Research On The Lightweight Cell Detection Model And Interpretability For Cervical TCT Pathology Image

Cervical cancer is one of the major tumor diseases that endanger women’s health and cause cancer deaths in women.TCT is a mainstream clinical screening method for cervical cancer with excellent screening results and low cost.However,the image resolution of TCT samples after enlarged scanning and imaging is more than 70,000 pixels,which makes it easy to miss and misdiagnose if pathologists diagnose it directly.In recent years,artificial intelligence(AI)technology has been widely used to assist in cervical cancer screening,however,there are still some limitations of AI applied to TCT pathology image auxiliary diagnostic of cervical cancer.For example,the weak generalization ability of the model due to the difficulty in acquiring training samples,and the large amount of time and computational resources required for the model to perform inference on the whole TCT pathology images.In addition,the deep learning-based TCT auxiliary diagnosis model has poor interpretability and low transparency due to its "black box" characteristics,and the diagnosis results are difficult to gain the trust of clinicians.To address the above problems,this thesis uses a publicly available cervical cancer TCT pathology image dataset with data enhancement by random rotation and random flip to study a lightweight cervical cancer TCT pathology image cell detection model based on deep learning and its interpretability.The main work of this thesis as follows:(1)A causal attention-based lightweight cell detection model Sim YOLOv5-CCAD for TCT pathological images of cervical cancer was proposed.The model is based on YOLOv5 to improve the detection accuracy and inference performance of the model,respectively.In terms of model detection accuracy,the deformable convolution module guided by causal attention mechanism improves the feature extraction ability and generalization ability of the model to accurately extract morphological features of cervical cancer cells in a refractory cell background.In terms of model inference performance,a slim neck network is designed by a lightweight convolutional structure GSConv,which improves the model inference speed and consumes less computational resources while ensuring the model accuracy.Experiments show that the model improves 1.9% and 1.3% in detection accuracy index m AP-0.5 and m AP-0.5:0.95,respectively,and reduces 22.3% in the number of parameters and 19.5% in the amount of computation in the model inference speed index,respectively.Additionally the single inference time is reduces from 5.9ms to 5.2ms.(2)The causal interpretability of cervical cancer cell detection model was studied.This thesis proposes a counterfactual and Grad-CAM based interpretation method and a Grad-CAM based causal visual feature extraction method from both intervention and non-intervention perspectives of model causal interpretability,respectively.And further proposed a joint causal interpretation method based on intervention and non-intervention,i.e.,a counterfactual interpretation method guided by causal visual features.The interpretation method based on counterfactual and Grad-CAM uses the attributional activation map obtained by Grad-CAM to guide the counterfactual region selection process,which improves the problems of the traditional counterfactual interpretation method such as strong randomness of counterfactual region selection and long interpretation time.The GradCAM-based causal visual feature extraction method uses the fusion gradient obtained by background category back-propagation to approximate the background features,and then separates the causal visual features from the fusion features obtained by Grad-CAM,providing a causal visual feature attribution activation map with stronger causality,more accurate localization and clearer hierarchy.The causal visual feature-based guided counterfactual interpretation uses causal visual features to guide the counterfactual region selection process on the basis of the first two,embedding model-intrinsic causality for the counterfactual region selection process,and realizing the full process causal interpretation of counterfactual region selection and causal intervention perturbation.(3)A trusted diagnosis decision system for TCT pathology images of cervical cancer was constructed.This system takes the model and interpretable method proposed in this paper as the core to realize the auxiliary diagnosis of TCT pathological images of cervical cancer,and gives the visualization results and diagnosis reports of model-assisted diagnosis.It also provides three model causal interpretation methods for different application scenarios,providing doctors with a diversified and multi-scale model-assisted diagnosis basis and improving doctors’ trust in the auxiliary diagnosis results.It helps doctors to judge and make decisions on complex conditions,and effectively improves the efficiency of clinical diagnosis.