Search

CN-122023936-A - Cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning

CN122023936ACN 122023936 ACN122023936 ACN 122023936ACN-122023936-A

Abstract

The invention discloses a cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning, which comprises the following steps of constructing and pre-training a basic feature extraction model, training a dynamic topological network, self-adapting Bayesian optimization, uncertainty quantization and reasoning, realizing image uploading and model selection through a graphical interface, displaying confidence distribution and feature heat map of classification results and assisting clinical diagnosis.

Inventors

  • ZHAO LILI
  • LI SHUYUE
  • LUO YIFENG
  • HANG YUEQIN

Assignees

  • 南通大学

Dates

Publication Date
20260512
Application Date
20260227

Claims (10)

  1. 1. A cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning is characterized by comprising the following steps: (1) Constructing and pre-training a basic feature extraction model, namely performing global feature extraction on cervical cell images by using a global feature extraction network, converting the extracted global features into a two-dimensional feature map suitable for a local feature extraction network through a feature adaptation layer, inputting an improved local feature extraction network for pre-training to obtain a baseline model; (2) Based on the obtained baseline model, counting the activation intensity of the neurons of the full-connection layer through verification set data, dynamically generating an input mask and an output mask according to the activation intensity, constructing a forward propagation path of the self-adaptive topological network by using the masks, and applying noise disturbance to the gradient of the key neurons in the training process; (3) The self-adaptive Bayesian optimization comprises the steps of replacing a standard normalization layer in a self-adaptive topological network with a Bayesian normalization layer, wherein the Bayesian normalization layer introduces a learnable noise intensity parameter in a forward propagation process, and fine-tuning the network to obtain a probabilistic neural network model; (4) Uncertainty quantification and reasoning, namely, for cervical cell images to be classified, carrying out forward propagation sampling for a plurality of times by using the obtained probabilistic neural network model, calculating the mean value of a plurality of prediction results as a classification result, and calculating variance as an uncertainty measure.
  2. 2. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference according to claim 1, wherein in the step (1), a global feature extraction network is ViT model, a local feature extraction network is an improved LeNet-DNN network, and a feature adaptation layer is a full connection layer for linearly transforming and remodelling a global feature vector output by the ViT model into a two-dimensional feature map adapted to the input of the LeNet-DNN network.
  3. 3. The cervical cell image classification method based on dynamic topology optimization and Bayesian inference according to claim 1, wherein in the step (2), the specific way of dynamically generating the input mask and the output mask is to screen out the high-activation neuron indexes respectively according to the average activation intensity of the full-connection layer neurons in the forward propagation process on the verification set, and generate the corresponding binary mask, wherein the mask is used for indicating the path selection of feature fusion in the adaptive topology network.
  4. 4. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference according to claim 1, wherein in the step (2), the self-adaptive topological network comprises an original input path, an intermediate transformation path and an output fusion path, the directional fusion of multipath features is realized through mask indexing, and the expression capability of the fused features is enhanced through a nonlinear activation function.
  5. 5. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference according to claim 1, wherein in the step (2), gradient noise disturbance only acts on the weight gradient corresponding to the key neurons screened by the activation intensity, wherein the noise is random noise.
  6. 6. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference as claimed in claim 1, wherein in the step (3), a learnable noise scaling factor is introduced into forward propagation by a Bayesian normalization layer, noise injection and gradient propagation are realized through a re-parameterization skill, and the fine tuning process adopts a progressive strategy to freeze or defrost network parameters in stages so as to gradually adapt to model changes caused by noise injection.
  7. 7. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference according to claim 1, wherein in the step (4), forward propagation samples are Monte Carlo samples for a plurality of times, probability distribution of a predicted result is obtained through independent forward propagation for a plurality of times, a mean value is used as a final classification result, a variance is used as an uncertainty index, and the method is used for identifying high uncertainty samples and prompting manual review.
  8. 8. The cervical cell image classification method based on dynamic topological optimization and Bayesian inference as set forth in claim 1, wherein the method further comprises the steps of realizing image uploading and model selection through a graphical interface, displaying confidence distribution and characteristic heat map of classification results, and assisting clinical diagnosis.
  9. 9. A computer readable storage medium, characterized in that a computer program is stored thereon, which, when being executed by a processor, implements the method according to any of claims 1-8.
  10. 10. An electronic device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the method of any of claims 1-8.

Description

Cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning Technical Field The invention relates to the technical field of medical image analysis and application deep learning intersection, in particular to a cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning. Background Cervical cancer is a global high-grade malignancy for women, early screening is an important means of detection, and cervical cell screening is a key technology for early diagnosis. Traditional screening relies on pap smear and mainly relies on manual analysis, which is time-consuming, limited by doctor experience and professional acquisition technology, and the easy stacking of cells is unfavorable for observation, resulting in low diagnosis rate. Therefore, how to balance screening efficiency and diagnostic accuracy with technical innovation becomes an important point of research. Based on this, the application of computer vision and deep learning techniques in cell image classification is becoming a research hotspot. With the development of deep learning techniques, leNet and ViT are increasingly being applied to cervical cell classification. However, the prior art has three defects that the traditional LeNet model is simple, but has small capacity, weak feature extraction capability, poor effect on high-precision tasks, easy overfitting when processing pictures with complex lines and complex backgrounds, viT model has strong global capability but high calculation complexity, insufficient capturing capability on image local feature details and is easy to be interfered by data noise, and cervical cell dataset has the problems of unbalanced category and small data volume, so that the model is overfitted, high confidence error judgment is easy to be generated, and the robustness requirement required by medicine cannot be met. In addition, although the existing Bayesian neural network can process uncertainty information, the network structure learning process is complex, the calculation cost is high, the problem that the dimension disaster is faced, and the combination of the dynamic feature selection mechanism is difficult is solved. Disclosure of Invention The invention aims to provide a cervical cell image classification method based on dynamic topology optimization and Bayesian reasoning, which improves the generalization capability of a model to a cell image by adopting a fusion strategy of a dynamic topology network and normalization layer optimization and solves the problems that the existing Bayesian neural network can process uncertainty information, but the network structure learning process is complex, the calculation cost is high, the dimension disaster is faced, and the combination with a dynamic feature selection mechanism is difficult. The cervical cell image classification method based on dynamic topological optimization and Bayesian reasoning comprises the following steps: (1) Constructing and pre-training a basic feature extraction model, namely performing global feature extraction on cervical cell images by using a global feature extraction network, converting the extracted global features into a two-dimensional feature map suitable for a local feature extraction network through a feature adaptation layer, inputting an improved local feature extraction network for pre-training to obtain a baseline model; (2) Based on the obtained baseline model, counting the activation intensity of the neurons of the full-connection layer through verification set data, dynamically generating an input mask and an output mask according to the activation intensity, constructing a forward propagation path of the self-adaptive topological network by using the masks, and applying noise disturbance to the gradient of the key neurons in the training process; (3) The self-adaptive Bayesian optimization comprises the steps of replacing a standard normalization layer in a self-adaptive topological network with a Bayesian normalization layer, wherein the Bayesian normalization layer introduces a learnable noise intensity parameter in a forward propagation process, and fine-tuning the network to obtain a probabilistic neural network model; (4) Uncertainty quantification and reasoning, namely, for cervical cell images to be classified, carrying out forward propagation sampling for a plurality of times by using the obtained probabilistic neural network model, calculating the mean value of a plurality of prediction results as a classification result, and calculating variance as an uncertainty measure. Further, in the step (1), the global feature extraction network is ViT model, the local feature extraction network is improved LeNet-DNN network, and the feature adaptation layer is full connection layer, which is used for linearly transforming and remodelling the global feature vector output by the ViT model into a two-dimensional feature map adapted to t