Auto-contouring for Image-Guidance and Treatment Planning

Manual contouring is a time-consuming task routinely performed in radiotherapy to identify each patient’s targets and anatomical structures. Auto-segmentation of targets and normal tissues has been growing in clinical use as it can mitigate the inter- and intra-observer differences of manual segmentation and significantly reduce contouring time. Auto-segmentation has gone through advances over the years as computer technology has improved. The first-generation techniques are low-level with no prior information included, such as intensity thresholds. The second-generation techniques use uncertainty models to mitigate issues arising from noise, intensity nonuniformity, and partial volume effect. The third-generation techniques incorporate higher-level knowledge, such as atlas-based contouring. The fourth generation is the most recent development and focuses on deep learning-based techniques. There are many different deep learning techniques, with convolutional neural networks being the most commonly used technique for segmentation tasks. Before implementation in clinics, careful QA must be carried out for auto-segmentation tasks, such as comparison with clinically approved manual contours. In a clinical setting, commissioning and continuous quality assurance must be performed on auto-segmentation systems to ensure safe and proper use. This chapter covers these topics from the generations of auto-segmentation to QA as well as provides some of the latest results for auto-segmentation of normal tissues and targets/tumor volumes.