CN-121999001-A - Medical image segmentation method and system based on Perturb-and-Denoise regularization frame

Abstract

The invention discloses a medical image segmentation method and a medical image segmentation system based on Perturb-and-Denoise regularization frames, and belongs to the field of medical image processing. The method and the device can effectively improve the precision and reliability of medical image segmentation. By performing intelligent disturbance and structure perception denoising on the filter base, the adaptive capacity of the constant-variation convolution network to noise and complex structures of images is enhanced, so that the constant-variation convolution network can obtain clearer and continuous segmentation boundaries on biomedical images with abundant details such as electron microscope images. The method improves the generalization performance of the model under the conditions of cross-sample and cross-data, simultaneously maintains the integrity of edges and fine structures, and finally provides more accurate and robust technical support for neuroscience and quantitative analysis of cell layers.

Inventors

• ZHENG XIN
• BAO RUI
• YIN QIAN
• ZHANG QIULI

Assignees

• 北京师范大学

Dates

Publication Date

20260508

Application Date

20260206

Claims (8)

1. 1. A medical image segmentation method based on Perturb-and-Denoise regularization framework, comprising the steps of: collecting a medical image to be segmented; constructing a medical image segmentation model based on Perturb-and-Denoise regularization framework; and inputting the medical image to be segmented into a medical image segmentation model to complete segmentation of the image to be segmented.
2. 2. The Perturb-and-Denoise regularization framework-based medical image segmentation method of claim 1, wherein the medical image segmentation model employs a constant-change convolution network, and the construction method comprises: obtaining a disturbance base by injecting noise based on an initial base designed for the constant convolution network; Acquiring an optimization base through content self-adaptive denoising processing based on the disturbance base; based on the optimization basis, constructing a constant-variation convolution network model for image segmentation.
3. 3. The method of medical image segmentation based on Perturb-and-Denoise regularization framework of claim 2, wherein the perturbation basis is obtained by step-wise injection of gaussian noise based on an initial basis designed for a constant convolution network.
4. 4. The method for medical image segmentation based on Perturb-and-Denoise regularization framework of claim 2, wherein a final optimization basis is obtained by denoising with a bilateral filtering operator according to the perturbation basis.
5. 5. A medical image segmentation method based on Perturb-and-Denoise regularization framework according to claim 2, characterized in that the optimization basis is obtained by a content-adaptive denoising process, corresponding to solving an optimization problem whose objective function includes data fidelity terms aligned to the perturbation basis, and regularization terms that penalize the lack of smooth solution.
6. 6. The Perturb-and-Denoise regularization framework-based medical image segmentation method of claim 4, wherein the bilateral filtering operator, when denoising, computes a weighted average based on spatial distance and pixel value differences.
7. 7. The method of claim 6, wherein the calculating the weighted average includes calculating the weighted average for the spatial position of the object in the basis vector based on the spatial distance weights and pixel value difference weights of points in the neighborhood window to obtain the denoised basis vector value.
8. 8. A medical image segmentation system based on Perturb-and-Denoise regularization framework, the system configured to implement the method of any of claims 1-7, comprising: The acquisition module is used for acquiring medical images to be segmented; the construction module is used for constructing a medical image segmentation model based on Perturb-and-Denoise regularization frames; The segmentation module is used for inputting the medical image to be segmented into the medical image segmentation model to complete segmentation of the image to be segmented.

Description

Medical image segmentation method and system based on Perturb-and-Denoise regularization frame Technical Field The invention relates to the field of medical image processing, in particular to a medical image segmentation method and a medical image segmentation system based on Perturb-and-Denoise regularization frames. Background In the field of medical image processing, especially in the field of connective tissue research in neuroscience, it is important to perform accurate cell and cell membrane segmentation on electron microscope images. In recent years, deep learning-based methods, such as U-Net and its variants, have made significant progress in this task. These models can automatically extract features and complete pixel-level classification by learning a large amount of annotation data. However, biomedical images are generally characterized by high noise, low contrast, and complex fine structures, so that the segmentation result of the general convolutional neural network is easily broken or blurred at the boundary, and the generalization ability may be reduced in the face of samples outside the training data distribution. In order to improve generalization and structural perceptibility of models, constant convolution networks are receiving more and more attention. Such networks can more efficiently utilize the inherent rules of data by explicitly embedding symmetry priors, such as rotation or reflection, in the architecture, enhancing robustness to geometric transformations. A convolution filter, one of its core components, is typically composed of a well-designed set of mathematically idealized basis functions. While such idealized filter bases have advantages in theoretical analysis, in practical applications, their overly smooth and highly idealized structure tends to be difficult to adequately adapt to complex, irregular, and often defective morphological features in real biological tissue. This lack of adaptation limits the ability of the network to capture subtle structural changes in the real data, possibly resulting in segmentation performance that fails to meet theoretical expectations. Therefore, how to optimize the design of the filter base in the isomorphism convolution network, so that the filter base can keep the theoretical elegance and isomorphism attribute, and can better model and adapt to the complex and noisy structural characteristics in the real-world medical image, and the filter base becomes a technical problem to be solved urgently. Existing methods typically use fixed idealized bases directly or do simple parameter tuning, lacking a systematic framework to robustly and generalize the base functions themselves. Disclosure of Invention In order to solve the technical problems in the background, the invention provides the following scheme: A medical image segmentation method based on Perturb-and-Denoise regularization framework, comprising the steps of: collecting a medical image to be segmented; constructing a medical image segmentation model based on Perturb-and-Denoise regularization framework; and inputting the medical image to be segmented into a medical image segmentation model to complete segmentation of the image to be segmented. Preferably, the medical image segmentation model adopts a constant-variation convolution network, and the construction method comprises the following steps: obtaining a disturbance base by injecting noise based on an initial base designed for the constant convolution network; Acquiring an optimization base through content self-adaptive denoising processing based on the disturbance base; based on the optimization basis, constructing a constant-variation convolution network model for image segmentation. Preferably, the perturbation basis is obtained by stepwise injection of gaussian noise based on an initial basis designed for a constant convolution network. Preferably, denoising is performed through a bilateral filtering operator according to the disturbance base, and a final optimization base is obtained. Preferably, the optimization basis is obtained by a content-adaptive denoising process, corresponding to solving an optimization problem, the objective function of which includes a data fidelity term aligned with the disturbance basis, and a regularization term punishing the lack of smooth solution. Preferably, the bilateral filtering operator calculates the weighted average according to the difference between the spatial distance and the pixel value when denoising is performed. Preferably, the calculation mode of the weighted average comprises the step of calculating the weighted average for the target space position in the base vector according to the space distance weight and pixel value difference weight of each point in the neighborhood window so as to obtain the denoised base vector value. The invention also provides a medical image segmentation system based on Perturb-and-Denoise regularization framework, which is used for the method a