Atlas selection methods for multi-atlas-based segmentation in breast cancer radiotherapy planning

We evaluated atlas selection methods for multi-atlas-based segmentation (MABS) in breast cancer radiotherapy planning. Forty-five patients were divided into 30 atlas and 15 test cases. The 30 atlases were stratified into three groups based on breast separation, height, and volume. Firstly, MABS was performed on each of the 30 atlas cases using the remaining 29 atlases. Secondly, MABS was performed on 15 test cases using the 30 atlases. The Dice similarity coefficient (DSC) was calculated to assess the agreement between MABS and manual segmentation. The DSC was found to increase as more atlases were selected. Although this led to an increase in the computational time, the implementation of patient stratification reduced the computational time compared with using the entire dataset. Atlas selection from the height-matched and volume-matched tertile datasets provided median DSC values > 0.9. Breast height may be a practical surrogate for breast volume which is unknown before segmentation.