Prediction Of Near-Term Breast Cancer Risk In Mammography

Breast cancer is one of the most common malignant tumor which seriously threatening female health.The mortality rate of breast cancer can be reduced through early diagnosis and treatment.Mammography is the most popular population-based breast cancer screening modality accepted in current clinical practice.However,the effectiveness of the universal screening for breast cancer is criticized due to the higher false-positive recall rates and lower detection sensitivity.In order to improve efficiency of screening,it is important to develop more effective personalized screening model.Establishing effective and reliable short-term breast cancer risk prediction model is the basis of personalized mammography screening.Traditional breast cancer risk methods have been developed and tested using risk factors,such as family history,genomic information and woman’s age,which unable to accurately predict short-term breast cancer risk.Thus,in this study,we analyzed how to develop short-term cancer risk assessment models to identify a small fraction of women with high short-term risk whom should be more frequently screened for helping reduce false-positive recalls.The main work and contribution of the thesis are as follows:In the procedure of image preprocesses,we used multidirection Gabor filtering approach to detect the breast skin-line for breast region segmentation.An additional step was required to segment breast region from MLO view images,removing pectoral muscle.In this study,we applied gradient and Hough methods to remove pectoral muscle.After breast region segmented,a hesisen matrix and intensity based method were adopt for identification of nipple.Then,the bilateral matched strip regions were anatomically segmented with an anatomical structure based method which take the nipple as the reference point.At the same time,the element regions were computed using the difference of Gaussian based approach to extract the positional information for prediction short-term breast cancer risk.In terms of performing short-term breast cancer risk models,this thesis mainly unfurls the study by four aspects:bilateral asymmetry,multiple-view of mammogramms,dynamic variations of bilateral asymmetry and technology perspective.1.The bilateral local tissue density asymmetry of mammograms is the important sign in clinical practice to assess cancer risk and detect suspicious lesions.While the previous studies only focused on the global bilateral asymmetry of mammogramms to predict short-term breast cancer risk.Thus,in this study,we applied bilateral asymmetry information including global-and local-based features to establish the generalized linear model and the multilayer perceptron based risk models respectively for assessment of short-term cancer risk.Results demonstrate that performance of local region-based bilateral mammographic features is higher than global image features When combining the local and global region based image features,the performance of risk model is improved significantly.2.Most prediction of breast cancer risk studies used features computed only from craniocaudal(CC)view mammography.There are usually two view images including the CC and mediolateral oblique(MLO)view images of two breasts which contain supplementary information.Physician usually diagnose breast cancer from both CC and MLO view images.As a result,in this study we used both CC and MLO view images for prediction of short-term breast cancer risk.Specifically,we first established the CC view based risk model and MLO view based risk model with bilateral asymmetry features.Then we developed two fusion risk models by fusing results of the CC and MLO based risk models with alpha-integration method and multi-agent approach respectively.Results indicate that performance of CC based risk model is higher than MLO based risk model.When adaptively integrating CC and MLO-based risk models could increase the predictive power for predicting short-term breast cancer risk.3.Because development of breast cancer is a gradually progressive process.In addition,the bilateral asymmetry of mammogram varies over time.Thus,we analyzed the relationship between the dynamic variability of bilateral asymmetry in mammogram and the short-term breast cancer risk.Specially,dynamic variability of bilateral asymmetry features were first computed from three most recent prior screening examinations.Then,the correlation between the dynamic variability of features and short-term risk were assessed with univariate regression analysis and multivariate regression analysis.In addition,three risk models were established with three dynamically varying feature sets,namely prior(2 to 1)set(#21),prior(3 to 2)(#32)set and prior(3 to 1)set(#31).Results show that the dynamic variability of bilateral asymmetry is correlated with the short-term breast cancer risk.In addition,performance of risk model built with dynamically varying feature set#31 was significantly higher than risk models established with#21 as well as#32 respectively.4.Recently deep learning technology,in particular convolutional neural networks(CNN)have become a methodology of choice to analyze the medical images.At present,deep learning method has not been applied in extraction of bilateral asymmetry from mammographic images to predict short-term breast cancer risk.Thus,in this study we developed and tested a novel computer-aided-diagnosis scheme with deep learning algorithm to automatically compute bilateral asymmetry for predicting short-term breast cancer risk.Specifically,after bilateral whole breast region and difference of bilateral matched central regions were computed,three deep CNN architectures including AlexNet,GoogLeNet and ResNets were applied to establish risk prediction models respectively.Then,three deep CNN based risk models were fused with multi-agent method for whole breast region and central region respectively.Finally,the whole breast region based risk model and the central region based risk model were integrated for further increasing the accuracy of short-term breast cancer risk prediction.Results show that using of deep learning technology is helpful in predictive of short-term breast cancer risk.In conclusion,the main innovations of this thesis are as follows.First,compared to exiting epidemiological long-term breast cancer risk prediction models,the models established in this thesis were used for prediction of short-term breast cancer risk.Second,we first introduced the local region based bilateral asymmetry information into short-term breast cancer risk assessment models.Third,we investigated a new scheme to improve the accuracy of short-term breast cancer risk prediction by using information on CC and MLO views of left and right breasts.Fourth,we investigated the relationship between dynamic variability of bilateral asymmetry features and short-term breast cancer risk.Fifth,we developed and tested a novel short-term breast cancer risk prediction model applied deep learning based algorithm.This study helps advance on individualized breast cancer screening to improve detection rate,reduce false positive rate and medical expenditure.