ESTRO 2024 - Abstract Book

S3149

Physics - Autosegmentation

ESTRO 2024

To verify the correlation between qualitative and quantitative scores we computed Kendall’s Tau correlation coefficient 7 between the grades given by ROs and 3 quantitative metrics: vDSC, sDSC and HD95 8 . sDSC demonstrated the highest correlation (0.38), vDSC the lowest (0.18) and HD95 (0.27) had a fair correlation. Only the cases with manual contours graded with 1 were used (gold standard). Despite having the lowest correlation, vDSC is the most computed metric in the literature and finding what vDSC scores correspond to grade 1 given by ROs is still of great interest. To further investigate the correlation between the grades given by the ROs and the vDCS, we separated the scans in 2 groups. The first group consists of all the AI contours graded as acceptable (grade 1) while the second group consists of all the remaining scans. Then we computed the median vDCS (between the AI contours and the gold standard) for each OAR in each of the 2 groups. For 9 OARs the median of the first group was higher than the second group by at least 0.02 vDCS. For 13 OARs the difference was below 0.02 Dice and for only 2 OARs the second group had higher median scores by 0.02 (Cochlea_R and OpticNerv_R). From these results, we can presume that the vDCS has a positive correlation with grades given by the doctors for most OARs.

By assuming a correlation between median vDSC and given grades of 1, we observe that a vDCS higher than 0.72 is clinically accepted for the next steps in radiotherapy.

Conclusion:

The study demonstrates AI's effectiveness in outlining OARs for radiotherapy planning, outperforming manual methods in 14 of 25 cases and equalling performance for the Brain and Eye_R. With median vDSC scores above 0.80 for most structures, and notable results for the Brain, Bone_Mandible, and Eye_L/R, AI proves to be accurate and reliable. Despite difficulties in visualizing the OpticChiasm, AI's performance is impressive compared to existing