Breast Mass Detection And BI-RADS Classification Algorithm Based On Multi-source Heterogeneous Data

Breast cancer is the most common malignancy in the world.The survival rate of breast cancer patients can be improved by the early accurate diagnosis.Mammography is an important method for early screening of breast cancer.Clinical imaging diagnosis doctors often judge mass by comparing observation and analysis of multi-view mammography,and make classification diagnosis of the mass found by Breast Imaging Reporting and Data System(BI-RADS).However,there is an urgent need to use computer-aided diagnosis system(CAD)to improve the work efficiency due to the shortage of radiologists with the rapid growth of the demand for imaging examination.Detection and classification of benign and malignant breast mass based on deep learning has become a research highlights in recent years.However,there exists the problem of the high false positive rate in most of the current methods where the single mammogram is used for the breast mass detection.On the other hand,the classification of benign and malignant mass is a little significance for the imaging diagnosis doctors.Therefore,the breast mass detection and breast imaging reporting and data system classification algorithm based on multi-source heterogeneous data are proposed in this thesis.The main contents of this thesis are as follows:(1)Aiming to the requirements of the high-precision and high-efficiency clinical in the breast mass detection tasks,a single-view mass detection algorithm based on deep learning is proposed.The YOLOv3 algorithm with better comprehensive performance was selected and improved through the theoretical analysis and experimental comparison of deep learning detection algorithm,and the improvements to YOLOv3 are as follows: a)A multi-scale local feature fusion module was designed to improve the detection sensitivity of small mass.b)The detection specificity is improved by adding the focusing parameters in the loss function for the problem of uneven positive and negative samples caused by small mass proportion.c)K-means algorithm was used to reset the anchor box to improve the accuracy of detection position in response to the diversity of mass size.The experimental results show that the method presented in this thesis can effectively improve the ability of detecting small mass.(2)In order to reduce the high false positive rate of single-view algorithm for breast mass detection,an algorithm for breast mass detection based on dual-view was proposed.False positives were screened by modeling on dual-view to expand the supervised learning information for the deep learning algorithm in the data layer: a)The Sech template was established to search the region of interest by cross-correlation method.b)A matching model was established on the dual-view to screen false positive according to the chest wall line and nipple position and the database of the dual-view was labeled together with the mass.c)The improved YOLOv3 algorithm was used to complete the training and modeling of the dual-view database.d)The coordinate normalization of test results is carried out at the decision-making layer to further reduce false positive.The experimental results demonstrate that the sensitivity can be improved by the dual-view method and the false positive rate can be reduced significantly.Moreover,the comprehensive performance of the model is improved considerably and the dual-view method has good universality in different database.(3)In response to the clinical needs of multi-classification of breast mass,a classification algorithm of BI-RADS for breast mass based on multi-source heterogeneous data was proposed.Multi-source heterogeneous data,composed of images and non-images and whose features were extracted by Graph Convolution Neural Network(GCN),were constructed to complete the BI-RADS classification of mass: a)A multi-source heterogeneous database was constructed based on the image of the mass region obtained by the dual-view detection algorithm and non-image data such as the patient’s age and breast density.b)The graph is constructed by multi-source heterogeneous data group.The feature of mass region image was extracted by residual network as node of the graph.Non-image data was used to interpret the correlation between target feature as the edge of the graph.c)The graph was fed into a GCN composed of two hidden layers and activation functions to complete the BI-RADS classification of the mass.Experimental results verify that the proposed method can effectively improve the accuracy of BI-RADS classification.The sensitivity and specificity of the proposed method reached 92.0% and 87.7%,with an average of 0.041 false positives per image and an average accuracy of 85.9%for BI-RADS classification.The proposed method has certain theoretical significance and clinical application value.