The benefit of automatic segmentation of intracranial organs at risk for radiation therapy: a multi-rater behavioral investigation

Abstract

Medical image segmentation has become a vital component of radiation therapy treatment planning over the past two decades. Segmentations are used both to target dose to tumors and assess dose to normal tissues, and they have been traditionally generated manually by human experts. In recent years there has been an abundance of scholarly activity in the development of computer algorithms to segment these tissues and quickly implement them into the clinical workflow. Comparatively little attention has been paid to the evaluation of such systems with respect to their potential clinical impact. Our group has developed atlas-based, image-registration-driven methods to segment the tissues of the brain, specifically the brainstem, optic chiasm, eyes, and optic nerves, critical to treatment planning. I have undertaken a multi-rater behavioral investigation to test the impact of segmentation differences with respect to the automatic system and experts both on a geometric level and with respect to radiation dosimetry, which is their end-use. I found with respect to both geometry and dosimetry that our automated methods produce segmentations that may be used as surrogates to those of the experts. These results also indicate that automated methods improve efficiency while reducing inter-rater variance and maintaining or improving accuracy, which have implications for both cost reduction and quality improvement of radiation therapy treatment planning.