Background: Recent studies have indicated that not just density but also texture in mammograms is associated with risk. Since ninety percent of breast cancers arise from the ductal and lobular glands also calcification appears as a parallel lines assocoiated with blood vessel hence we choose to investigate texture features describing the local orientation of breast tissue in an anatomically oriented breast coordinate system that is more accurate than traditional image coordinate system. We further discover the region on mammogram where changes due to breast cancer are most like to occur and are not visible by a naked eye to a radiologist in a case-control study design. Evaluation: We use a machine learning system consisting of (1) the Gaussian derivative features on four scales extracted with respect to (2) an anatomically oriented breast coordinate system; (3) Sequential Forward Selection of features, and (4) committee of kNN classifiers. The use of the anatomical breast coordinate system defines matching locations across mammograms and makes it possible to investigate what are the mammogram regions that show significantly different classification scores between the cancer and control group. We use independent training and testing sets from the case-control study of 495 women including 245 diagnosed with breast cancer 2-4 years after the baseline mammograms, used in this work. Discussion: Our experiments suggest that significant regions for breast cancer risk assessment are the glandular tissue region behind the nipple as well as the lobule close to the pectorial muscle in the breast, which both can be anatomically justified. Moreover, this work suggests which regions should be more carefully considered to improve both preclinical and follow-up screening phases of cancer progression. Conclusion: The study demonstrated the benefits of including breast coordinate over traditional coordinate system considering the anatomy of the breast in mammogram. It also suggests which regions should be more carefully considered to improve both preclinical and follow-up screening phases of cancer progression. However, larger study population is warrented to validate this hypothesis.
Discovery of Mammogram Regions That Show Early Changes Due to the Development of Breast Cancer: A Preliminary Work
G. Karemore, S. Brandt, N. Karssemeijer and M. Nielsen
Annual Meeting of the Radiological Society of North America 2011.