Deep Learning Based Breast Lesion Detection And Classification On Transversal And Longitudinal Ultrasound Sequences

Breast cancer is the most common cancer among women in the world,and it is also has the highest mortality rate.The incidence and mortality of breast cancer are increasing every year.In recent years,China ranks among the top 10 in the world in the new incidence and mortality of breast cancer.Breast cancer has become the leading killer of women’s life and health.Unlike most cancers,early-stage breast cancer is treated well.A large number of scientific research literature and clinical experience show that early stage screening and treatment is the important method to improve the survival rate of patients,"early detection,early diagnosis,early treatment" is particularly important,which is the key to reduce the mortality rateCommon clinical screening methods include mammography,breast ultrasound and magnetic resonance imaging.Ultrasound imaging has become an important supplementary method for mammography examination because of its advantages such as convenience,low cost,no radiation and real-time ability.In clinical practice,reading ultrasonic imaging calls for doctor who is clinical experienced.In order to improve diagnostic accuracy and reduce the misdiagnosis and misdiagnosis in early screening stage,two or more experienced doctors are needed to give the multi view and multi-modal comprehensive diagnosis of the breast lesion area,which increases the reading time and leads to the variance of the diagnosis sensitivity.In addition,due to the different setting conditions of multi-brand devices and different image preprocessing methods,the geometric center and the contrast of the images are greatly different,which affects the final diagnosis result.Due to the urgent need of clinical diagnosis,CAD,a computer-aided diagnosis system,which can automatically locate and diagnose the focal areas in ultrasound images,was invented.The CAD can automatically segment and classify breast lesions,assist diagnosis,greatly shorten the reading time for imaging,improve the efficiency of breast diagnosis and improve the survival rate of patients Traditional breast CAD is based on manually designed feature descriptors and machine learning classification methods.It takes a lot of time and experience to design features in the field of breast tumors,which is a time consuming process.There are too many human intervention factors in traditional feature engineering,which can only be designed for a specific data set and lacks the model generalization abilitySince 2012,deep learning technology has been an important role in all walks of life,and its performance has surpassed that of traditional algorithms in many fields.A major advantage of deep learning technology is that it can automatically learn and extract high-level semantic features from data such as images,freeing researchers from the tedious work of feature engineering.Researchers can now focus on the process of modeling with mathematics.Although many scholars have applied deep learning tools to the task of benign and malignant discrimination of breast tumors and achieved good results,previous work did not cover effective feature fusion for multi-view and multi-modal breast ultrasound images.As we all know,deep learning is a data-driven science,and the quality and quantity of data sets affect the final performance of the model extremely.Multi-view and multi-mode data supply a lot of input images to train the model effectively.At the same time,the way that combine different features from different modes has become an urgent problem to be studied and solvedIn this thesis,multi-view and multi-mode ultrasonic images are preprocessed to distribution uniformly.Then,a feature fusion network based on multiple models was built to fuse the features learned from multimodal ultrasonic images.Experiments were conducted to compare different fusion strategies,including simple fusion methods such as mean,maximum and sum,as well as fusion methods based on convolution and bilinear method.In order to provide doctors with object visualization results and improve the accuracy of feature fusion,an ultrasonic image segmentation and classification network based on object detection is designed in this study,so that the features of tumor objects can be fused and background noise can be reduced.The network can provide visualization results of high-resolution segmentation of tumor objects,and the segmentation results can be involved in subsequent classification tasks to get a better classification result.By adding the object detection module into the network,the tumor object can be tracked effectively and the characteristics of tumor ROI can be learned.Based on the object detection module,the background noise can be filtered out and leave a pure feature of the object itself.The fusion experiment is conducted in this study.The experimental results show that the classification performance of ultrasonic image based on object detection and segmentation is better than that of the simple fusion network and achieve an accuracy of 94.59%.At the same time,it provides visualization results,which can provide auxiliary diagnosis to doctors The classification of ultrasonic image based on object detection and segmentation is more efficient and excellent than common image classification method.