Multimodality Computerized Diagnosis of Breast Lesions Using Mammography and Sonography1

摘要: Rationale and Objectives The purpose of this study is to investigate the use computer-extracted features lesions imaged by means two modalities, mammography breast ultrasound, in computerized classification lesions. Material Methods We performed analysis on a database 97 patients with total 100 (40 malignant, 40 benign solid, 20 cystic lesions). Mammograms ultrasound images were available for these There was an average three mammographic per lesion. Based seed points indicated radiologist, computer automatically segmented from parenchymal background extracted set characteristic each For feature, its value averaged over all pertaining given lesion input Bayesian neural network classification. also investigated different approaches combine image-based into by-lesion analysis. In that analysis, mean, maximum, minimum feature values considered representing performance using leave-one-lesion-out approach, based image alone (two five features), (three four combination both modalities total). Results task distinguishing cancer other abnormalities lesion-based single modality, areas under receiver operating curves (A z values) increased significantly when selected manner (mean, minimum, or maximum) which combined features. highest found automated selection (resulting being used). That A 0.92, showing statistically significant increase achieved either alone. Conclusion Computerized improved combined. Classification depended specific methods combining multiple These results are encouraging warrant further exploration multimodality imaging.