Brain Tumor Segmentation And Recognition In Computer-aided Diagnosis

Brain tumor is one of the most serious disorders globally,and can result in several health problems even death.Precision diagnosis is key to provide treatment and prevention to patients,and can be divided into two categories: invasive and non-invasive diagnosis.In clinical settings,non-invasive diagnosis by brain medical images is more common and intuitive since it does not harm body in any way.Unfortunately,long-term image visual inspection is extremely laborious for physicians,and leads to misdiagnosis.This has motivated researchers to develop computer aided system for tumor diagnosis.Recently,brain tumor segmentation and recognition are two hot research topics in tumor diagnosis.In this thesis,two brain tumor segmentation methods are firstly proposed based on Geodesic Graph Cuts and Convolutional Neural Networks(CNN),respectively.Then,a malignant brain tumor recognition method is proposed which combines the optimized CNN model with Support Vector Machine(SVM).Finally,a brain tumor aided diagnosis system is developed based on the aforementioned methods.The main work completed in this thesis is listed as follows: 1.Brain tumor segmentation1).A graph-based brain tumor segmentation method is proposed to tackle the challenge of algorithm flexibility.Specifically,an interactive algorithm based on Geodesic Graph Cuts is proposed,which incorporates expert knowledge and does not require model training.Experimental results show that the proposed method not only achieves high accuracy,but also fits for practical applications.2).A CNN-based deep learning network is proposed for brain tumor segmentation.In traditional methods using handcrafted features,it is hard to yield significant improvement in terms of accuracy.To address this problem,a two-scale multi-modal CNN model is proposed to segment brain tumor utilizing self-extracted features.This model automatically extracts spatial features from different scales and modalities,in order to capture both local and global characteristics.Extensive experimental results demonstrate that the proposed model is superior to the baseline methods.2.Malignant brain tumor recognitionAn integrated brain tumor recognition approach is proposed to distinguish between the benign and malignant brain tumors.This method combines the optimized CNN model with SVM to fully utilize the strengths of them.Experimental results justify the effectiveness of the proposed model.3.Design and implementation of brain tumor aided diagnosis systemA brain tumor aided diagnosis system is designed and implemented based on the aforementioned methods.The system is constructed using B\S framework and contains two function modules,namely tumor aided diagnosis module and brain image management module,respectively.Practical application demonstrates that our developed system can provide integrated healthcare information to physicians for diagnosis,and hence improve the clinical treatment of brain tumor disease.