Search

CN-121982313-A - Weak supervision brain tumor segmentation method

CN121982313ACN 121982313 ACN121982313 ACN 121982313ACN-121982313-A

Abstract

The invention belongs to the technical field of medical image processing and artificial intelligence, and particularly relates to a weak supervision brain tumor segmentation method which comprises the following steps of obtaining a low-level glioma segmentation dataset of a cancer genome map, randomly dividing the dataset into a training set and a testing set according to patient division, carrying out migration learning by utilizing pre-trained ResNet as a backbone network based on image data in the dataset, carrying out self-adaptive optimization and intelligent screening on characteristics extracted based on a traditional class activation mapping technology by utilizing a global optimization mechanism of SZIO algorithm, realizing fine fusion and semantic enhancement of multi-level characteristics, and carrying out fine adjustment on the ResNet backbone network with a preset initial learning rate. The invention can realize the performance breakthrough of weak supervision brain tumor segmentation, has stronger practicality, robustness and clinical transformation potential, and provides a high-efficiency and reliable intelligent segmentation solution for the field of medical image analysis.

Inventors

  • CHEN ZHIQING
  • ZHANG YUEJIN
  • SHEN ZHENG

Assignees

  • 华东交通大学

Dates

Publication Date
20260505
Application Date
20260123

Claims (10)

  1. 1. The weak supervision brain tumor segmentation method is characterized by comprising the following steps of: obtaining a low-level glioma segmentation dataset of a cancer genome map, and randomly dividing the dataset into a training set and a test set according to patient division; based on the image data in the dataset, performing migration learning by using the pre-trained ResNet as a backbone network; utilizing a global optimization mechanism of SZIO algorithm to perform self-adaptive optimization and intelligent screening on the characteristics extracted based on the traditional class activation mapping technology, and realizing fine fusion and semantic enhancement of multi-level characteristics; performing fine tuning on the ResNet backbone network at a preset initial learning rate; deep mining complementary feature information contained in each residual convolution block in four stages in the ResNet network through a hierarchical feature search strategy of the SZIO algorithm, and synchronously optimizing a foreground threshold value, a background threshold value and iteration number parameters of the GrabCut algorithm to capture tumor boundaries; And generating a brain tumor segmentation result based on the optimized characteristics and parameters, and verifying the performance of the model on the test set.
  2. 2. The method of claim 1, wherein the ResNet backbone network uses ImageNet pre-training weights to fine tune 25 epochs at an initial learning rate of 0.0001.
  3. 3. The method of claim 1, wherein the SZIO algorithm comprises four coupling stages of a solicitation matrix, a tactical model, a strategic decision and an individual elimination mechanism; Tactical modes include a soldier mode and a wonder mode, by tactical success rate When it is determined that the local search is in a dead state Judging as the state of convergence of premature ripening, switching to the soldier mode when the state of convergence of premature ripening is the state of convergence ≤ ≤ Maintaining the soldier mode when in use, wherein 。
  4. 4. The method of claim 1, wherein the strategic decision stage calculates a window alertness factor when the cumulative number of attempts reaches Determined evaluation window Triggered at time, calculating tactical success rate based on search data within the current evaluation window Tactical improvement rate And deducing a dynamic decision threshold according to the tactical success rate SR and the tactical improvement rate delta F, and judging a tactical mode of the next stage.
  5. 5. The method of claim 1, wherein the individual elimination mechanism is periodically operated by For interval execution, the current population is sorted according to the fitness, nonlinear elimination probability is calculated, and the population scale is dynamically reduced.
  6. 6. The method according to claim 1, wherein the parameter optimization range of the GrabCyt algorithm is an integer of a foreground threshold value E [0.5,0.9], a background threshold value E [0.1,0.5], and a iteration number E [3,10 ].
  7. 7. The method of claim 1, wherein the optimizing process of SZIO algorithm uses the intersection ratio between the segmentation result and the real label as the fitness function to guide the searching direction of the feature subset and the parameter.
  8. 8. The method of claim 1, wherein the random sampling verification scheme is to randomly extract 100 MRI slices from the training set as a search set for optimizing feature subset-parameter combinations, randomly extract 100 independent slices from the test set as a verification set, and migrate and apply an optimal strategy for performance verification.
  9. 9. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any one of claims 1 to 8.
  10. 10. A computer device comprising a memory, a processor and a computer program stored on the memory, wherein the processor implements the method of any one of claims 1 to 8 when executing the program.

Description

Weak supervision brain tumor segmentation method Technical Field The invention belongs to the technical field of medical image processing and artificial intelligence, and particularly relates to a weak supervision brain tumor segmentation method. Background Most of the optimization problems in the real world are complex, diverse, discrete or continuous, constrained or unconstrained, static or dynamic, single-mode or multi-mode, single-target or multi-target. Mathematical-based optimization methods, such as integer programming, linear programming, convex optimization, and the like, have not been adequate to solve the complex and diverse practical problems. More and more people use a group intelligent optimization algorithm to solve the problem of reality optimization, and compared with a mathematical method, the group intelligent optimization algorithm is a random optimization method based on group synergism and adopting various heuristic rules to iterate continuously, and the method can obtain an acceptable solution in a shorter time although the method does not necessarily converge to a globally optimal solution every time. Problems of the prior art: Although the swarm intelligent optimization algorithm has been widely used because of its excellent performance, it has drawbacks. When the search step size and the search speed are not matched, local optimum or premature results. In order to obtain a faster searching speed, a larger searching step is needed, global optimum is easily missed when the step is too large, and a better solution cannot be obtained in a plurality of iterations when the step is too small, so that the searching speed is influenced. Disclosure of Invention The invention aims to provide a weak supervision brain tumor segmentation method, which can realize the performance breakthrough of weak supervision brain tumor segmentation, has stronger practicality, robustness and clinical transformation potential, and provides a high-efficiency and reliable intelligent segmentation solution for the field of medical image analysis. The technical scheme adopted by the invention is as follows: A weakly supervised brain tumor segmentation method comprising the steps of: obtaining a low-level glioma segmentation dataset of a cancer genome map, and randomly dividing the dataset into a training set and a test set according to patient division; Based on the image data in the data set, utilizing pre-trained ResNet to serve as a backbone network for transfer learning, wherein the ResNet backbone network uses an ImageNet pre-training weight to finely adjust 25 epochs at an initial learning rate of 0.0001; the method comprises the steps of performing self-adaptive optimization and intelligent screening on the characteristics extracted based on the traditional class activation mapping technology by utilizing a global optimization mechanism of SZIO algorithm to realize fine fusion and semantic enhancement of multi-level characteristics, wherein the SZIO algorithm comprises four coupling stages of a solicitation and soldier arrangement, a tactical mode, a strategic decision and an individual elimination mechanism, the tactical mode comprises a solicitation mode and a soldier mode, and the success rate of tactical is the same as that of tactical When it is determined that the local search is in a dead stateJudging as the state of convergence of premature ripening, switching to the soldier mode when the state of convergence of premature ripening is the state of convergence≤≤Maintaining the soldier mode when in use, wherein; The strategic decision stage calculates the window alertness factor when the accumulated number of attempts reachesDetermined evaluation windowTriggered at time, calculating tactical success rate based on search data within the current evaluation windowTactical improvement rateDeriving a dynamic decision threshold according to the tactical success rate SR and the tactical improvement rate DeltaF, and judging a tactical mode of the next stage; The individual elimination mechanism is regular Is performed for intervals. Sorting the current population according to the fitness, calculating nonlinear elimination probability, and dynamically reducing the population scale; performing fine tuning on the ResNet backbone network at a preset initial learning rate; Deep mining complementary characteristic information contained in each residual convolution block in four stages in a ResNet network through a hierarchical characteristic search strategy of the SZIO algorithm, and synchronously optimizing a foreground threshold value, a background threshold value and iteration frequency parameters of a GrabCut algorithm to capture tumor boundaries, wherein the parameter optimization range of the GrabCut algorithm is an integer of the foreground threshold value epsilon [0.5,0.9], the background threshold value epsilon [0.1,0.5] and the iteration frequency epsilon [3,10 ]; In the optimization process of the SZIO al