Click here to


Are you sure ?

Yes, do it No, cancel

Using Knowledge Based Planning to Determine Whether Geometric Variability in Segmentation Correlates with Dosimetric Variability

S Berry1*, C Senra2 , R Haq3 , H Veeraraghavan4 , (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) Hofstra Northwell School of Medicine, Hempstead, New York, (3) Memorial Sloan-Kettering Cancer Center, New York, NY, (4) Memorial Sloan Kettering Cancer Center, New York, NY


(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 5

Purpose: Segmentation of normal tissues in radiotherapy planning (RTP) is a means to the end of reducing and documenting doses to organs-at-risk (OARs). We evaluate whether the use of standard geometric segmentation quality metrics (SQM’s) is sufficient for judging segmentation quality for RTP.

Methods: 40 head and neck (H&N) patient RTP scans were retrospectively analyzed. Each scan contained 1 expert manual segmentation (EMS) and 1 atlas-based auto-segmentation (AS) for 8 H&N OARs. A commercial knowledge based planning (KBP) algorithm was used to estimate the mean and max doses to each segmentation version. The following SQM’s were calculated for each EMS and AS organ pairing: Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and 95th percentile of the HD (95%HD). The corresponding differences in KBP estimated mean (∆Davg) and maximum (∆Dmax) doses for each pairing were placed on scatter plots against each SQM and the Pearson correlation coefficients were calculated.

Results: Overall, there are no statistically significant strong (r > +/-0.6 ) correlations between geometric SQM’s and ∆Davg for any OARs. Moderate correlations with the 95%HD for the ipsilateral (r=0.428, p<0.001) and contralateral (r=0.551, p=0.001) parotid ∆Davg were observed, as was a moderate correlation between the DSC and the brainstem ∆Davg (r=-0.450, p=0.027) and ipsilateral submandibular gland ∆Davg (r=-0.422, p=0.006). No statistically significant moderate (r > +/-0.4 ) correlations between geometric SQM’s and ∆Dmax were observed for any OARs.

Conclusion: A lack of strong correlations indicates that evaluating OAR contours exclusively based on geometry will not necessarily inform the clinician about the clinical repercussions of a segmentation discrepancy. Therefore, it is important to also directly consider the dosimetric consequences of segmentation variability when evaluating contours for RTP.

Funding Support, Disclosures, and Conflict of Interest: This research was partially supported by Varian Medical Systems


Segmentation, Radiation Therapy


IM/TH- image segmentation: CT

Contact Email