Feature and knowledge based analysis for reduction of false positives in the computerized detection of masses in screening mammography

Previously we presented a morphologic concentric layered (MCL) algorithm for the detection of masses in screening mammograms. The algorithm achieved high sensitivity (92%) but it also generated 3.26 false positives (FPs) per image. In the present study we propose a false positive reduction strategy based on using an artificial neural network that merges feature and knowledge-based analysis of suspicious mammographic locations. The ANN integrates two types of information regarding the suspicious candidates: (i) directional and fractal neighborhood analysis features, and (ii) knowledge-based analysis using an information-theoretic similarity metric. The study hypothesis is that the synergistic application of feature and knowledge-based analysis will be an effective strategy to reduce false positives while still maintaining sufficiently the detection rate for true masses. The study was performed using mammograms from the Digital Database of Screening Mammography. Using the fusion ANN decision strategy 56% of the FPs were reduced while maintaining 95% of the true masses.