Investigating performance of a morphology-based CAD scheme in detecting architectural distortion in screening mammograms

This study evaluates the performance of computer-aided detection (CAD) scheme for architectural distortion (AD) signs in screening mammograms The scheme targets the morphological effects of breast parenchyma due to cancer infiltration. The underlying hypothesis is that cancer growth evolves in concentric layers around a focal region. Overall, the CAD scheme operates by providing a topographic map of the breast region to identify candidate masses for further analysis. Initially, the centroid of suspected areas is identified based on five main morphological characteristics. By monitoring the topographic representation of the breast parenchyma from these seed locations for areas of increased concentric activity, fading around the focal layer, we identify seeds with more suspicious behavior. Subsequently, relative incidences of those suspicious seeds are inspected to keep only those with intuitive rare incidence range. Previously, the CAD scheme was developed and optimized for the detection of malignant masses. In the present study, key CAD parameters were further optimized using mammography cases that include architectural distortion. The study was performed using 78 AD cases from the Digital Database of Screening Mammography (DDSM). These cases contain 80 biopsy-proven malignant masses; 13 masses, 38 masses with associated AD, and 29 abnormalities where parenchymal distortion was the only sign of malignancy. The study showed that our CAD scheme detects AD with 82.3%/91.3% sensitivity rate at false positive per image (FPI) rate of 7.3/9.1. Detection performance was robust among masses, masses with associated AD, and pure AD abnormalities.