude of the low-frequency component of the source domain image after small batch perturbation, represents the amplitude of the high-frequency component of the source domain image, M(x, y, z) represents the rectangular perturbation mask, and r represents the Beta distribution with hyperparameters (α, α).

[0032] Preferably, the loss function in step 4 is expressed as:

[0033] Loss=DiceLoss(f θ (X + ),Y)+λ·MSELoss(f θ (X + ),f θ (X));

[0034] Among them, f θ represents the neural network model, Y represents the segmentation mask, λ represents the trade-off parameter for balancing different training losses, DiceLoss represents the objective function of the segmentation task, and MSELoss is used to ensure that the new image X after perturbation + and the output consistency of the source domain image X.

[0035] As a second aspect of the present application, the present invention further discloses a domain-generalized three-dimensional medical image segmentation system based on dilation amplitude transformation, comprising a data acquisition module, a dilation amplitude transformation module, a training module, and a verification module; wherein the dilation amplitude transformation module comprises a small batch perturbation unit, a three-dimensional image decomposition unit, a three-dimensional image adjustment unit, and a three-dimensional image synthesis unit;

[0036] A small batch perturbation unit is used to perform a small batch perturbation on the input source domain image data to perform a random shuffling operation;

[0037] A three-dimensional image decomposition unit is used to calculate the amplitude and phase of the source domain image data after three-dimensional discrete Fourier transform of the source domain image data after small batch perturbation;

[0038] A three-dimensional image adjustment unit is used to decompose the source domain image data and the small batch perturbed source domain image data through a cuboid mask to obtain the high-frequency and low-frequency component amplitudes, and then expand and fuse them to obtain the perturbed amplitude;

[0039] The three-dimensional image synthesis unit is used to synthesize the disturbed amplitude and phase into a new disturbed image through three-dimensional discrete Fourier inverse transform.

[0040] As a third aspect of the present application, the present invention further discloses an electronic device, comprising:

[0041] at least one processor, and a memory communicatively coupled to the at least one processor;

[0042] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the above-mentioned domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transform.

[0043] As the fourth aspect of the present application, the present invention also discloses a computer storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the above-mentioned domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transformation are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention provides a domain-generalized 3D medical image segmentation method based on dilated amplitude transform (DAT). This method can achieve 3D medical image segmentation without the need for domain labels, expanding the model's application potential in unknown domains. A 3D medical image dataset with unknown domain labels is obtained and then separated into source and target domain image data using a leave-one-out method. The source domain image data is then input into a DAT framework to train a 3D medical segmentation model. Finally, the 3D medical segmentation model is optimized and validated before output. The DAT framework includes small-batch perturbation, 3D image decomposition, 3D style adjustment, and 3D image synthesis. Specifically, the input source domain image is first preprocessed using small-batch perturbation to improve the randomness of the data. Next, the source domain image is decomposed into content features and style features. The domain-specific style features are then perturbed using a cuboid mask to simulate potential domain differences. Finally, the processed features are recombined through 3D image synthesis to generate medical images adapted to the unknown target domain. This paper utilizes cuboid perturbation technology to perturb the style characteristics of a specific domain without changing the domain-invariant content features. This effectively improves the model's generalization ability, enabling it to adapt to diverse clinical environments. It is also applicable to various 3D medical image segmentation applications, demonstrating excellent versatility and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings constituting a part of this application are used to provide a further understanding of this application and make other features, purposes and advantages of this application more apparent. The drawings and descriptions of the exemplary embodiments of this application are used to explain this application and do not constitute an improper limitation on this application.

[0047] In the attached figure:

[0048] Figure 1 A connection diagram of the main steps of the domain-generalized three-dimensional medical image segmentation method in an embodiment of the present invention;

[0049] Figure 2 This is a flowchart of the overall steps of the domain-generalized three-dimensional medical image segmentation method in an embodiment of the present invention;

[0050] Figure 3 This is a structural diagram of the expansion amplitude transformation framework in an embodiment of the present invention;

[0051] Figure 4 This is a structural diagram of a three-dimensional style adjustment part in an embodiment of the present invention;

[0052] Figure 5 Schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0054] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other.

[0055] The present invention discloses a domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transformation. The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. The present invention includes an intervention utility identification part, a blocking set update part, and an intervention utility set determination part. Figure 1 and Figure 2 As shown, it mainly includes the following steps:

[0056] Step 1, in response to obtaining a three-dimensional medical image dataset with unknown domain labels;

[0057] Step 2: The 3D medical image dataset is divided into source domain and target domain image data by leaving one domain out;

[0058] Step 3: Input the source domain image data into the dilation amplitude transform framework to train the 3D medical segmentation model;

[0059] Step 4: Using the source domain image data as a training set, the parameters of the 3D medical segmentation model are optimized in combination with the loss function;

[0060] Step 5: Use the target domain image data as a test set to verify the generalization performance of the 3D medical segmentation model and then output it.

[0061] First, we need to obtain a 3D medical image dataset with unknown domain labels. Domain labels are used to identify the specific domain or subset to which the data belongs. Their role is to distinguish the source, generation environment, or application scenario of the data. Generally speaking, in clinical medical scenarios, patients may be divided into many potential domains with unknown labels based on different hospitals, equipment, or operator experience. For example, different medical institutions use different imaging equipment (such as CT, MRI, X-ray, etc.), and their imaging principles, resolutions, and parameter settings vary. In addition, different hospitals have different requirements for the technical level of operators, which leads to inconsistent data collection standards, etc. These differences make domain division complex and difficult to clearly define. At the same time, due to factors such as the multi-source heterogeneity of medical data, privacy protection requirements, and annotation costs, the actual domain labels cannot be obtained.

[0062] We divide the acquired 3D medical image dataset with unknown domain labels into source domain and target domain using the Leave One Domain Out method. Leave One Domain out is achieved by systematically excluding target domain data, that is, the source domain includes multiple independent data domains (such as different hospitals, equipment or patient groups), and the target domain means selecting one domain from the 3D medical image dataset as the test set each time, and merging the remaining domains into the source domain and executing it in a loop. That is, assuming that the 3D medical image dataset contains N independent domains, the i-th domain is taken as the target domain in turn, and the remaining domains are merged into the source domain and represented as N-1. Specifically, the source domain represents the data set used to train the initial segmentation model, and the target domain represents the data set with domain differences from the source domain. They both have no domain label information.

[0063] Then, the 3D medical image dataset is subjected to amplitude dilation perturbation operation through the dilation amplitude transformation framework. Figure 3 As shown, the expansion amplitude transformation framework includes small batch perturbation, 3D image decomposition, 3D style adjustment and 3D image synthesis parts. First, small batch perturbation is performed on the source domain image data. Small batch perturbation refers to improving model performance or algorithm effect by introducing controllable randomness. In the present invention, we perform a random shuffling operation on the source domain image data without domain labels, so as to improve the randomness of the source domain image data. Then the source domain image data is decomposed and the amplitude and phase are obtained by calculation, which specifically includes the following steps:

[0064] Step 3.1.1: Perform a 3D offline Fourier transform on the source domain and the mini-batch perturbed source domain image data to obtain a frequency domain representation.

[0065] Step 3.1.2, decompose the Fourier transform into real and imaginary parts based on the frequency domain representation;

[0066] In step 3.1.3, the amplitude and phase of the source domain image and the source domain image after mini-batch perturbation are calculated according to the real and imaginary parts according to the complex operation.

[0067] Specifically, the source domain image data undergoes a 3D offline Fourier transform. 3D offline Fourier transform is a frequency domain analysis method for static 3D data, converting source domain image data (such as volume data, 3D images) into a frequency domain representation. The conversion from the spatial domain to the frequency domain reveals its spatial frequency components, which can be expressed as:

[0068]

[0069] Among them, X(d,h,w) represents the dimensions of the source domain image data, namely w (width), d (depth) and h (height).

[0070] Indicates the summation of the three dimensional variables of the source domain image data. Represents a complex exponential function that includes periodicity and frequency information. The frequency domain representation after Fourier transform is then decomposed into real and imaginary parts. The real part represents the amplitude information of the even-symmetric component in the signal, corresponding to the cosine wave component, and the imaginary part represents the amplitude information of the odd-symmetric component in the signal, corresponding to the sine wave component. It can be expressed as:

[0071] F(x,y,z)=F Re (x,y,z)+iF Im (x,y,z);

[0072] Among them, F Re Expressed as the real part of the three-dimensional offline Fourier transform, F Im It is expressed as the imaginary part after the three-dimensional offline Fourier transform, where i is a complex unit. Finally, the amplitude A and phase P are calculated according to the complex operation, which are expressed as:

[0073]

[0074] The amplitude A represents the energy intensity of the signal at the corresponding frequency, and the phase P reflects the time offset characteristics of each frequency component in the signal. The amplitude of the source domain image and the amplitude of the source domain image after small batch perturbation are calculated as A and A respectively. T .

[0075] In 3D style adjustment, such as Figure 4 As shown in Figure 1, a cuboid mask is used to perturb the style features of a specific domain while retaining the domain-invariant content features to simulate potential domain differences. The following steps are also included:

[0076] Step 3.2.1, set the mask of the cuboid mask perturbation;

[0077] Step 3.2.2: The amplitude of the source domain image is masked by a rectangular block to obtain the amplitudes of its low-frequency and high-frequency components;

[0078] In step 3.2.3, the amplitude of the source domain image after small batch perturbation is then masked with a cuboid to obtain the amplitude of its low-frequency and high-frequency components;

[0079] In step 3.2.4, the amplitude of the source domain image and the low-frequency component of the amplitude of the source domain image after small-batch perturbation are used to obtain the perturbed amplitude through expansion fusion.

[0080] Specifically, the cuboid mask is used in 3D object detection or point cloud segmentation tasks to define a cuboid region in 3D space using bounding box parameters (center point, length, width, height, and rotation angle). In the present invention, the cuboid mask M is represented as:

[0081]

[0082] L is a hyperparameter that determines the magnitude of the low-frequency component. (x, y, z) represent the horizontal, vertical, and depth directions of the cuboid, respectively. L·D≤x≤(1-L)·D indicates that the range of x is constrained by the depth D and the hyperparameter L. Similarly, the range of y is constrained by the height H and the hyperparameter L, and the range of z is constrained by the width W and the hyperparameter L.

[0083] Then, we first pass the source domain image amplitude through the cuboid mask to obtain the low-frequency component amplitude A l (x,y,z) and high-frequency component amplitude A h (x,y,z), expressed as:

[0084]

[0085] Where ° represents element-wise multiplication. Low-frequency components correspond to areas in the image where grayscale or color changes slowly, while high-frequency components correspond to areas in the image where grayscale or color changes rapidly. Represents the element-wise multiplication of the source domain image amplitude and the cuboid mask. This means that the source domain image amplitude is multiplied element-by-element by the inverted rectangular mask to achieve the rectangular mask operation. After the source domain image amplitude is passed through the above formula, the low-frequency component amplitude of the source domain image can be obtained. and high-frequency component amplitude Next, the source domain image after small batch perturbation is also subjected to three-dimensional offline Fourier transform in sequence, and the amplitude A of the source domain image after small batch perturbation is calculated. T , and then calculate the amplitude of the low-frequency component of the source domain image after small batch perturbation according to the above formula and high-frequency component amplitude Finally, according to the amplitude of the low-frequency component of the source domain image and the amplitude of the low-frequency component of the source domain image after small batch perturbation Calculate the amplitude A of the rectangular block after disturbance * The specific formula is:

[0086]

[0087] Among them, r represents the Beta distribution with hyperparameters (α, α), and the rectangular mask M determines the weight. and the amplitude of the low-frequency component of the source domain image after small batch perturbation First perform weighted summation to obtain the total low-frequency component amplitude A l* , and then use the total low-frequency component amplitude A l* and the high-frequency component amplitude of the source domain image The weighted summation is used to obtain the amplitude A after disturbance* 。 Through the above formula, and By performing random fusion expansion, the total low-frequency component amplitude A can be increased l* diversity in order to better simulate domain differences.

[0088] Finally, in the three-dimensional image synthesis, the amplitude A after the disturbance of the cuboid is obtained * The image corresponding to the phase P is synthesized, which is represented as the inverse operation of the three-dimensional image grading. Specifically, the following steps are included:

[0089] Step 3.3.1, synthesize the corresponding frequency domain representation of the perturbed amplitude and phase;

[0090] In step 3.3.2, the frequency domain representation is synthesized by three-dimensional offline inverse Fourier transform to obtain the new perturbed image.

[0091] Specifically, the amplitude and phase of the perturbed cuboid are first synthesized into the corresponding frequency domain representation, which is expressed as:

[0092] F=A * cos(P)+iA * sin(P);

[0093] Among them, A * cos(P) and A * sin(P) represents the cosine component and the sine component, that is, the real part and the imaginary part of the frequency domain representation. The synthesized frequency domain representation is then transformed into the corresponding image X using the three-dimensional discrete Fourier inverse transform. + , expressed as:

[0094]

[0095] The perturbed image X + By adjusting the frequency domain characteristics of Fourier transform to simulate the domain differences of different imaging devices or acquisition centers, the generalization ability of the model in unknown target domains is enhanced.

[0096] Finally, we use the source and target domains divided by the leave-one-out method described above to train and validate the trained 3D medical segmentation model. Specifically, in each experiment, a certain target domain image data is used as an independent test set, and a mixture of multiple source domain image data is used as the training set. The model parameters are optimized in the source domain using a loss function. The loss function is expressed as:

[0097] Loss=DiceLoss(f θ (X + ),Y)+λ·MSELoss(f θ (X +),f θ (X));

[0098] Among them, f θ represents the neural network model, Y represents the segmentation mask, λ represents the trade-off parameter for balancing different training losses, DiceLoss represents the objective function of the segmentation task, and MSELoss represents the function used to ensure that the new image X after perturbation + And the output consistency of the source domain image X. After optimization, the generalization performance of the model is verified on the unseen target domain test set.

[0099] Application of the embodiment

[0100] In this example, we prepared a public domain label-agnostic 3D prostate cancer image dataset consisting of six central datasets and containing only T2WI images. As part of the preprocessing phase, the 3D images were normalized to a resolution of 16×128×128. Table 1 below shows the details of the prostate generalized segmentation dataset.

[0101]

[0102] Table 1

[0103] We used the U-Net architecture as the backbone of the 3D medical segmentation model. All results were obtained by averaging the results of 5 independent runs to ensure the robustness and reliability of the results. During training, we used a batch size of 4 and trained for 100 epochs on all datasets. We used the SGD optimizer with an initial learning rate of 0.28 and a weight decay of 2×10 -5 , to optimize the performance of the model. In addition, to ensure training stability, we adopted a polynomial learning rate decay scheme, that is, the learning rate will decay over time.

[0104] The model validation indicators selected are Dice similarity coefficient (Dice) and average symmetric surface distance (ASSD). Among them, Dice similarity coefficient is a commonly used metric used to quantify the similarity between the true value and the predicted segmentation mask. The larger its value, the better the segmentation performance. ASSD is a metric used to evaluate the average surface distance between the ground truth value and the predicted segmentation. The smaller its ASSD value, the better the segmentation performance. DeepAll is the most basic performance reference method for DG tasks. Its core idea is to not adopt any special methods for domain differences, but directly train a unified model on the source domain (usually using UNet as the backbone network), and then test it on the target domain.

[0105] The experimental results are shown in Tables 2 and 3 below.

[0106]

[0107] Table 2

[0108]

[0109] Table 3

[0110] The experimental results above show that compared to the baseline method DeepAll, the proposed method achieves absolute improvements of 6.14% and 5.63% in the Dice similarity coefficient of M-avg. and m-avg., respectively, and relative improvements of 9.73% and 8.73%. In terms of the ASSD average surface distance metric, the proposed method achieves absolute improvements of 0.83mm and 0.78mm in M-avg. and m-avg., respectively, and relative improvements of 25.78% and 27.37%.

[0111] The present application also discloses a domain-generalized three-dimensional medical image segmentation system based on dilation amplitude transformation, which includes a data acquisition module, a dilation amplitude transformation module, a training module and a verification module. Among them, the dilation amplitude transformation module includes a small-batch perturbation unit, a three-dimensional image decomposition unit, a three-dimensional image adjustment unit and a three-dimensional image synthesis unit. The small-batch perturbation unit is used to perform a small-batch perturbation on the input source domain image data to perform a random shuffling operation. The three-dimensional image decomposition unit is used to calculate the amplitude and phase of the source domain and the source domain image data after small-batch perturbation after three-dimensional discrete Fourier transform. The three-dimensional image adjustment unit is used to decompose the source domain and the source domain image data after small-batch perturbation through a rectangular mask to obtain the high-frequency and low-frequency component amplitudes, and then expand and fuse them to obtain the perturbed amplitude. The three-dimensional image synthesis unit is used to synthesize the perturbed amplitude and phase into a new perturbed image through a three-dimensional discrete Fourier inverse transform.

[0112] In order to implement the above embodiment, the present application also discloses an electronic device. Figure 5 As shown, the electronic device 500 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 508 into a random access memory (RAM) 503. Various programs and data required for the operation of the electronic device 500 are also stored in the RAM 503. The processing device 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0113] Typically, the following devices may be connected to the I / O interface 505: an input device 506 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 507 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 508 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 509. The communication device 509 may allow the electronic device 500 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 5 The electronic device 500 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead. Figure 5 Each block shown in the figure may represent one device, or may represent multiple devices as needed.

[0114] In particular, according to some embodiments of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, some embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer storage medium, and the computer program includes program code for executing the method shown in the flowchart. In some such embodiments, the computer program can be downloaded and installed from the network via the communication device 509, or installed from the storage device 508, or installed from the ROM 502. When the computer program is executed by the processing device 501, the above-mentioned functions defined in the method of some embodiments of the present disclosure are performed.

[0115] It should be noted that the computer storage medium described above in some embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0116] In some embodiments of the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of the present disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0117] In some embodiments, the client and server can communicate using any currently known or future developed network protocol, such as HTTP (Hypertext Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0118] The computer storage medium may be included in the electronic device or may exist independently and not incorporated into the electronic device. The computer storage medium carries one or more programs that, when executed by the electronic device, enable the electronic device to implement a domain-generalized three-dimensional medical image segmentation method based on dilation amplitude transform.

[0119] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0120] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the part of the module, program segment or code includes one or more executable instructions for realizing the logical function of the specification. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings.

[0121] For example, two boxes shown in succession may actually be executed substantially in parallel, or they may sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of boxes in the block diagram and / or flow chart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or may be implemented using a combination of dedicated hardware and computer instructions. The units described in some embodiments of the present disclosure may be implemented in software or in hardware. The units described may also be provided in a processor, and the names of these units do not, in some cases, constitute limitations on the units themselves.

[0122] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0123] The above description is only an illustration of some preferred embodiments of the present disclosure and the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A domain-generalized 3D medical image segmentation method based on dilation amplitude transform, characterized by: The steps include: Step 1, in response to obtaining a three-dimensional medical image dataset with unknown domain labels; Step 2: dividing the three-dimensional medical image dataset into source domain and target domain image data by leaving one domain out; Step 3: inputting the source domain image data into a dilation amplitude transform framework to train a three-dimensional medical segmentation model; Step 4, using the source domain image data as a training set and optimizing the parameters of the 3D medical segmentation model in combination with a loss function; Step 5: Using the target domain image data as a test set, verify the generalization performance of the 3D medical segmentation model and then output it; The dilated amplitude transform framework includes small-batch perturbation, 3D image decomposition, 3D style adjustment and 3D image synthesis parts.

2. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 1, characterized in that: The small batch perturbation part in step 3 is to perform random small batch perturbation on the input source domain image data to obtain small batch perturbation source domain image data, which is used to improve the randomness of the data.

3. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 2, characterized in that: The three-dimensional image decomposition part in step 3 further includes the following steps: Step 3.1.1, performing a three-dimensional offline Fourier transform on the source domain and the small batch perturbed source domain image data to obtain a frequency domain representation; Step 3.1.2, decomposing the Fourier transform based on the frequency domain representation to obtain the real part and the imaginary part; Step 3.1.3, calculate the amplitude and phase of the source domain and the source domain image after small batch perturbation according to the real part and the imaginary part according to the complex operation.

4. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 3, characterized in that: The three-dimensional style adjustment part in step 3 further includes the following steps: Step 3.2.1, set the mask of the cuboid mask perturbation; Step 3.2.2, applying a rectangular mask to the source domain image amplitude to obtain the amplitudes of its low-frequency and high-frequency components; In step 3.2.3, the amplitude of the source domain image after small-batch perturbation is subjected to a rectangular mask to obtain the amplitudes of its low-frequency and high-frequency components. In step 3.2.4, the amplitude of the source domain image and the low-frequency component of the amplitude of the source domain image after small-batch perturbation are used to obtain the perturbed amplitude through expansion fusion.

5. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 4, characterized in that: The three-dimensional image synthesis part in step 3 further includes the following steps: Step 3.3.1, synthesize the amplitude and phase after the disturbance into the corresponding frequency domain representation; In step 3.3.2, the frequency domain representation is synthesized by three-dimensional offline inverse Fourier transform to obtain a new perturbed image.

6. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 4, characterized in that: In step 3.2.4, the source domain image amplitude and the low-frequency component of the source domain image amplitude after small-batch perturbation are used to obtain the perturbed amplitude through expansion fusion, which is expressed as: in, represents the amplitude of the low-frequency component of the source domain image, represents the amplitude of the low-frequency component of the source domain image after small batch perturbation, represents the amplitude of the high-frequency component of the source domain image, M(x, y, z) represents the rectangular perturbation mask, and r represents the Beta distribution with hyperparameters (α, α).

7. The domain-generalized 3D medical image segmentation method based on dilation amplitude transform according to claim 5, characterized in that: The loss function in step 4 is expressed as: Loss=DiceLoss(f θ (X + ),Y)+λ·MSELoss(f θ (X + ),f θ (X)); Among them, f θ represents the neural network model, Y represents the segmentation mask, λ represents the trade-off parameter for balancing different training losses, DiceLoss represents the objective function of the segmentation task, and MSELoss is used to ensure that the new image X after perturbation + and the output consistency of the source domain image X.

8. A domain-generalized 3D medical image segmentation system based on dilation amplitude transform, characterized by: It includes a data acquisition module, an expansion amplitude transformation module, a training module and a verification module; wherein the expansion amplitude transformation module includes a small batch perturbation unit, a 3D image decomposition unit, a 3D image adjustment unit and a 3D image synthesis unit; A small batch perturbation unit is used to perform a small batch perturbation on the input source domain image data to perform a random shuffling operation; A three-dimensional image decomposition unit is used to calculate the amplitude and phase of the source domain image data after three-dimensional discrete Fourier transform of the source domain image data after small batch perturbation; A three-dimensional image adjustment unit is used to decompose the source domain image data and the small batch perturbed source domain image data through a cuboid mask to obtain the high-frequency and low-frequency component amplitudes, and then expand and fuse them to obtain the perturbed amplitude; The three-dimensional image synthesis unit is used to synthesize the disturbed amplitude and phase into a new disturbed image through three-dimensional discrete Fourier inverse transform.

9. An electronic device, characterized in that: include: at least one processor, and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the steps of the method according to any one of claims 1 to 7.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps according to any one of claims 1 to 7 are implemented.

Images