Development and Evaluation of Computerized Segmentation Algorithm for 3D Multimodality Breast Images
thesisposted on 2014-06-20, 00:00 authored by Hsien-Chi Kuo
Breast cancer is the 2nd most common cancer among US women. Until now, mammography still has been a widely accepted screening tool for breast cancer. However, mammography projects 3D tissue structures of the breast onto a 2D plane and results in superimposition effect which leads to misdiagnoses. Recently, researchers have been developing CT systems (bCT) and automated 3D breast ultrasound (ABUS) dedicated solely for breast imaging. Such imaging modalities generate 3D image volumes that completely resolve breast tissue structures and avoid the superimposition effect. However, it also produces large amount of image data that the radiologists need to review. Such data explosion could make image interpretation task even more difficult and time consuming. Therefore, CAD (computer-aided detection/diagnosis) technology is expected to alleviate the burden. In this study, a technique for CAD application to bCT and ABUS is developed. We aim to propose an automated segmentation procedure and computer classification for the images where the lesion center has been labeled. The lesion segmentation algorithm combines radial gradient index segmentation and modified level set based active contour algorithm. Satisfactory segmentation results were obtained on both contrast-enhanced and unenhanced bCT, as well as ABUS in terms of the measure of the overlap of computer segmentation and manually-delineated lesion outlines. The computer classification was done lesion feature analysis. Among those extracted features, lesion shape features (irregularity measures) showed that the proposed segmentation algorithm was able to capture sufficient shape information (are under the receiver operating characteristic curve, AUC = 0.81), which is considered one important factor for differentiating tumors. In addition, we firstly developed a new 3D spiculation feature for bCT image volumes in order to further improve the classification performance. This new spiculation feature utilizes the 3D structural information in the lesion neighborhood to analyze the lesion surface and evaluate the degree of spiculation. By adding the new spiculation feature, AUC was improved from 0.81 to 0.85 significantly. The results suggest that the development of such feature which utilizes 3D information resolved by 3D imaging modalities should be further investigated for future CAD application.