Research Of Key Technologies Of Digital Mammography Mass Classification

Breast cancer is one of the leading causes of death among women worldwide.In2020,the number of new cases of breast cancer globally has surpassed that of lung cancer,and breast cancer has become the world’s largest cancer.Early screening and treatment are considered to be the main ways to improve the survival rates of breast cancer.Mammography is currently the most reliable and cost-effective method for diagnosing breast cancer,and it is widely used to detect abnormalities in the breast.Among the multiple abnormalities of the breast(e.g.,masses,microcalcifications,architectural distortions,and asymmetry),breast masses have the characteristics such as irregular size,low contrast,and easy to be surrounded by similar backgrounds,all of which can affect the accuracy of diagnostic results.In this paper,we take into account the shape irregularity and low contrast of breast masses,and address the problem of segmentation and classification of breast masses by combining ideas in computer vision,machine learning and other fields.The main research contents of this paper are as following:1)We proposed a classification model of benign and malignant breast masses based on feature fusion.An existing problem with current convolutional neural networks(CNNs)is the lack of local invariant features to effectively respond to changes caused by geometric transformations or imaging angles of mammogram images.Therefore,a new model with texton representation and deep CNN feature representation is proposed in this paper for the breast mass classification task.The rotation-invariant features provided by the maximum response filter bank are incorporated into a CNN-based classification,and the methods of direct fusion and dimensionality reduction and fusion are used to address the defects of CNN in extracting mass features.This method is validated on CBIS-DDSM and a combined dataset(mini-MIAS and INbreast).Based on the CBIS-DDSM dataset,in terms of AUC(0.97),accuracy(94.30%)and specificity(97.19%),the fusion after implementing the reduction approach is better than other mass classification models.2)We proposed a model of breast mass segmentation and shape classification based on adversarial learning.Aiming at the problem of breast mass shape classification,we propose an accurate mammography mass segmentation model based on improved conditional generative adversarial networks(c GAN).In the decoder of c GAN,a superpixel average pooling layer is introduced,and the superpixels are used as a pooling layout to enhance the boundary segmentation on the mass.In addition,a multiscale input strategy is used to enable the network to learn robust and scaleinvariant breast mass features.The segmentation model was evaluated on two public datasets: CBIS-DDSM and INbreast.Dice and Jaccard scores of 93.37% and 87.57%,respectively,are obtained for the CBIS-DDSM dataset.The Dice and Jaccard scores for the INbreast dataset are 91.54% and 84.40%,respectively,which exceeded the current state of the art mass segmentation.The superpixel average pooling layer and multi-scale input module increased the Dice and Jaccard scores of original c GAN by7.8% and 12.79%,respectively.Based on the generated breast mass segmentation mask,a CNN-based shape classification model is proposed.The generated segmentation mask is divided into four types of mass shapes: round,oval,irregular and lobular.The proposed shape descriptor was trained on CBIS-DDSM and produced an overall accuracy of 80%,exceeding the current state-of-the-art.