Purpose: Breast cancer is a common but treatable disease for adult women. Improvements in breast cancer detection and treatment have helped to lower mortality, but there is still a need for further improvements, particularly for better computer-aided diagnosis (CADx) and computer-aided detection (CADe).
Methods: Two new CAD prototypes, one CADx and one CADe prototype, were evaluated. The core modules are segmentation of lesions, feature extraction, and classification. The evaluation of microcalcifications and mass lesions is based on the currently largest publicly available Digital Database for Screening Mammography (DDSM) with digitized film mammograms and a smaller data source with high-quality mammograms from digital mammography devices. Two different image analysis approaches used by the respective CAD prototypes were examined and compared. These include the 'machine learning' approach and the new 'knowledge-driven' approach. Particular emphasis is put on a profound discussion of statistical methods with recommendations for their proper application in order to avoid common errors including feature selection, model fitting, and sampling schemes.
Results: The results show that the classification performance of the investigated CADx prototypes for microcalcifications produced a higher AUC =.777 for 44 machine learning features than for 10 knowledge-driven features (AUC =.657). A combination of both feature sets (53 features) did not substantially raise the classification performance (AUC =.771). These analyses were based on 1,347 and 1,359 ROIs, respectively. Evaluating the CADx prototype with 242 machine learning features on DDSM masses data resulted in an AUC of .862 using 1,934 ROIs. Furthermore, a CADe prototype was applied to three own databases giving information about the true positive detection rate for mass lesions. Depending on the definition of a true positive detection, it produced AUC values of .953, .818, and .954 using 12, 17, and 18 features, respectively.
Conclusion: The comparison of CAD prototypes revealed that the quality of results is highly dependent on the correct usage of statistical models, feature selection methods, and evaluation schemes.