Search

CN-122024216-A - Hysteromyoma segmentation and prognosis analysis method based on diffusion model

CN122024216ACN 122024216 ACN122024216 ACN 122024216ACN-122024216-A

Abstract

The invention provides a method for dividing and prognostic analysis of hysteromyoma based on a diffusion model. The method comprises the steps of collecting and preprocessing 3D T2 MRI image data of a patient suffering from hysteromyoma, converting original data into machine-available data, designing a dividing model of the hysteromyoma in a 3D MRI image based on a diffusion model, optimizing model parameters in a denoising process by designing a function combining a Dice loss, a binary cross entropy loss and a position loss, extracting key features of a focus area, providing redundant features, and further utilizing a 3D-ResNet network to provide the key features of the divided image for prognostic analysis to accurately predict the curative effect of HIFU.

Inventors

  • FAN BINGJUAN
  • WU JUN
  • WANG JIA
  • HE YUANYUAN
  • ZHENG YANWEN

Assignees

  • 杭州市萧山区第一人民医院
  • 浙江大学滨江研究院

Dates

Publication Date
20260512
Application Date
20251226

Claims (9)

  1. 1. A method for dividing and prognostic analysis of hysteromyoma based on diffusion model is characterized by comprising the following steps: step one, collecting 3D T2 MRI image data of a patient suffering from hysteromyoma and preprocessing the data to solve the problems of inconsistent spatial direction, more noise and different resolution in original data; Step two, realizing accurate segmentation of hysteromyoma in the 3D MRI image based on the diffusion model, and identifying and removing background information and non-focus tissues which are irrelevant to HIFU reaction in the image, so that the model is more focused on a distinguishing area in the characteristic learning process; Step three, designing a multi-loss joint optimization model parameter for accurate segmentation of hysteromyoma, and performing joint optimization on the model parameter by adopting Dice loss, binary cross entropy loss and position loss to directly optimize a final segmentation result; step four, extracting key features of the segmented image based on the 3D-ResNet network, and outputting a segmentation mask of the diffusion model As a region of interest (ROI), a myoma region is cropped from the original MRI image X For prognostic analysis.
  2. 2. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 1, wherein the specific way of preprocessing the image data in the first step is as follows: The image correction and the space alignment are carried out, namely an N4 offset field correction algorithm is adopted, a non-parameter non-uniform intensity normalization-based iterative optimization method is adopted, a low-frequency intensity non-uniform field in the image is estimated and compensated, then all the images are unified to a standard coordinate system, and a space reference is established for batch processing; noise reduction, namely adopting Gaussian filtering to carry out smooth noise reduction; image interpolation and resampling, namely resampling all data to the same isotropic voxel size through a tri-linear interpolation algorithm to calculate the intensity value of a new voxel point by linear interpolation in X, Y, Z directions of a three-dimensional space, and resampling all images to a preset uniform isotropic voxel size.
  3. 3. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 1 or 2, wherein the specific way of achieving accurate segmentation of hysteromyoma in 3D MRI image based on diffusion model in the second step comprises forward denoising and backward denoising, and mask is segmented to reality during forward denoising Gradually adding Gaussian noise until completely losing the structural information of the original mask, only retaining noise characteristics, and removing noise from the mask in the reverse denoising process Starting, step-by-step denoising recovery 。
  4. 4. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 3, wherein the specific way of forward noise addition in the second step is as follows: For segmentation mask In other words, at each step t, data From previous step data Generating: Wherein, the Is the total number of steps of forward diffusion, Is a predefined variance schedule that increases as t increases, then by re-parameterization techniques, from Directly obtain : Wherein, the , , When (when) When it is sufficiently large that the number of the cells is small, Approximately obeys a standard gaussian distribution.
  5. 5. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 3, wherein the inverse denoising in the second step is performed by inputting the original 3D MRI image X as a condition, splicing with a noise mask, and inputting the spliced image into the model: Wherein the denoising process of each step obeys the distribution: Wherein, the In order to adjust the variance of the signal, For the mean of model predictions, it computes a conditional-dependent prediction network Here, training of the model is achieved by directly predicting the true mask: Wherein, the Is a learnable noise reduction network.
  6. 6. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 1 or 2, wherein the method for optimizing model parameters by using the Dice loss in the third step is as follows The overlap degree between the predicted hysteromyoma region and the real label is calculated, the method is suitable for solving the problem of unbalanced classification between myoma and background tissues, and can effectively promote the model to generate a segmented region with high spatial consistency, and the specific calculation mode is as follows: Wherein, the And Are respectively pixels True and predicted values of (a).
  7. 7. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 6, wherein the third step uses binary cross entropy loss The specific mode of optimizing the model parameters is as follows, two independent classifications are carried out on each pixel in an image, the classification accuracy of a segmentation result at the pixel level is ensured by comparing the prediction probability with a real label pixel by pixel, the accuracy of the segmentation boundary details of hysteromyoma is ensured, and the specific calculation mode is as follows: 。
  8. 8. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 7, wherein the joint optimization of model parameters by position loss in the third step is performed by the following specific method The accuracy of the segmentation result on the anatomical position is ensured by minimizing the Euclidean distance between the predicted myoma centroid and the real centroid, and the calculation mode is as follows: Wherein, the And Representing the coordinates of the predicted myoma centroid and the real centroid respectively; Finally, the total loss of the model is the sum of three losses, namely, the Dice loss, the binary cross entropy loss and the position loss, namely: 。
  9. 9. The method for segmentation and prognosis analysis of hysteromyoma based on diffusion model according to claim 1 or 2, wherein the method for extracting key features of segmented image based on 3D-ResNet network in the fourth step comprises the following steps of using 3D ResNet as an infrastructure, and extracting the key features from the segmented image based on 3D-ResNet network Deep image histology features including texture, shape, and intensity distribution, etc.). The advantage of the 3D ResNet network is that the three-dimensional medical image data can be effectively processed; In obtaining features of regions of interest Then, the nonlinear correlation between the image characteristics and the disease curative effect is further extracted by using a deep neural network to realize the prediction of the HIFU curative effect, and then the value is converted into the probability prob of poor prognosis by using the Sigmoid function of the output layer: 。

Description

Hysteromyoma segmentation and prognosis analysis method based on diffusion model Technical Field The invention relates to the technical field of digital medical treatment, in particular to a method for segmenting and prognostic analysis of hysteromyoma based on a diffusion model. Background Uterine fibroids are the most common benign pelvic neoplasms of women of childbearing age, whose occurrence is closely related to estrogen levels, with high clinical morbidity, and major symptoms including menorrhagia and menostaxis, which may lead to complications such as anemia. In addition to menstrual changes, uterine fibroids can also cause various degrees of pelvic compression symptoms due to volume increase or special location, causing frequent urination, urgency and even incontinence. In addition, myomas, if located submucosally, can interfere with fertilized egg implantation or block the fallopian tubes, thereby affecting fertility in females. Current treatment protocols are primarily surgical, including laparoscopic or hysteroscopic hysteromyectomy and hysterectomy. With the advancement of medical ideas and the improvement of life quality requirements of patients, minimally invasive and noninvasive treatment modes which keep uterus and pay attention to organ functions are increasingly paid, and in this context, high-intensity focused ultrasound (HIFU) technology under ultrasound or magnetic resonance guidance is developed. HIFU, a non-invasive treatment means, has the advantages of small trauma, rapid recovery, few complications and the like, and is now one of the safe and effective treatment options for symptomatic uterine fibroids. However, not all uterine fibroids are suitable for HIFU ablation, and their efficacy is closely related to factors such as location, size, blood supply, and acoustic environment of the fibroid, and accurate preoperative evaluation is critical to the efficacy of HIFU treatment. The preoperative risk assessment helps to screen appropriate cases, formulate personalized treatment regimens, and thereby maximize therapeutic efficacy and reduce risk of recurrence. However, most of the current related researches rely on manually screened features for analysis, the feature screening process is complicated, the subjectivity is high, deep image histology features which are highly related to prognosis are easily omitted, and the accuracy of prognosis is limited. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide a diffusion model-based uterine fibroid segmentation and prognosis analysis method, which is characterized in that 3DT2 weighted MRI image data of a patient is collected, a diffusion model-based uterine fibroid segmentation model is designed, irrelevant information in MRI is removed, a key focus area is reserved for a prognosis analysis model, and features are automatically extracted from accurately segmented focuses by utilizing 3D-CNN, so that more objective and more accurate prediction of HIFU curative effect is realized. In order to achieve the aim, the invention provides the following technical scheme that the method for segmenting and prognostic analysis of hysteromyoma based on a diffusion model is characterized by comprising the following steps: Step one, collecting 3DT2MRI image data of a patient suffering from hysteromyoma and preprocessing the data to solve the problems of inconsistent spatial direction, more noise and different resolution in the original data; step two, realizing accurate segmentation of hysteromyoma in the 3DMRI image based on the diffusion model, and identifying and removing background information and non-focus tissues which are irrelevant to HIFU reaction in the image, so that the model is focused on a distinguishing area in the characteristic learning process; Step three, designing a multi-loss joint optimization model parameter for accurate segmentation of hysteromyoma, and performing joint optimization on the model parameter by adopting Dice loss, binary cross entropy loss and position loss to directly optimize a final segmentation result; step four, extracting key features of the segmented image based on the 3D-ResNet network, and outputting a segmentation mask of the diffusion model As a region of interest (ROI), a myoma region is cropped from the original MRI image XFor prognostic analysis. As a further improvement of the present invention, the specific way of preprocessing the image data in the first step is as follows: The image correction and the space alignment are carried out, namely an N4 offset field correction algorithm is adopted, a non-parameter non-uniform intensity normalization-based iterative optimization method is adopted, a low-frequency intensity non-uniform field in the image is estimated and compensated, then all the images are unified to a standard coordinate system, and a space reference is established for batch processing; noise reduction, namely adopting Gaussian filtering t