The goal of the proposed research is to develop computer-aided diagnosis (CAD) schemes in order to improve the diagnostic accuracy of breast cancer in mammography.
Four specific aims are included: (l) development of computer programs for the detection and characterization of microcalcifications, (2) development of computer programs for detection and characterization of masses, (3) implementation of the CAD algorithms in a dedicated workstation to perform a pilot preclinical testing of the accuracy of the CAD programs, and (4) evaluation of the effects of the CAD schemes on radiologists' performance. The proposed CAD schemes will aid radiologists in screening mammograms for suspicious lesions and provide estimate of the likelihood of malignancy for the detected lesions. The information is expected to reduce the miss rate and to improve the positive predictive value of the mammographic findings. A data base of clinical mammograms which include malignant and benign microcalcifications and masses will be established. Physical measures which characterize the significant image features of the lesions will be developed. Based on these measures, linear discriminant classifiers or neural network classifiers will be optimized using a genetic algorithm approach to classify true and false signals and to estimate the likelihood of malignancy for each type of lesions. For automated detection and classification of microcalcifications, we will investigate the usefulness of multiresolution analysis for enhancement of the signal-to-noise ratio of the microcalcifications and for improvement of feature extraction techniques. Physical characteristics such as size, shape, frequency spectrum, spatial distribution, clustering properties, and texture features will be extracted and analyzed with the classifiers. For automated detection and classification of masses, we will improve the background correction and signal segmentation techniques, and develop effective false-positive reduction methods. Adaptive filtering, edge enhancement, and clustering segmentation methods will be developed for extraction of the mass margins. Physical characteristics such as size, density, edge sharpness, calcifications, shape, lobulation, spiculation, and multiresolution wavelet texture features will be extracted from the masses and analyzed with the classifiers. The algorithms will be implemented in a dedicated CAD workstation and preclinical testing will be conducted. The performance of the programs in a clinical setting will be assessed. The algorithms will be revised and improved based on the information obtained with the preclinical testing. The study is a vital step for the development of a clinically reliable CAD scheme. Observer performance studies using receiver operating characteristic (ROC) methodology will be conducted to evaluate the effects of the CAD schemes on radiologists'performance.
Showing the most recent 10 out of 36 publications
