IMPROVEMENT OF VISUAL SIMILARITY OF SIMILAR BREAST MASSES SELECTED BY COMPUTER-AIDED DIAGNOSIS SCHEMES

We developed a new method to improve visual similarity between a queried mass region and a set of reference regions selected by computer-aided diagnosis (CAD) schemes. For each queried region, CAD scheme first segmented the region, detected its boundary spiculation level, and computed 14 image features. The scheme then used a k-nearest neighbor algorithm to select a set of 25 most "similar" regions with the same computed spiculation level from a large reference library. The scheme computed the mutual information (MI) between the queried region and each of these 25 reference regions. The scheme finally selected and displayed six reference regions with the highest MI scores. In an observer preference study involving 85 queried regions, two sets of reference regions selected by this new scheme and the previously developed interactive method were randomly displayed with the queried region. Four observers participated in the study to select the more visually similar reference image set. On average for 54.1% of the queried regions, four observers preferred the automated selected reference region sets as being more visually similar to the queried region. The results suggested that both this automated and the interactive methods achieved the comparably visual similarity, which is significantly higher than the traditional CAD schemes