Knowledge-based detection of mammographic masses: Analysis of the impact of database comprehensiveness

Previously, we presented a knowledge-based computer-assisted detection (KB-CAD) algorithm for mammographic masses. Given a query case, the algorithm compares it with templates stored in a databank with known ground truth. A decision index is calculated to measure how well the query matches the stored mass templates relative to templates depicting normal parenchyma. The underlying hypothesis is that if the query region contains a mass, it should match the mass templates better than the normal templates, thus resulting in a higher decision index. Such CAD system takes advantage of growing image libraries without further retraining. However, performance strongly depends on the comprehensiveness of the knowledge database. The purpose of this study was to assess the impact of database comprehensiveness by measuring the changes in overall detection performance as the size of the knowledge databank progressively increased. The study was based on the Digital Database of Screening Mammography. Starting with a database of 681 mass and 656 normal templates, we observed that the KB-CAD performance improved when increasing the size of the database. The system achieved the best overall performance when the knowledge databank contained 600 templates. Breast density and mass shape did not affect the detection performance. However, the margin characteristics of the query mass did. Fewer (100) mass templates were required for accurate detection of circumscribed masses while more (400) mass templates were required for detection of ill-defined query masses. In conclusion, it is important to balance comprehensiveness, performance, and maintenance requirements when developing knowledge databases for CAD.