Computer Diagnosis Algorithms For Breast Masses Based On Visual Cognition Model

Breast cancer is one of the leading women's cancers worldwide.It is important to detect,diagnosis and treat cancer at their early stage for improving the cure rate of breast cancer and the quality of life.Medical images are suitable for detecting occult disease in early stage.Therefore,a very important aspect to develop computer aided diagnosis systems is the assistance they provide to doctors.Breast mass is one kind of sign to assess the state of breast cancer,and it is a hotpot that how to apply computer to detect and diagnose breast cancer.At present,the existing computer-aided diagnosis methods for breast masses have a number of problems,such as high false positive rate,low recognition rate.Therefore,these methods can't meet the demand of clinical practice.The major reasons are organized as follows:Firstly,the diagnosis of masses is challenge due to the adhesive property between the mass-es and normal tissue,the varying size,shape and appearance of the masses.Secondly,the computation model of bottom-up mechanism is adopted by traditional computer diagnosis systems which lack the priori knowledge of radiologists and can not produce the consistent results with radiologists.Accordingly,a deep research for breast masses diagnosis is made from the perspective of visual cognition model on X-ray image and Magnetic Resonance Image.The research contents include:mass detection,mass segmentation,benign or malig-nant diagnosis of mass.Overall,two complete systems are presented in this paper to solve problems in different application areas.System One is called mass diagnosis method based on multi-source image features.Firstly,visual rules are modeled based on Gestalt psychol-ogy to achieve mass detection and segmentation.Secondly,a weighted ensemble method for semantic feature vectors is proposed for mass diagnosis across image modality.It can also be used to other tumors diagnosis,which is of both theoretical and clinical importance.System Two is called mass diagnosis framework based on multi-view fusion and deep learn-ing for breast X-ray images.It models the contextual information from multi-view images and improve the performance of mass diagnosis.It is meaningful for the application of deep learning in medial image processing.The main innovations and contributions are as follows:(1)In order to solve the problem of high false positive rate in the methods of mass detection,an automatic mass detection method based on Gestalt psychology is presented.Firstly,the characteristic of reading mammograms for radiologists are analyzed.Secondly,the visual characteristic of radiologists is modeled based on Gestalt visual rules.Kinds of visual rules in the proposed framework are used to screen and identity the candidate mass regions.Not only the method makes use of the advantage of machine learning methods,but the visual rules based on Gestalt is also modeled,which achieves significant improvement in reducing false positives and improving sensitivity.(2)In order to achieve the goal of automated segmentation of the breast mass,we propose a novel breast mass segmentation method based on prior experience of radiologists and visual patches.First,the visual patches are generated as basic processing units which extends the semantic expressive power of pixels in an image.Second,the segmentation process is performed by profiting from the idea of graph cut algorithm.Prior experience of radiologists is also integrated into the proposed method.The final segmentation results are obtained by a global optimization method.The experimental results show that the proposed method yields higher decent segmentation performance and stronger robustness than other algorithms.(3)For the purpose of integrating the multi-source features derived from multi-source med-ical images and promoting the performance of mass diagnosis,a weighted ensemble clas-sification framework for groups of semantic feature vectors is proposed.Inspired by the cognition process of the radiologists,the semantic features are extracted from multi-source images.In order to solve the challenge of feature heterogeneity,different kinds of semantic features are grouped and modeled by different kernels.Lastly,a weight integration strate-gy is used for the integration of different kinds of semantic features and for the diagnosis of breast masses.Experiments have demonstrated that the proposed algorithm outperforms several stage-of-art classification algorithms.(4)In clinical practice,radiologists combine information from multi-view mammograms to make mass diagnosis.Inspired by this fact,a hybrid deep network framework is presented,aiming to efficiently integrate and exploit information from multi-view data.Innovatively,the proposed framework solve the difficulties of information integration from different in-puts in traditional deep learning methods.Firstly,a generic Convolutional Neural Network(CNN)is adopted to extract features from single view mammogram.Secondly,mass features from multi-view data are effectively aggregated by a Recurrent Neural Network(RNN).The RNN learns the semantic label dependency among different views and makes a final mass classification from multi-view data.We justify the proposed framework through extensive experiments.We achieve a better performance than stage-of-art classification algorithms.(5)To address the problem of "black-box" operation in existing CNNs,a attention-based deep learning method is proposed for masses diagnosis,after performing a series of visual-ization analysis on CNNs.Firstly,the attention-based method distills the mass features such as shape and location information of the masses.Secondly,the proposed method is used to refine the mass features of CNN.It helps CNN shift its attention to the regions of mass and extract informative features for diagnosis.After plenty of validation,the proposed method greatly improve classification performance on both single-view data and multi-view data.This content of this dissertation belongs to the interdisciplinary of information engineering and biomedical science.The methods proposed in the dissertation partly solve some prob-lems in computer diagnosis for breast masses.Starting from human vision and radiologists'knowledge,promising results of mass diagnosis are obtained based on human vision cogni-tion model.These methods improve the accuracy and efficiency of mass diagnosis,to some extent,which meet the radiologists' requirements,and have a good prospect for clinical ap-plication and social benefits.