Detection And Recognition Of Architectural Distortion In Mammography

Breast cancer is a leading health threaten for women in the world,early diagnosis and treatment are critical.Mammography is one of the most commonly used screening methods.Among the several abnormalities observable on mammograms,architecture distortion is one of the most difficult to detect due to its subtlety.Computer-Aided diagnosis technology can improve the performance of diagnosis in breast cancer.Machine learning has been widely applied,such as Bayes classifier,artificial neural network and so on.SVM is the most commonly used technology when dealing with small samples problem.In recent years,Twin SVM has been developed rapidly as a new support vector machine and has been widely used in many fields.Two bound support vector machines(TBSVM)is the development of TWSVM.In this thesis,a new automatic architectural distortion detection method for breast cancer in mammographic images is proposed.Firstly,use Gabor filters and phase portrait analysis to locate the suspicious area.For each suspicious region,several features are extracted.However,not every extracted feature contributes to the classification accuracy.We proposed a novel feature selection method for TBSVM and utilized it for the architectural distortion detection named Multiple Twin Bounded Support Vector Machines Recursive Feature Elimination(MTBSVM-RFE).The results showed that our proposed method can detect the region of architecture distortion with high accuracy.In recent years,deep learning has become the mainstream technology in large data and artificial intelligence research.Deep learning requires a larger training data set.However,in Digital Database for Screening Mammography(DDSM)dataset and mini-MIAS dataset,the sample of architectural distortion is insufficient.In this thesis,we combine the convolutional neural network(CNN)and twice transfer learning method to recognition architectural distortion.We first use model trained on the Image Net to initialize the CNN model and use the mass and normal breast to fine-tuning the model(the first transfer learning),then,use architectural distortion and normal breast to fine-tuning the CNN model again((the second transfer learning)).The result shows the optimization of the CNN model is more suitable for the recognition of architectural distortion by fine tuning the network twice.