Research On Breast-ultrasound Image Classification Based On Semi-Supervised Deep Learning

Breast cancer is the most common and deadly malignancy in women worldwide.Prompt and accurate diagnosis of benign and malignant breast tumors can reduce the risk of patients’ lives.Breast-ultrasound imaging is an indispensable tool in breast cancer diagnosis,but interpreting ultrasound images and making diagnostic conclusions is laborintensive.Recent advances in deep learning have made it possible to assist in the diagnosis of medical images.However,the training of deep models relies on a large number of high-quality labeled samples,and collecting accurate labeled breast-ultrasound images requires patients to undergo surgical biopsy,which is costly,while obtaining unlabeled data only requires scanning and imaging.It is relatively easy.In order to reduce the need for labeled images and save medical costs,the thesis uses both labeled and unlabeled data to automatically classify breast-ultrasound images using a semi-supervised deep learning method.The thesis proposes a benign and malignant breast-ultrasound image classification algorithm named MeanMatch based on semi-supervised deep learning image methods,which integrates the key elements of a variety of semi-supervised classification algorithms,and takes into account the entropy minimization,consistency regularization and generalization regularization.They are three important regularization principles of semi-supervised deep learning.MeanMatch utilizes feature distribution information in unlabeled data,reducing the need for labeled images,and outperforming supervised and other semi-supervised methods on the test set.In addition,in view of the problem that the classification accuracy is not high due to the lack of labeled samples,this paper uses the idea of ensemble learning to divide the labeled data of breast-ultrasound images into multiple labeled training sets and their corresponding validation sets.Single classifiers fused by an ensemble strategy using relative majority voting,get performance gains for both supervised and semi-supervised methods.Based on the medical pathological knowledge of breast-ultrasound images,the thesis combines the BI-RADS features of breast-ultrasound images with benign and malignant features,and proposes a semi-supervised breast ultrasound image classification method named Deep SSL-BI.The semi-supervised method pre-trains the BI-RADS level classifier of breast-ultrasound images using MeanMatch method,then uses the BI-RADS feature mapping matrix to map the BI-RADS levels to benign and malignant probabilities,and combines image-based methods for feature fusion of breast-ultrasound images,and Adding the consistency constraint of BI-RADS features to unlabeled data is more in line with the medical characteristics of breast-ultrasound images,thereby improving the classification performance.The experimental results on the test set verify the effectiveness of the method.In addition,the thesis adds the consistency constraints of BIRADS features to other semi-supervised algorithms,and the experimental results verify its effectiveness.Finally,the thesis verifies through experiments that the mapping matrix of BI-RADS features can achieve the best results when both benign and malignant mapping probabilities are taken into account,and the method that the weight coefficient of BI-RADS features decreases linearly in the training process is better.