Automatic Segmentation of Low-Contrast Head and Neck Tumors for Palliative Radiotherapy
S Gay1*, C Cardenas1, C Nguyen1, T Netherton1, A Aggarwal2, K Naidoo3, H Simonds3, B Beadle4, L Court1, (1) The University of Texas MD Anderson Cancer Center, Houston, TX, (2) Guys Cancer Centre, London, UK, (3) Stellenbosch University, Stellenbosch, ZA, (4) Stanford University, Stanford, CA
Presentations
TU-IePD-TRACK 4-2 (Tuesday, 7/27/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: To automatically contour gross tumor volumes (GTVs) for palliative head-and-neck (HN) radiotherapy at multiple treatment sites from single-modality, low-contrast scans.
Methods: 110 palliative and 234 curative treatment plans (CT scans and physician-approved GTV contours) for nine HN treatment sites were retrospectively obtained. Mean target volume was 88.9 (σ 123.3) cc, and both palliative and curative cases were evenly split 80%/20% into training-validation and test sets, respectively. First, state-of-the-art CNNs (e.g. U-Net) were evaluated. Then, a novel, coarse-to-fine cascading 3D Attention U-Net was developed to segment the GTVs. Model performance was evaluated by comparing Dice similarity coefficient (DSC) and mean surface distance (MSD) to measure target coverage and structure similarity of predicted and clinical GTVs.
Results: Over the entire test dataset (curative & palliative cases), the best scores achieved with the conventional U-Net architecture were median DSC and median MSD of 0.65 (σ 0.22) and 2.08 (σ 7.65) mm, respectively. For the cascading 3D Attention U-Net, these scores changed to 0.67 (σ 0.23) and 2.26 (σ 4.27) mm; thus, this approach improved both median DSC and the extent of MSD outliers. Segmentation of large, high-visibility tumors performed better than smaller or low-visibility tumors. Across treatment sites, base of tongue, mandible, and thyroid median DSC scores were 0.70 or greater, with mandible achieving the highest median DSC (0.74). Additionally, median DSC scores for glossopharyngeal sulcus, head and neck, thyroid, tonsil, and other were 0.60 or greater.
Conclusion: We developed a novel 3D segmentation network to find GTV delineations for palliative HN treatments. To our knowledge, this is the first work developed for multiple HN treatment sites.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by a grant from the Wellcome Trust.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email