ESTRO 2024 - Abstract Book

S3142

Physics - Autosegmentation

ESTRO 2024

The AI-generated PBS underwent a visual assessment by an experienced oncologist, with any deviations from corresponding manually delineated contours being rated as no adjustments needed, minor adjustments needed, only adjustments of the cranio-caudal extension needed or major adjustments needed Furthermore, we compared the AI-generated PBS and the manual delineations using various metrics. Here Surface Dice 1mm (SDSC1) is reported as the median [range], and mean Hausdorff Distance (mm) (HD) are reported as median[range]. In the trial, two plans were optimized for all patients: one standard plan (SD) optimized without inclusion of the bone substructures and one (BS) including the substructures in the optimization process. In order to quantify the dosimetric implication of using AI generated structures in the clinic DVH’s of the uncorrected AI generated structures were extracted and compared to DVH’s of clinically used structures for the 10 patients. That was done for both the SD and the BS plan. Mean doses for each bone structure were plotted as scatter plots (Manual versus AI) and compared using Pearson correlation coefficient.

Results:

Median SDSC1 and median HD for all structures are shown in table 1. SDSC1 was lowest for the SI joints L/R: 0.86 [0.72-0.92] and 0.89 [0.72-0.93], and the median HD was also highest for SI joints L/R 0.89 and 0.82 mm. For all other structures median SDSC1 was above 0.92. The less favourable AI performance for the SI joints is due to the definition of the joint, as a volume comprising 1 cm to each side of the joint, that has no borders visual on CT.