Computerized Interpretation of Dynamic Breast MRI
Abstract
In the past three years we have investigated computerized methods for analyses and interpretation of breast MR images. We investigated an automatic method for correcting intensity inhomogenieity artifacts in breast MR images. We developed a fuzzy c-means (FOM) based method for 3D lesion segmentation which is a key procedure in computerized interpretation of breast MR images including differential diagnosis and assessment of response to therapy. The computerized segmentation yielded 07% of 121 lesions having overlap larger than 0.4 with an experienced radiologist's manual outlining. We investigated computerized methods for automatic identification of characteristic kinetic curves (0KG) from the segmented 3D lesion and automatic extraction of kinetic features from the identified OKOs. The automatic method significantly improved the performance of kinetic features in the task of distinguishing between malignant and benign lesions. We also investigated a volumetric texture analysis method for classifying breast lesions on MRl as malignant and benign. The 3D method was found to have significantly better classification performance than the traditional 2D method. Finally, we assessed the relative importance of the various features-kinetic, morphological and texture-in differential diagnosis and combined multi-category features using a Bayesian neural network the performance of which was evaluated using two breast MR databases. Our research could potentially expedite the standardization of guidelines for interpretation of dynamic breast MRl of breast.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2006
- Accession Number
- ADA456375
Entities
People
- Weijie Chen
Organizations
- University of Chicago