(Monday, 7/15/2019) 3:45 PM - 4:15 PM
Room: Exhibit Hall | Forum 2
Purpose: Automated segmentation has the potential to address manual contouring challenges but, to date, performance of available solutions showed expert-level performance on only a small number of organs, thus requiring further editing before being considered clinically acceptable. During the last few years, deep learning-based algorithms have proven capable of delivering significantly better results , especially fully convolutional neural network (CNN) architectures. In this study we investigate the clinical performance of a custom-built fully CNN in delineating a wide range of important organ-at-risks (OARs)for thoracic cancer radiotherapy treatments.
Methods: A fully CNN has been trained using a multi-institutional, heterogeneous, dataset consisting of 550 CT scans of patients with different age, sex, histology and tumor location. The model has been tested against a set of independent CT images previously unseen to the model; the set originates from a research study [1] and it includes a cohort of 28 patients treated for lung disease. To demonstrate the generalizability of models, we also curated a test set of 25 open source CT scans available from The Cancer Imaging Archive (“TCIA test set�). To account for human variability, TCIA cases were manually segmented by experienced radiologists and reviewed by specialist oncologists.
Results: The deep learning model automatically detected, localized, and segmented most structures for all the 53 test patients. A generated segmentation is shown in Figure 1. The achieved surface distances on a subset of 20 cases are shown in Figure 2.
Conclusion: A DL model for thorax was trained, and its performance compared against two subgroups of test cases. Our initial results showed an overall good agreement between DL-originated structures and manually contoured volumes, with small differences for structures that are not clearly visible on CT scans, like for instance esophagus and great vessels.
Funding Support, Disclosures, and Conflict of Interest: All authors are employed by Varian Medical Systems and the research was performed in cooperation with it.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
