ESTRO 2024 - Abstract Book

S3108

Physics - Autosegmentation

ESTRO 2024

contours (GSC) in the test cohort were reviewed and adjusted by a board-certified radiation oncologist, and geometrically compared with the AI-generated contours (AIC). For quantitative evaluation, dice similarity coefficients (DSC) and 95% Hausdorff distances (HD95) were calculated in MATLAB. The model performances were compared using Bonferroni-corrected Wilcoxon signed-rank test.

Results:

AIC were generated in a median [range] time of 197 [163–221] s, 862 [859–865] s, and 307 [245-398] using CT, CTMR or MRI, respectively. CT-based AIC showed a maximum median DSC of 0.95 for liver and minimum median DSC of 0.87 was found for spinal cord. In the case of the MRCT- and MRI-based AIC, corresponding maximum and minimum median DSC values with 0.95 for the liver and 0.81 for the stomach were found (cf. Fig. 1). For HD95, the maximal and minimal median HD95 of all models were detected for spinal cord with 2.6 mm (CT), 2.6 mm (CTMR), and 2.4 mm (MRI) and for stomach with 7.9 mm (CT), 9.3 mm (CTMR), and 8.4 mm (MRI), respectively.

While good performance was observed for AIC based on CT, GTV delineation quality was poor (DSC=0, HD95=30.8 mm). For the MRCT model, GTV DSCs ranged from 0.10 to 0.93 with a median value of 0.73. The median [range] HD95 was 3.4 [1.3-86.8] mm. MR-only auto-segmentation showed a median [range] DSC of 0.89 [0.32-0.95]. The median [range] HD95 was 1.9 [1.0-111.2] mm. Fig. 2 qualitatively visualizes the best scored patient for each model. Over the entire test cohort, the MRI-only model showed the globally highest performance.