Classification of Masses

   Malignant and benign masses are classified using morphological and texture features [12]. The ROI containing the mass is processed with a background-correction technique to reduce the structured background variation. A migrating-means clustering algorithm is then employed to segment the ROI into a mass object and breast tissue. A rubber-band straightening transform (RBST) technique is used for mapping a band of pixels that contains the mass margin into the Cartesian plane. The resulting RBST image carries useful directional information that can be extracted more easily than from the original image. Texture features are extracted from the RBST image. A second-stage segmentation extracts the detailed margin of the mass. Morphological features are extracted from the segmented mass. Linear discriminant analysis or neural network then classifies the malignant and benign masses.

Fig. 12. ^ Schematic diagram for computerized classification of masses.

Original       First-Stage      Second-Stage    ROI         Segmentation   Segmentation Malignant mass: Intraductal and Infiltrative Ductal Carcinoma Benign Cluster: Fibrosis Fig. 13. ^ Examples of segmentation for a malignant and a benign mass. Fig. 14. < Comparison of the average ROC curves for six MQSA-approved radiologists in classification of masses. The improvement in Az was statistically significant (p=0.007) [13].