Breast density: Computerized analysis on digitized mammograms

Abstract

Mammographically dense breast tissue is related to a higher risk of breast cancer. We aim to evaluate a computerized system, assess whether it can provide an accurate and objective estimation of the breast density and if it can accurately classify the mammograms according to the ACR/BIRADS system. We retrospectively reviewed the mediolateral oblique (MLO) and cranial-caudal (CC) views of 83 normal mammograms and classified them, both manually and with the use of computerized texture analysis (CTA), according to their density. We grouped the mammograms either into two (ACR 1–2, ACR 3–4) or four categories (ACR 1 to 4). An inter-rater reliability analysis was performed using the kappa statistic to determine consistency among the radiologist and the CTA. The best matching was observed for the MLO view when the classification involved 2 groups (94%). The equivalent matching for the CC view was 92.8%. When we used all 4 ACR categories the matching was lower: i.e. 84.3% for the MLO view and 79.5% for the CC view. For older patients (>50 years old) the best matching was for the MLO views while for the younger patients equal matching was observed for both views. Overall, substantial to almost perfect agreement was observed between the two methods of assessment. CTA is a reliable and accurate form of computerized assisted diagnosis. If a single view is to be used, it should be the MLO view since the addition of CC view does not seem improve the sensitivity of the method.