Investigating the Classification Robustness of a Computer Aided Diagnosis System Using Mammographic Images

Objectives: To test a microcalcification classification algorithm and determine whether it correctly classifies detected microcalcifications in a selected region of a mammogram. Methods: The proposed algorithm was tested in three regions per mammogram (one -biopsy proven-abnormal and two random ones) indicated from collaborating doctors in a sample of 50 mammograms collected during the systemʹs clinical trial in Athens Hippocratioʹs Hospital Breast Unit. Every microcalcification is tested for seven characteristics and is assigned a value for each one of them in order to calculate the final risk and to assign the microcalcification in one of the five designated risk categories. Furthermore, all microcalcifications belonging to a specific risk category are coloured using a colour-based classification making it easier for the doctor to identify microcalcifications from different risk categories. Results: 150 regions from a total number of 50 mammograms accompanied from biopsy test results have been tested. The average number of microcalcifications per cm2 detected in random areas was 12.55 while for the biopsy proven abnormal areas this number was 21.17. The average number of high risk microcalcifications per cm2 was 1.32 for the random regions tested, while the same number for the biopsy proven abnormal areas was 6.96. Conclusion: The results show that the algorithm classifies correctly the microcalcifications of suspiciousʹ areas clusters while the number of high risk microcalcifications is substantially higher in abnormal regions than in random ones.