Research On The Lightweight Classification Model And Interpretability Of TCT Images Of Cervical Cancer

Cervical cancer is one of the common malignant tumors in women,ranking second among female tumors.Clinical follow-up observations have shown that cervical cancer can be prevented and cured through early screening and treatment.There are many ways to screen for early cervical cancer,among which TCT image detection is currently the most advanced and widely used cytological examination method because of its high accuracy,low cost,and wide application in early screening of cervical cancer.However,with the popularization of the two-cancer policy,the demand for TCT image screening for cervical cancer is increasing day by day,and the traditional manual screening method can no longer meet the current screening needs.Currently,the application of artificial intelligence technology to actual screening tasks still faces many challenges.Cervical cancer images have high resolution and a large number of slices.Existing complex classification models have a large number of parameters and calculations,which need to be deployed on highperformance operation devices,which is not conducive to actual clinical screening.In addition,the classification model based on deep learning is a black box model,which cannot give the decisionmaking process of classification and is difficult to gain the trust of doctors.Therefore,researching and designing lightweight and interpretable classification models for cervical cancer TCT images has vital significance for the practical application of cervical cancer TCT image screening.This paper proposes for the first time a lightweight cervical cancer TCT image classification model based on Ghost Module and ECA attention mechanism to address the characteristics of high resolution and a large number of slices.The proposed model significantly reduces the number of parameters and calculations while maintaining the predictive accuracy.Furthermore,to address the unreliable issue brought by the classification model,a Noise-based Grad-Cam interpretable method is proposed to output the interpretability of the model’s classification results,providing reliable evidence for pathologists’ review and diagnosis.The main work focuses on the following aspects:(1)A lightweight cervical cancer TCT image classification model based on Ghost Module and ECA efficient channel attention mechanism is proposed in this paper.According to the characteristics of cervical cancer TCT images,we adopted a lightweight model architecture design that uses depthwise separable convolution instead of traditional convolution and introduces the inverted residual linear bottleneck structure,significantly reducing the number of parameters and calculations of the model.In order to optimize redundant convolution features that produce redundant feature maps during feature extraction of sample images,Ghost Module is used in the model.In addition,in order to solve problems such as accuracy reduction caused by using lightweight architecture and interference caused by sample image background,we added an efficient channel attention mechanism based on ECA to the model,which significantly improves model classification accuracy with fewer parameters and calculations.Finally,comparative experiments and ablation experiments have demonstrated the effectiveness of the proposed lightweight classification model for cervical cancer TCT image classification tasks.(2)A Noise-based Grad-Cam interpretable method for cervical cancer TCT image classification models is proposed in this paper.The cervical cancer TCT image classification model is a black box model that cannot provide the basis for the model’s judgment while outputting the model’s predicted results.This paper uses a Noise-based Grad-Cam interpretable method to output the interpretability in the form of heat maps of the predicted results of the model.This method first adopts a gradient strategy based on Smooth Grad to optimize the redundant noise points in the input sample images.Then,a gradient strategy based on Noise Grad is introduced to optimize the weights of the classification model,in order to enhance the model’s interpretability.Finally,comparative experiments and ablation experiments have demonstrated the effectiveness of the proposed interpretable method for cervical cancer TCT image classification models.(3)A cervical cancer TCT image classification system is designed and implemented in this paper.Firstly,the demand analysis for the cervical cancer TCT image screening process is carried out,and then an image classification system is designed and implemented to assist doctors in sample diagnosis.The system applies the proposed lightweight cervical cancer TCT image classification model and interpretable method to the sample image classification prediction module.After sampling the images,the system uses image processing techniques to quickly slice the samples,and then uses the lightweight classification model to predict the classification of each slice sample.Finally,the proposed interpretable method is used to output the interpretability in the form of heat maps of the classification results.Pathologists can judge and review the diagnostic results through the output results,which improves the efficiency of cervical cancer TCT image screening in clinical applications.