2243 - Evaluation of an Externally Trained Deep-Learning Based GTV Auto-Segmentation Model for Head and Neck Cancer

J. C. M. Rwigema¹, A. Bilaal², R. C. Tegtmeier³, J. Holmes⁴, J. Qian⁵, L. A. McGee¹, S. H. Patel¹, Y. Rong⁶, and Q. Chen⁶; ¹Department of Radiation Oncology, Mayo Clinic, Phoenix, AZ, ²Mayo Clinic Arizona, Phoenix, AZ, ³Department of Medical Physics, University of Wisconsin - Madison, Madison, WI, ⁴Department of Radiation Oncology, Mayo Clinic Arizona, Phoenix, AZ, ⁵Department of Radiation Oncology, Mayo Clinic, Rochester, MN, ⁶Department of Radiation Oncology, Mayo Clinic AZ, Phoenix, AZ

Purpose/Objective(s): While Deep-learning based approach has started to be adopted in clinical practice for organs-at-risk, there has been limited success in auto-segmentation(AS) of target structures. Our goal is to evaluate a Head & Neck (HN) target AS model trained on public dataset on the PET-CT scans from our local institution. Materials/

Methods: The model was trained and validated on the ‘HECKTOR’ Challenge dataset. Fifty HN cases with PET-CT scans registered with planning CT were collected. Both primary (GTV_p) and nodes (GTV_n) segmentation were produced based on registered PET and planning CT. Quantitative metrics including Dice similarity Coefficient (DSC) and 95-percentile Hausdorff distance (HD95) were computed against the clinical contours. Subjective analyses were also performed to evaluate the clinical acceptability of the AS target contours.
Results: For GTV_p, AS contour achieves a DSC of 0.72±0.17 and HD95 of 8.3±11mm on our data, poorer than that achieved on Challenge (0.80±0.14 and 5.7±3.4mm). A case-by-case review identified 13 obvious outliers where the software either picked or included one of the high standardized uptake value (SUV) nodes as the primary, or skipped the region in primary tumor that has low SUV. When we exclude these outliers, the remaining 37 cases achieves a DSC of 0.80±0.06 and HD95 of 4.0±1.4mm, similar to performance observed on Challenge. For PTV_n, DSC of 0.66±0.17 and HD95 of 16.0±16.0mm was obtained by the AS. Both false-positive and false-negative were seen in the PTV_n AS.
Conclusion: The HN target AS model showed promise on primary target segmentation. However, further improvements are needed to improve the generalizability, especially for the positive node identification and segmentation.