Could it be so that when there is errors in the arm position with the DICOM, you simply acquire new reference surface for the setup when you may see the entire arm. New reference surface should be taken with the accurate patient position- and posture after online corrections, when there was not any new errors in the images causing new systematic errors.
With the AlignRT, with the DICOM, in FB breasts, where the scan was acquired beginning from the mandibula, we have had surprisingly good reproducibility in the shoulder joint position. Systematic error is still larger than the random error indicating often the need for the new reference surface concerning (also) arm setup. In DIBH, where we always acquire new reference surface for the FB setup, results are also good nowadays. In both cases we setup the arm with the aid of treatment capture. In comparison to arm ROI, we have found that treatment capture is better in the results which demonstrates shoulder joint accuracy.
I would not acquire CT-scan for breast patients which goes through the brains, since there is in problematic arm positioning cases an other way to solve the problem, which is a new reference surface for the setup.