Zero-shot MRI brain tumor image generation method

Abstract

The present invention discloses a zero-shot MRI brain tumor image generation method, belonging to the field of image generation. The method comprises the following steps: S1: acquiring and preprocessing data without a brain tumor sample; S2: constructing a brain tumor shape generation method; S3: constructing a brain tumor texture generation method; S4: constructing a method for simulating tumor mass effect and capsule effect; and S5: generating zero-shot MRI brain tumor images based on the aforementioned brain tumor shape generation method, texture generation method, and mass effect and capsule effect simulation methods. First, a lesion-free brain tumor dataset undergoes brain extraction and white matter segmentation. In step S2, a mathematical model is established to simulate tumor shape, generating core tumor regions, edema regions, and complete tumor regions. In step S3, texture simulation is performed using Gaussian noise and cubic spline interpolation, and Gaussian filtering is used to avoid oversharpening of the generated texture. Finally, in step S4, local scaling and warping are used to simulate tumor mass effect and capsule effect, respectively. This method generates zero-shot MRI brain tumor images using traditional image processing algorithms rather than deep neural networks, eliminating the need for tumor samples. The method increases the controllability of image generation, allowing precise adjustment of specific features in the image, such as tumor size, shape, and location.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and designs a zero-sample MRI brain tumor image generation method, belonging to the field of image generation. Background Art

[0002] In the field of medical imaging, especially MRI brain tumor imaging, obtaining high-quality data is challenging, primarily because this process typically requires expensive equipment and specialized operations, limiting the quantity and diversity of the data. Furthermore, the precise annotation of medical images is time-consuming and costly, requiring experienced physicians or radiologists to identify pathological features within the images. In particular, for rare or difficult-to-diagnose cases, these images can be extremely scarce in databases, posing significant challenges to image-based research and analysis. These factors collectively contribute to the scarcity of medical image data, limiting the potential for in-depth research and development using this data.

[0003] Deep learning technology has demonstrated significant potential in the field of medical image segmentation, particularly brain tumor segmentation. Advanced machine learning techniques such as convolutional neural networks (CNNs) can accurately identify and segment tumor regions, providing support for diagnosis, treatment planning, and pathological research. Soheila Saeedi, Sorayya Rezayi, and others used convolutional neural networks (CNNs) to classify gliomas, meningiomas, and pituitary tumors. However, the difficulty of acquiring medical image data, the high cost of annotation, and the scarcity of annotated data have all limited the development of deep learning technology in this field, particularly its application and widespread adoption in clinical practice.

[0004] Given these limitations, the importance of zero-shot MRI brain tumor image generation has become increasingly prominent. This technology can generate images of specific brain tumor types without the need for a sample. It not only reduces the need and cost of data annotation but also accelerates medical research, providing more diverse data for deeper understanding of disease characteristics. In clinical applications, these generated images can aid diagnosis and treatment planning, particularly in treating uncommon brain tumor cases.

[0005] Some progress has been made in using deep learning methods, particularly those based on generative adversarial networks, to generate brain tumor images. For example, Debadyuti Mukherkjee et al. proposed a novel aggregation method called AGGrGAN, which combines synthetic images from different GAN models. The study also employed style transfer techniques to capture localized information from the source images. Experiments have shown that this method achieves promising results on brain tumor datasets.

[0006] Compared to traditional methods that rely on deep learning, the unique brain tumor image generation method adopted by the present invention does not rely on deep neural networks, but uses traditional image processing algorithms and does not require any brain tumor samples. This method provides higher controllability when generating brain tumor images, allowing precise adjustment of specific features in the image, such as tumor size, shape, and location. In addition, compared with methods based on generative adversarial networks (GANs), this method has obvious advantages in the diversity of generated images and can cover a wider range of tumor types and characteristics. In summary, this traditional image processing technology not only provides a high degree of controllability, but also broadens the range of diversity of image generation, making it superior to deep learning-based methods in specific application scenarios. Summary of the Invention

[0007] In order to address the shortcomings of the prior art, the present invention provides a zero-sample MRI brain tumor image generation method, which can complete the generation of MRI brain tumor images without using a dataset containing brain tumor lesions. The specific solution of the present invention is as follows:

[0008] The zero-sample MRI brain tumor image generation method includes the following steps:

[0009] S1: Data acquisition and preprocessing of brain tumor-free samples;

[0010] S2: Construct a brain tumor shape generation method;

[0011] S3: Construct a method for generating brain tumor textures;

[0012] S4: Constructing methods to simulate tumor mass effect and capsule effect;

[0013] S5: Based on the above-mentioned brain tumor shape generation method, texture generation method and method of simulating mass effect and capsule effect, zero-sample MRI brain tumor image generation is performed.

[0014] Furthermore, in the experiment of step S1, the IXI dataset was used, which was jointly created by medical imaging research centers such as Hammersmith Hospital, Guy's Hospital and Institute of Psychiatry in the UK. This dataset mainly contains brain magnetic resonance imaging (MRI) images of normal volunteers with a wide range of age and gender coverage, and does not contain brain tumor lesion samples. The preprocessing of this dataset includes extracting brain tissue from the data. Considering the difference in the growth rate of tumors in gray matter and white matter, it is necessary to segment the white matter and gray matter. The purpose of this processing method is to complete the generation of zero-sample MRI brain tumor images in the absence of brain tumor samples.

[0015] Furthermore, step S2 constructs a method for generating a brain tumor shape, specifically:

[0016] S21. Brain tumors spread in gray and white matter. First, we need to randomly sample the initial point of the tumor. We randomly select an initial point in the gray and white matter of the brain to obtain the initial coordinates.

[0017] S22. Construct a mathematical formula to simulate the spread of brain tumors as follows:

[0018]

[0019] In the mathematical model for simulating brain tumor spread, D(x,y,z) represents the probability of a pixel at coordinate (x,y,z) becoming a tumor cell. w and p g Represents the percentage of white matter and gray matter at the coordinate point, respectively. This ratio is assigned according to the white matter and gray matter probability map segmented in step S1. w (x,y,z) and D g (x, y, z) represents the diffusion rate of tumor cells in white matter and gray matter, respectively, and Ω represents the domain of white matter and gray matter in the dataset.

[0020] S23, starting diffusion simulation based on the initial coordinates obtained in S21, respectively, the left and right pixel points of the coordinates on the x, y, and z axes are as follows:

[0021] (x-1,y,z),(x+1,y,z),(x,y-1,z),(x,y+1,z),(x,y,z-1),(x,y,z+1)

[0022] In the S22 stage of the brain tumor simulation, initial coordinates are input into the mathematical model to calculate the probability of a tumor cell appearing at that location. Based on this probability, a roulette wheel strategy is used to determine whether a pixel will mutate into a tumor cell. After each round of tumor diffusion, the newly added tumor cells are considered the new tumor boundary. In subsequent iterations, these tumor boundary pixels serve as the starting coordinates for diffusion. To smooth the tumor edge and prevent it from being overly sharp, a Gaussian filter is applied to the tumor boundary in the final stage to achieve a more natural visual effect.

[0023] Furthermore, in step S3, a method for generating a brain tumor texture is constructed. In step S2, the pixels contained in the first generated tumor shape are filled with the average pixel value. Subsequently, in order to simulate the texture of the tumor, 3D Gaussian noise is generated. In order to enhance the realism of the texture, the cubic spline interpolation method is used to carefully interpolate the Gaussian noise along the x, y, and z axes. This interpolation not only enhances the continuity and naturalness of the texture, but also improves the fidelity of the simulated texture. Finally, in order to avoid excessive sharpening of the generated tumor texture, a Gaussian filter is applied to the tumor area. This Gaussian filter is intended to smooth the tumor texture and reduce the unnatural sharpening effect, thereby making the simulation of the tumor more realistic and natural.

[0024] Furthermore, the method for constructing the simulation of tumor mass effect and capsule effect in step S4 is as follows:

[0025] S41. Simulate tumor mass effect. Tumor mass effect refers to the compression and displacement of surrounding brain tissue caused by tumor growth. Use the local scaling warp method to achieve tumor mass effect, as shown below:

[0026]

[0027] In the above formula, γ is the distance between the original tumor pixel and the tumor center, γ′ is the distance between the tumor pixel and the tumor center after scaling, and γ max is the farthest distance from the tumor pixel to the tumor center. I∈[0,100] is used to control the expansion intensity. The larger I is, the stronger the distortion is.

[0028] S42: Simulate the tumor capsule effect. This refers to the formation of a relatively dense layer of tissue around the tumor, which makes the boundary between the tumor and surrounding normal tissue more distinct. Highlight the tumor edge to simulate the capsule effect. Increase the HU intensity of the tumor edge pixels, then apply a Gaussian filter to the tumor edge. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Algorithm flow chart

[0030] Figure 2 Schematic diagram of brain tumor formation DETAILED DESCRIPTION

[0031] S1: Data acquisition and preprocessing of brain tumor-free samples;

[0032] S11, data acquisition stage:

[0033] In this paper's experiments, the IXI dataset was used as the basis for the experimental dataset. This dataset contains imaging data from multiple modalities, including T1 and T2 MRI sequences. This study primarily focused on T1 and T2 image data, with each image having a resolution of 256 × 256 × 150 pixels. These images are normal brain images without lesions, representing healthy brain images without tumor lesions.

[0034] S12, data preprocessing stage:

[0035] During data preprocessing of the IXI dataset, since the dataset contains full head scans, including non-brain tissue, and this experiment focuses on applying algorithms to the brain, we first used the Brain Extraction Tool (BET) in the FSL (FMRIB Software Library) tool to extract the brain from the dataset. BET is a morphological approach that utilizes edge detection techniques and shape models to identify and segment brain tissue, effectively separating it from the full head scan. Given the differential growth rates of brain tumors in gray and white matter, accurate segmentation of these two types of brain tissue is crucial. In this experiment, FMRIB's Automated Segmentation Tool (FAST) in the FSL tool was used to segment white and gray matter. FAST is based on a Bayesian probabilistic framework that combines morphological and statistical methods. By modeling the signal intensity of different tissue types in MRI images, it can automatically identify and segment different tissue types, such as white matter, gray matter, and cerebrospinal fluid. This method ultimately generates tissue probability maps and segmentation maps, providing the necessary foundational data for tumor growth simulation.

[0036] S2. Constructing a brain tumor shape generation method

[0037] S21. When simulating the spread of brain tumors, the selection of the tumor starting point is crucial. This step involves using a random sampling method to determine the tumor's starting point in the gray and white matter regions of the brain. In this way, a point can be randomly selected from the gray and white matter and its initial coordinates can be obtained;

[0038] S22. Construct a mathematical formula to simulate the spread of brain tumors as follows:

[0039]

[0040] When constructing a mathematical model for simulating brain tumor spread, we define the function D(x,y,z) to represent the probability of a tumor cell forming at the pixel coordinate (x,y,z). This probability is based on the proportion p of the location of the point in the white matter and gray matter. w and p g , and the corresponding tumor cells in the white matter D w (x,y,z) and gray matter D g The diffusion rate in (x, y, z). The values of these ratios and diffusion rates are determined based on the probability map obtained by segmentation of the MRI image. Ω represents the white matter and gray matter domains defined in the dataset. In this model, if the coordinate point is within Ω, that is, within the white matter or gray matter region, the probability of being a tumor cell is calculated based on the ratio of white matter and gray matter and their respective diffusion rates. If the coordinate point is not within Ω, the probability of the location being a tumor cell is zero;

[0041] S23, starting diffusion simulation based on the initial coordinates obtained in S21, respectively, the left and right pixel points of the coordinates on the x, y, and z axes are as follows:

[0042] (x-1,y,z),(x+1,y,z),(x,y-1,z),(x,y+1,z),(x,y,z-1),(x,y,z+1)

[0043] The initial coordinates obtained in stage S21 are input into the mathematical model established in stage S22 to calculate the probability that a location will develop into a tumor cell. Based on this probability, a roulette wheel strategy is used to determine whether a pixel at that location will develop into a tumor cell. After each round of tumor diffusion, the newly added tumor cells are considered the new tumor boundary. In subsequent iterations, the pixels at these tumor boundaries serve as the starting coordinates for diffusion. Diffusion occurs in three stages: first, a necrotic tumor core is formed, then an enhancing tumor region is formed, and finally, an edema region is formed. At the end of each stage, the pixels are recorded to generate a corresponding regional mask, which serves as the starting point for the next stage of diffusion. To avoid edge sharpening, a Gaussian filter is finally applied to the tumor boundary to smooth the edges and enhance visual naturalness. This model takes into account the dynamic nature of tumor growth and the manifestation of different biological characteristics in the image, and can achieve a refined simulation of tumor morphology through iteration.

[0044] S3. Construct a method for generating brain tumor texture. In step S2, the pixels in the generated tumor shape are filled to match the average pixel value, thereby simulating the basic structure of the tumor. Subsequently, in order to simulate the texture of the tumor, 3D Gaussian noise is generated. In order to enhance the realism of the texture, the Gaussian noise is carefully interpolated along the x, y, and z axes using cubic spline interpolation. This interpolation not only enhances the continuity and naturalness of the texture, but also improves the level of detail of the simulated texture. Finally, in order to avoid over-sharpening of the generated tumor texture, a Gaussian filter is applied to the tumor area. This Gaussian filter is designed to smooth the tumor texture and reduce the unnatural sharpening effect, thereby making the simulation of the tumor more realistic and natural.

[0045] S4. Construct a method to simulate tumor mass effect and capsule effect.

[0046] S41. Simulate tumor mass effect. Tumor mass effect refers to the compression and displacement of surrounding brain tissue caused by tumor growth. Use the local scaling warp method to achieve tumor mass effect, as shown below:

[0047]

[0048] In the above formula, γ is the distance between the original tumor pixel and the tumor center, γ′ is the distance between the tumor pixel and the tumor center after scaling, and γ max is the farthest distance from the tumor pixel to the tumor center. I∈[0,100] is used to control the expansion intensity. The larger I is, the stronger the distortion is.

[0049] S42: Simulate the tumor capsule effect. This refers to the formation of a relatively dense layer of tissue around the tumor, which makes the boundary between the tumor and surrounding normal tissue more distinct. Highlight the tumor edge to simulate the capsule effect. Increase the HU intensity of the tumor edge pixels, then apply a Gaussian filter to the tumor edge.

Claims

1. A zero-sample MRI brain tumor image generation method, characterized in that: The method comprises: S1: Data acquisition and preprocessing of samples without brain tumors. The experimental data set uses the IXI data set. There are no brain tumor lesion samples in the data set. The data set is preprocessed. First, brain tissue is extracted and white matter and gray matter are segmented. When preprocessing the IXI data set, the BET function in the FSL tool is first used to extract the brain in the data set. Based on the morphological method, edge detection technology and shape model are used to identify and segment brain tissue, thereby effectively separating brain tissue from the complete head scan image. The FAST function in the FSL tool is used to segment white matter and gray matter. Based on the Bayesian probability framework, combined with morphological and statistical methods, the signal intensity of different tissue types in the MRI image is modeled to automatically identify and segment white matter, gray matter and cerebrospinal fluid tissue, and finally generate a probability map and segmentation map of the tissue; S2: Construct a brain tumor shape generation method; S3: Construct a method for generating brain tumor textures; S4: Constructing methods to simulate tumor mass effect and capsule effect; S5: Generate zero-sample MRI brain tumor images based on the above-mentioned brain tumor shape generation method, texture generation method, and method for simulating mass effect and envelope effect; The step S2 constructs a method for generating a brain tumor shape, specifically: S21. Brain tumors spread in gray and white matter. First, we need to randomly sample the initial point of the tumor. We randomly select an initial point in the gray and white matter of the brain to obtain the initial coordinates. S22. Construct a mathematical formula to simulate the spread of brain tumors as follows: ， in: The coordinates are The probability that a pixel will become a tumor cell, and Respectively represent the coordinates The percentage of white matter and gray matter at the pixel point is assigned using the white matter to gray matter probability obtained by step S1. and represent the diffusion rates of tumor cells in white matter and gray matter, respectively. represents the domains that are white matter and gray matter in the dataset; S23, starting diffusion simulation based on the initial coordinates obtained in S21, respectively, calculating the coordinates of the left and right pixel points of the coordinates on the x, y, and z axes according to the following calculation method: ， These coordinates are substituted into the mathematical model in S22 to calculate the probability of becoming a tumor cell. Then, a roulette wheel method is set according to the obtained probability to determine whether the pixel point develops into a tumor cell. After each round of development, the newly added tumor cells in that round are used as the tumor boundary. The next round of diffusion starts with the pixel points at the tumor boundary as the initial coordinates. Finally, in order to avoid the sharpness of the tumor boundary, a Gaussian filter is used to perform a filtering operation on the tumor edge.

2. The zero-sample MRI brain tumor image generation method according to claim 1, characterized in that: In step S3, a method for generating brain tumor texture is constructed. The pixels contained in the tumor shape generated in step S2 are filled with the average pixel value, and 3D Gaussian noise is generated to simulate the texture of the tumor. To make the texture more substantial, the Gaussian noise is interpolated in the x, y, and z directions using cubic spline interpolation. To avoid the sharpness of the generated tumor texture, a Gaussian filter is used to perform a Gaussian filter on the tumor area.

3. The zero-sample MRI brain tumor image generation method according to claim 1, characterized in that: The step S4 is to construct a method for simulating tumor mass effect and capsule effect, specifically as follows: S41. Simulate tumor mass effect. Tumor mass effect refers to the compression and displacement of surrounding brain tissue caused by tumor growth. Use the local scaling distortion method to achieve tumor mass effect, as shown below: ， In the above formula is the distance between the original tumor pixel and the tumor center, is the distance between the tumor pixel and the tumor center after scaling, is the farthest distance from the tumor pixel to the tumor center, Used to control the expansion strength, The larger it is, the stronger the distortion; S42. Simulate the tumor capsule effect. The tumor capsule effect refers to a relatively dense tissue layer formed around the tumor. This tissue layer makes the boundary between the tumor and the surrounding normal tissue clearer. The edge of the tumor is highlighted to simulate the tumor capsule effect. The HU intensity of the pixels at the edge of the tumor is increased, and then a Gaussian filter is used to perform Gaussian filtering on the edge of the tumor.

4. The zero-sample MRI brain tumor image generation method according to claim 1, characterized in that: The brain data without brain tumor lesions in step S1 is extracted, and then the shape of the tumor and the coordinates of the edge are obtained using step S2. Then, the texture of the tumor is generated according to step S3. Finally, the mass effect and capsule effect of the tumor are simulated according to step S5, thereby completing the generation of a high-fidelity MRI brain tumor image.

Images