A Study On The ELM-based Breast Tumor Detection Technology

The breast cancer is becoming one of the most common malignant tumors that seriously affect both the mental and physical health of women, even risking their lives. The key point on enhancing the cure rate and reducing the death rate is to detect the tumor in the early stage. Mammography has been the most widely used detection method in clinical on woman breast disease. However, manually analyzing the mammograms is always time-consuming and human-fatigued, and this will result in the high missed diagnosis and misdiagnosis rate. Tumor, as the most common feature of the breast cancer, can be detected by using computer-aided breast tumor detection system (BTDS), which helps radiologists to shorten the image analyzing time, reduce the fatigue, and improve the precisely diagnosis rate. However, traditional machine learning based method on computer-aided breast tumor detection (BTD) can be improved in accuracy and training efficiency, and currently no good solution in large-scale dataset. The Extreme Learning Machine (ELM) has provided an effective way to BTD and it can overcome the above problems. In this thesis, a series of ELM-based BTD technologies are presented based on the current tumor detection and machine learning. The main contributions of this thesis are listed below:(1) A unilateral single view (USV) ELM-based BTD method. Base on the mammograms, we construct a feature model on the suspended tumor region (including shapes and textures). Hence, a unilateral single view ELM-based BTD method is proposed. Compared to Support Vector Machine (SVM), it has significant improvement in training speed, detection accuracy and etc.(2) A unilateral double views (UDV) ELM-based method. Combining CC view and MLO view, we construct a feature model on the feature vector of unilateral single view and unilateral double views. After the feature selection and the ELM-based tumor detection, the accuracy has been improved compared to the unilateral single view BTD method.(3) A bilateral single view (BSV) ELM-based BTD method. A similarity feature dataset was constructed and the double sides of the unilateral single view feature were combined, we construct BSV feature model. After the feature selection and the ELM-based tumor detection, the accuracy has been improved compared to the unilateral single view BTD method.(4) A distributed ELM method based on MapReduce framework. Base on the analysis, we find that the most computational consuming part among ELM calculation is in matrix multiply operation and theoretically prove that this operation can be isolated. We hence introduce a distributed ELM (ELM*) method based on MapReduce framework and experimental results show that this method can efficiently learn among large-scale training dataset.(5) A performance enhancement approach for ELM. We prove in theory that the most computational consuming part among ELM* calculation can engage incremental, decremental and error corrected calculation. We hence propose a performance enhancement approach for ELM* and experimental results show that this method can enhance the learning functionality of the ELM* on dataset updates.(6) An ELM-based BTD prototype system. We design and implement an ELM-based BTD prototype system, and through the system verify the effectiveness of the proposed theory and method.Overall, this thesis studies the key technology on ELM-based BTD and introduces ELM-based BTD on USV, UDV, BSV, and effectively enhances the training speed and detection accuracys. Moreover, we present ELM* and performance enhancement approach for ELM and utilize them to solve the problems on large-scale dataset learning and incremental learning, and hence enhance the extension of the traditional ELM. Finally, we implement an ELM-based BTD prototype system and prove the effectiveness of the above presented methods by applying a mass of experiments and analysis on the system. The study of these methods has significant academic value and practical application on computer-aided BTD technology.