Mass Segmentation of Dense Breasts on Digitized Mammograms: Analysis of a Probability-Based Function

摘要

In this study, a segmentation algorithm based on the steepest changes of a probabilistic cost function was tested on non-processed and pre-processed dense breast images in an attempt to determine the efficacy of pre-processing for dense breast masses. Also, the inter-observer variability between expert radiologists was studied. Background trend correction was used as the pre-processing method. The algorithm, based on searching the steepest changes on a probabilistic cost function, was tested on 107 cancerous masses and 98 benign masses with density ratings of 3 or 4 according to the American College of Radiology’s density rating scale. The computer-segmented results were validated using the following statistics: overlap, accuracy, sensitivity, specificity, Dice similarity index, and kappa. The mean accuracy statistic value ranged from 0.71 to 0.84 for cancer cases and 0.81 to 0.86 for benign cases. For nearly all statistics there were statistically significant differences between the expert radiologists.