Improving Clinical Diagnosis Through Change Detection in Mammography
Abstract
Temporal change of mass lesions overtime is a key piece of information in computer-aided diagnosis of breast cancer and treatment monitoring. For a specific patient, change detection is a critical step to select lesion candidates for follow-up diagnosis performed by either clinicians or computers. In the third year of this project, we developed a hybrid image registration technique to align temporal sequences of the same patient and detect changing lesions, and developed a neural network based classifier to derive the probabilities of true masses. In particular, we developed: (1) mPAR and MLP- based registration algorithm to recover non-rigid deformation; (2) a new change detection scheme using independent component analysis of image sequences; (3) a feature extraction algorithm to obtain discriminative imagery features of true masses against mass-like normal tissues; and (4) a neural network based decision support system for mass detection. Our preliminary studies have shown a very good performance of the mass detection system consisting of 91 mammograms. The performance was initially 0.78-0.80 for the areas A(sub z) under the ROC curves using the conventional neural network, and later being improved to A(sub z) values of 0.84-0.89 when using the newly developed multiple circular path neural networks.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2001
- Accession Number
- ADA398681
Entities
People
- Yue J. Wang
Organizations
- The Catholic University of America