Computer-Aided Approach for BI-RADS Breast Density Classification: Multicentric Retrospective Study
| Authors | |
|---|---|
| Year of publication | 2024 |
| Type | Chapter of a book |
| MU Faculty or unit | |
| Citation | |
| Description | Assessing mammographic breast density, a crucial risk determinant for breast cancer, is typically conducted by radiologists through a visual examination of mammography images using the Breast Imaging and Reporting Data System (BI-RADS) breast density classification. However, significant interobserver variability among radiologists leads to inconsistency and potential inaccuracy in breast density assessments and consequent risk predictions. To address this, we analyzed 3835 Full-Field Digital Mammography (FFDM) studies from three mammographic centers. A team of 10 radiologists with experience in breast imaging ranging from 2 to 27 years evaluated these studies, establishing a ground truth for 2127 cases. We utilized 1122 (BI-RADS A: 356, BI-RADS B: 356, BI-RADS C: 356, BI-RADS D: 54) of the studies for training and 122 (BI-RADS A: 39, BI-RADS B: 39, BI-RADS C: 39, BI-RADS D: 5) for testing our Deep-Learning-based Automatic Detection (DLAD) algorithm. The proposed DLAD demonstrated an overall high accuracy (0.853), with balanced accuracy (BA) scores of 0.899 for BI-RADS Category A, 0.838 for Category B, 0.900 for Category C, and 0.900 for Category D. Our findings suggest that the proposed DLAD model can serve as a substantial support in the evaluation process, introducing an additional layer of analysis. |
| Related projects: |