Computer-aided diagnosis of breast lesions using a multifeature analysis procedure

We have developed a family of objective features in order to provide non-invasive, reliable means of distinguishing benign from malignant breast lesions. These include acoustic features (echogenicity, heterogeneity, shadowing) and morphometric features (area, aspect ratio, border irregularity, margin definition). These quantitative descriptors are designed to be independent of instrument properties and physician expertise. Our analysis included manual tracing of lesion boundaries and adjacent areas on grayscale images generated from RF data. To derive quantitative acoustic features, we computed spectral parameter maps of radio-frequency (RF) echo signals (calibrated with system transfer function and corrected for diffraction) within these areas. We quantified morphometric features by geometric and fractal analysis of traced lesion boundaries. Although no single parameter can reliably discriminate cancerous from non-cancerous breast lesions, multifeature analysis provides excellent discrimination of cancerous and non-cancerous lesions. RF echo-signal data used in this study were acquired during routine ultrasonic examinations of biopsy-scheduled patients at three clinical sites. Our data analysis for 130 patients produced an ROC-curve area of 0.9164 +/- 0.0346. Among the quantitative descriptors, lesion heterogeneity, aspect ratio, and a border irregularity descriptor were the most useful; some morphometric features (such as the border irregularity descriptor) were particularly effective in lesion classification.