Research On Bias Field Correction And Lesion Classification In Mri Medical Image Diagnosis

In recent years,the incidence and mortality of liver cancer and breast cancer have been increasing year by year.As the main method of clinical auxiliary examination,medical imaging technology provides a reliable basis for early cancer detection and diagnosis.Among them,due to high soft tissue resolution and no radiation,Magnetic Resonance Imaging(MRI)has become one of the most important screening tools for clinical liver cancer and breast cancer.At present,Computer Aided Diagnosis(CAD)system based on MRI aims at assisting doctors in making an efficient and accurate diagnosis is widely used.However,due to the poor quality of collected images or low classification accuracy of the existing CAD for those hard-to-spot lesions,there are still some challenges in the research on CAD.To this end,this paper focuses on the study of liver MRI bias field correction and breast dynamic contrast enhanced MRI(DCE-MRI)benign and malignant lesions classification in the construction of CAD system.And the main work is as follows:(1)Bias field correction of liver MRI based on original intensity preserved.To correct the bias field of liver MRI,an automatic fuzzy c-means clustering algorithm with constraint term of bias filed is proposed.The proposed method firstly applies fuzzy membership mask to remove background noise.Then the local intensity and the global intensity information is taken into account to ensure the smoothness of bias field.And the spatial weight is combined to prevent liver edge from blurring.Furthermore,a novel constraint term is proposed to ensure gray level of corrected image similar with original image.And the results of experiment demonstrate the advantages of the proposed method compared with other bias field correction methods.(2)Comprehensive hierarchical CAD system based on multi-features of breast DCE-MRI for cancer diagnosis.To improve the classification accuracy of those hard-to-spot lesions,a comprehensive hierarchical breast CAD based on multi-features is proposed.The experiment results demonstrate that the proposed method is not only for common lesions but also for those hard-to-spot lesions.Furthermore,a better classification performance is obtained with the two proposed new texture features which effectively represent the characteristic of lesion margin.(3)Research on data augmentation in breast benign and malignant classification based on deep learning model.Due to the small number of medical images with specific diagnosis conclusions,the development of deep learning in building a CAD of medical images is limited.Based on the data augmentation methods commonly used in medical images,three augmentation methods are used in this paper: rotation at a large angle,rotation at a small angle and horizontal translation in order to study the performance of deep learning model with datasets obtained from different methods.The experimental results show that under the same number of datasets,the deep learning model trained by the dataset obtained from horizontal translation has a higher test classification accuracy than the other two augmentation methods.