Research On X-ray Breast Mass Segmentation Based On Multi-scale Adversarial Networks

Breast cancer is the most common malignant tumor disease threatening the lives and health of women around the globe.According to epidemiological statistics,the morbidity of breast cancer has been rising in successive years and the age of onset has become younger.There is no ideal treatment plans for patients with advanced breast cancer in modern clinical medicine.And timely detection and intervention in earlystage is still the most scientific and effective way to save the lives of patients.Known as one of the most reliable noninvasive breast examination methods in the world,mammography has the characteristics of clear imaging and high sensitivity,which is able to accurately detect small nidus and hidden nidus that are difficult to find by means of traditional palpation.At present,mammographic X-ray images seriously rely on the analysis and diagnosis of radiologists,which leads to low film reading efficiency and easy to be affected by subjective factors.Moreover,the structures of mammary glands are complex,and the size,shape,and texture of breast masses are rather changeful.In the process of reading films,doctors’ energy will be greatly consumed,which may even affect the accuracy of diagnosis results.With the advent of the era of big data,the acquisition cost of medical images is rapidly reduced,and the computer-aided diagnosis(CAD)system emerges as the times require.Breast mass image segmentation can quickly and accurately extract mass areas from the complex and changeable medical images,which greatly reduces the burden of doctors on film reading,and provides an important reference for the benign and malignant judgment of breast masses.The traditional techniques for breast mass segmentation mainly adopt hand-crafted features to mark the boundary of the mass.Limited to the expression capability of hand-crafted features,these methods are usually difficult to detect the complex edge.The powerful feature extraction of convolutional neural networks can make them break through the bottleneck of traditional techniques.However,the way of independent pixel classification forces the convolutional neural network unable to maintain the high-order consistency of images,which hinders the further improvement of segmentation accuracy.In order to handle the above issues,this paper introduces the idea of adversarial learning into breast mass image segmentation,and designs and implements a novel multi-scale adversarial network for precise segmentation of breast mass in X-ray images.This network model is composed of one segmentation network and one discrimination network.The segmentation network is responsible for dividing breast mass areas from original input images,and the discrimination network correspondingly scores output maps from the segmentation network,so as to further optimize the adjustable parameters.The main contents of this paper are as follows:(1)Summary of the research background and purpose,and a review of related work in recent years.This paper lays special stress on analyzing the development trend of breast mass segmentation techniques and discussing the advantages and disadvantages of traditional image segmentation techniques and image segmentation based on deep learning.(2)An introduction to basic theoretical knowledge of deep learning.We mainly focus on convolution neural network,generative adversarial network,and their relevant background theories.(3)The details of the breast mass segmentation method based on multi-scale adversarial learning.To capture long-range dependency and reinforce spatial contiguity of output maps produced by segmentation network,this work proposes to introduce adversarial training mode to semantic segmentation.We employ a set of multi-scale discriminators to adapt breast mass with different size and shape.A weighted loss function is designed to alleviate the class imbalance,and the EM distance metric is used to narrow the data distribution of the segmentation mask and the real label.(4)Demonstration and analysis of experiments.We design some comparison experiments on two publicly available mammography screening datasets,INbreast and CBIS-DDSM,to verify the effectiveness of the proposed method.