ABSTRACT
Auto-segmentation research in radiation oncology has been greatly advanced due to the availability of publicly available datasets curated to develop auto-segmentation algorithms. While these datasets are often carefully curated to ensure the high-quality of the segmentations provided, they are often sparse in patient numbers, limiting the range of patient anatomical variability that is captured while developing an algorithm. Similarly, as these datasets are curated, they typically only consider a single anatomical site (i.e. thorax, abdomen), limiting the ability to assess the model’s performance when imaging from other anatomical sites is used on a specific site model. In this chapter, the 2017 AAPM Thoracic Auto-segmentation Challenge data is used as a case study to train a commonly used deep learning method and evaluate its performance on a variety of sites (i.e. non-thorax) and clinical presentations (i.e. atelectasis).
