Automatic Tumor Detection for Dynamic Contrast-Enhanced Magnetic Resonance Imaging

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is technique to form images according to the variation information of concentration of the contrast agent, which shows different enhancement in different kinds of tissues. A four-dimensional (4-D) image from each patient is acquired by scanning patient at different time and a 4-D image contains dozens of three-dimensional (3-D) images. In recent studies, DCE-MRI is the most sensitive tool to detect breast cancer by radiologists for clinical practice. Because an acquisition of 3-D DCE-MRI consists of dozens of 2-D images, it is time-consuming to find the tumor or diagnose the cancer from such large amount of images by physicians and the misdetection might occur due to physicians’ fatigues. In order to make inspection of DCE-MRI images more efficient, computer-aided detection (CADe) systems have been proposed to interpret the images. CADe can not only shorten the interpretation time but also probably detect the tumor which could be not found by physicians. However, it is difficult to implement a CADe system to detect tumors efficiently from a large number of 3-D DCE-MRI images. To date, no related studies for development of CADe systems on DCE-MRI images have been reported to detect breast lesions. In the thesis, we proposed two CADe systems for DCE-MRI images. In the first study, the tumor detection based on the two-stage technique was applied to distinguish breast masses from normal tissues using binary logistic regression. The tumor regions which consist of enhanced tissues were detected at first stage and then the second stage were combined to identify the suspected regions which could not be identified easily. The results from two-stage detection algorithm were considered as the detected masses. In the second study, the tumor detection based on fuzzy c-mean (FCM) was proposed to locate enhanced regions firstly on DCE-MRI. Because of the blob-like characteristic of breast mass, we detected masses from the enhanced regions by the Hessian method. The results showed that the two CADe systems have high detection performance for the breast tumor detection on DCE-MRI images.誌謝 II 摘要 III Abstract V Contents VIII List of Figures XII List of Tables XV Chapter 1 Introduction 1 1.1. Research Motivation 1 1.2. Issue Descriptions 2 1.2.1. Logistic Regression Technique 3 1.2.2. Fuzzy C-means Technique 4 1.2.3. Hessian Analysis Technique 4 1.3. Organization 5 Chapter 2 Review of Three-Dimensional DCE-MRI Techniques 6 2.1. Introduction 6 2.2. Related Studies 7 2.3. CADe System Procedures 9 2.3.1. Preprocessing 10 2.3.1.1. Motion Correction 10 2.3.1.2. Nonuniformity Bias Correction 11 2.3.2. Segmentation 11 2.3.3. Detection 13 2.3.4. Reduction 13 Chapter 3 Computerized Breast Lesions Detection Using Kinetic and Morphologic Analysis for Dynamic Contrast-Enhanced MRI 15 3.1. Data Acquisition 16 3.2. Motion Correction 18 3.3. Breast Region Segmentation 18 3.4. Feature Extraction 22 3.5. Suspicious Tissue Detection 22 3.6. False Positive Reduction 24 3.7. Statistical Analysis 27 3.8. Results 28 3.8.1. Detection results of trained cases 28 3.8.2. Detection results of test cases 29 3.8.3. False positive reduction by morphologic features 30 3.8.4. Detection performance analysis by FROC curve 31 3.8.5. Analysis for number of suspicious regions 31 3.9. Discussion 42 3.10. Conclusion 46 Chapter 4 Computerized Breast Mass Detection Using Multi-Scale Hessian-Based Analysis for Dynamic Contrast-enhanced MRI 48 4.1. Materials 49 4.2. Methods 49 4.2.1. Motion and Nonuniformity Bias Correction 52 4.2.2. Breast Region Segmentation 54 4.2.3. Suspicious Tissue Detection 56 4.2.4. Breast Mass Detection 60 4.2.5. Mass Selection from Candidates 62 4.2.5.1. Mass Selection by Blob and Enhancement Features 63 4.2.5.2. Mass Selection by Morphologic Features 64 4.2.5.3. Mass Selection by Texture Features 65 4.2.6. Statistical Analysis 67 4.3. Results 68 4.3.1. Statistical Analysis for Mass Selection 68 4.3.2. Mass Selection by Different Features 71 4.3.2.1. Mass Selection 72 4.3.2.2. Detection Performance Analysis with FROC Curves 73 4.3.2.3. Detection Performance Analysis According to Different Size Groups 76 4.4. Discussion 76 4.5. Conclusion 82 Chapter 5 Conclusion and Future Directions 83 Publications 10