Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Deep-Learning Auto-Segmentation Vs. Manually Segmented Delineations for Head and Neck Cancer Patients

T Bejarano*, V Leandro Alves, K Ward, D Cousins, E Nesbit, E Asare, B Walker, J Siebers, C Luminais, University of Virginia Health System, Charlottesville, VA


PO-GePV-M-6 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To assess the performance of a commercial CT-based deep learning auto-segmentation method for head-and-neck (HNC) OARs compared to manually segmented delineations. .

Methods: One hundred twenty three HNC DICOM datasets from The Cancer Imaging Archive (TCIA), annotated as NRRD format by the University of California, were obtained from the Public Domain Database for Computational Anatomy (PDDCA). TCIA manually delineations (MD) were intercompared with AI-based INTContour (Carina Medical LLC, Lexington, KY) delineations using the dice-similarity-coefficient (DSC), Hausdorff-distance (HD) and 95th percentile HD (HD95). Additionally, a blinded delineation rating study was completed for 5 randomly selected patients to determine delineation method preference for the 13/25 delineations with the poorest mean DSC. Eight experienced delineators (dosimetrists and physicians) were asked to select delineation “A” (e.g. manual), “B” (e.g. AI), “neither A or B,” or “either A or B” that best represented the specific OAR.

Results: The AI delineations demonstrated good geometric accuracy on OARs, with min/mean±sd DSC for Bone_Mandible, Lobe_Temporal_L, Lobe_Temporal_R, Brainstem, Parotid_L, Parotid_R, Eye_R, and Eye_L of 0.55/0.78±0.05, 0.28/0.75±0.07, 0.65/0.77±0.04, 0.75/0.84±0.03, 0.66/0.83±0.04, 0.61/0.82±0.05, 0.72/0.83±0.03 and 0.72/0.84±0.03, respectively. HD95 for same delineations were 0.81±0.44, 1.42±0.43, 1.34±0.33, 0.38±0.12, 0.70±0.57, 0.82±0.63, 0.30±0.06, 0.30±0.08, respectively. The delineation rating study found that neither MD nor AI delineations were unanimously preferred by all delineators for any OAR; the mixed preference on all OARs indicates subjectivity in the preference selection. Overall, the MD were preferred for Glnd_Thyroid(71.4%), Optic_Chiasm(55%), OpticNrv_L(42.5%), and OpticNrv_R(45%). AI delineations were preferred for InnerEar_L(60%), InnerEar_R(60%), Joint_TM_L(62.9%), Joint_TM_R(62.9%), Lens_R(35%) and Spinalcord(75%). Neither structure was preferred for Lens_L and the Pituitary.

Conclusion: AI delineations display good geometric accuracy for the HNC OARs. For all OARs, preference is mixed between MD and AI, indicating clinical acceptability for either delineation. Delineation rating study indicates AI is preferred by delineation experts for delineations with high contrast such as bone.

Funding Support, Disclosures, and Conflict of Interest: Grant project # 5R01CA222216-03


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email