ABSTRACT
With the emergence of deep learning as the most accurate machine learning method, its applications in radiation oncology have exploded. Most notable successes have been demonstrated in the segmentation of various organs from medical images. Deep learning requires large, well-curated datasets for extracting very high dimensional model representations for producing robust inference to generate such segmentations from the underlying imaging data. Creation of such datasets requires labor-intensive and manual effort by domain experts, which creates the data curation problem. A second problem occurs in the clinical commissioning stage of the algorithm because of lack of consistent guidelines for judging when a method is performing at an appropriate level for routine clinical use. In this chapter, we expand on both of these issues, and present some guidelines for clinical commissioning of auto-segmentation methods.
