| Cervical cancer is one of the most common malignant tumors in women,and it is also one of the few malignant tumors with a clear etiology.TCT is a mainstream cervical cancer screening method.Although the screening rate is relatively high and the cost is relatively low,it involves multiple manual procedures and requires professional pathologists to operate.In recent years,deep learning methods have been widely used in cervical cancer auxiliary screening.However,there are still some limitations in the development of current medical AI technology,such as over-reliance on training samples.The characteristics of less annotation and unbalanced categories of cervical cancer pathological images have brought new challenges.Semi-supervised object detection can effectively improve the performance of the model by using unlabeled data.However,there are still problems in the application of existing semi-supervised object detection methods on unbalanced cervical cancer pathological image datasets.In addition,since the medical field is a highly security-sensitive field,the interpretability of the model is also crucial.The diagnostic model based on the "black box" parameter training mode is opaque and incomprehensible,so it is difficult to gain the trust of doctors and patients.In this paper,a semi-supervised object detection model for cervical cancer pathological images based on RetinaNet is proposed to solve the problems of few annotations,unbalanced categories,and opaque and untrustworthy black-box models of cervical cancer pathological image data.The automatic detection and model interpretation of cancer pathological images can effectively improve the efficiency of clinical diagnosis.The main work includes the following aspects:1)Semi-supervised object detection model for pathological images of cervical cancer based on RetinaNet.The model is an end-to-end semi-supervised object detection framework based on Teacher-Student,which can be trained with a small amount of annotation.In terms of feature extraction,the prediction of the model is guided by deformable convolution and attention mechanism.In order to address the impact of category imbalance on the quality of pseudo-labels,this paper uses Mean-threshold to measure the difference between categories and reweights the contribution of each category.This paper conducts experiments on cervical cancer pathological image datasets with imbalance ratios of 120 and 200 and annotation ratios of 10% and 5%.The experimental results show that,compared with the existing single-stage semi-supervised object detection models CSD(300/512)and T-SET,the proposed method has obvious advantages in the rare category of cervical cancer cells.Higher m AP is obtained on the unbalanced cervical cancer pathology image dataset.2)Cervical cancer detection model interpretable method.Among them,the LIME method generates new sample data based on perturbation to fit a simple model to approximate the prediction of the complex model in the test set,and gives the interpretation result of the actual prediction by analyzing the weight of local features in the simple model.On this basis,the SHAP method further introduces the SHAPLY value to distribute the feature contribution fairly.These two methods belong to the second level of the causal ladder.In order to obtain interpretation results that are more in line with human thinking,this paper further explains the model prediction results from the third level of the causal ladder,that is,counterfactuals.Using SLIC superpixel segmentation to divide the pixels in the image,and then continuously perturb to find the segment that has an important impact on the model prediction,and finally attribute the result to the segment with the smallest granularity.The prediction results of the model are explained through a variety of interpretable methods,which enhances the credibility of the model and greatly reduces the potential risk of the model in clinical diagnosis.3)Cervical cancer pathological image automatic detection system.According to the needs of pathological image detection of cervical cancer,a user management module and a patient information management module are designed and implemented for information management of doctors and patients respectively.At the same time,the trained detection model is deployed to the system through flask to help clinicians make pathological diagnoses and generate pathological reports.In the system,for the pathological images of cervical cancer with doubtful model detection,the interpretable method is applied,and a reasonable model explanation is given while improving the diagnostic efficiency. |
PDF Full Text Request