Date of Degree
PhD (Doctor of Philosophy)
Joseph M. Reinhardt
Breast cancer is the second leading cause of cancer deaths in women today. Currently, mammography is the primary method of early detection. However, research has shown that many cases (10-30%) missed by mammography can be detected using breast MRI (BMRI). BMRI is more difficult to interpret than mammography because it generates significantly more data. Also, there are fewer people qualified to use it for diagnosis because it is not the standard breast imaging modality.
Our goal is to develop and test a CAD system to aid and improve the performance of radiologists with different levels of experience in reading breast MR images. Part of the CAD system is an image loader and viewer capable of displaying multiple sequences simultaneously, with standard region of interest and high level analysis tools. We propose a semi-automatic segmentation method that identifies significant lesions. Then, 42 shape, texture, and enhancement kinetics based features were computed. The top 13 best features were selected and used as inputs to three artificial classifiers: a backpropagation neural network (BNN), a support vector machine (SVM), and a Bayesian classifier (BC). Each one was trained using pathology results as the gold standard. Five human readers (a BMRI expert, two mammographers, and two body imaging fellows) manually classified 75 BMRI datasets (80 lesions), both with and without CAD system assistance. The performance of the computer classi- fiers and human readers were compared using ROC curves, and the human readers' performance was also evaluated using MRMC analysis.
The ROC curve analysis showed that the BNN system significantly outperformed the other two classifiers with Az = 0.970, and p < 0.05, and a sensitivity of 91.3% with zero false positives. Also, all human readers significantly improved when aided by the CAD system (p < 0.05). The MRMC analysis showed that the human reader performance with and without CAD system assistance can be generalized over the population of cases and still maintain a statistically significant improvement (F(1, 74) = 6.805, p = 0.0110 < 0.05). These results show significant advantages to using CAD systems in classifying BMRI lesions.
2, xiii, 148 pages
Includes bibliographical references (pages 140-148).
Copyright 2005 Lina Arbash Meinel