Computerized analysis and interpretation of breast MR images

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) of the breast is being increasingly used in clinical breast cancer imaging. However, interpretation of 4D DCE-MRI data is labor-intensive and challenging for radiologists. The overall objective of this research is to investigate an automatic and efficient computerized system for DCE-MRI interpretation in terms of an estimate of the probability of malignancy.; Two clinical breast DCE-MRI databases were acquired using a T1-weighted 3D spoiled gradient echo sequence. Database 1 included 77 malignant and 44 benign breast lesions and Database 2 included 97 malignant and 84 benign breast lesions, all of which were pathologically verified for benignity/malignancy truth.; The main contributions of the dissertation are summarized as follows. (1) A novel lesion segmentation algorithm was developed and evaluated by comparing it with manual delineation by an expert radiologist. The algorithm significantly improved the performance of a previously developed volume-growing method. (2) An automatic and efficient approach was investigated to identify the characteristic kinetic curve (CKC) from the 4D lesion image data. The diagnostic performance of kinetic features extracted from computer-identified CKCs was found to be significantly better than those extracted from averaging over the entire lesion, and was found to be comparable to features extracted from CKCs generated from manual ROIs. (3) A volumetric texture analysis method in the gray-level co-occurrence matrix framework was developed to quantify lesion enhancement patterns. Differential diagnosis performance of texture features based on 3D segmented lesions was found to significantly improve that based on 2D rectangular ROIs. (4) A computerized benignity/malignancy classification scheme using an automatic relevance determination Bayesian neural network was evaluated with two clinical databases. Cross-validation results showed that the classification system is efficient and robust across different MRI scanners. Az values of 0.87 (s.e. 0.04) and 0.83 (s.e. 0.03) were achieved in independent testing on two clinical breast MR databases respectively.; The significance of this research is that it provides a computerized system for breast DCE-MRI lesion segmentation and characterization that could potentially aid radiologists in achieving an improved interpretation and workup of breast MR images in terms of efficiency, consistency, and accuracy.