3353 - Impact of Deep Learning Auto Segmentation Accuracy on Duodenal Dose-Volume Parameters for Pancreas SBRT Patients Treated with MR-Guided Online Adaptive Radiotherapy

Materials/

Methods: Nine pancreas SBRT patients treated on a 1.5T MR-Linac were included in this study. Two research DLAS models, DLAS_1 and DLAS_2, were used to segment the duodenum on daily fat-suppressed btFE images for each of the 9 patients. Manually drawn contours were used as ground truth (GT). Geometric metrics, Dice similarity coefficient (DSC) and mean distance to agreement (MDA), were used to determine contour accuracy. Dosimetric parameters, D_0.03cc, D_0.5cc, D_2.0cc, and D_5.0cc, were used to evaluate dose differences in the duodenum from DLAS_1 and DLAS_2 with the delivered plan. The Wilcoxon matched-pairs signed rank test was used to determine if significant dose differences existed between the DLAS and reference contours. A Spearman correlation analysis was used to determine the correlation between dose differences, ?D_0.03cc, ?D_0.5cc, ?D_2.0cc, and ?D_5.0cc, and geometric metrics.

Results: DLAS_2 demonstrated significantly better accuracy than DLAS_1 with average DSC and MDA of (0.73, 0.57, p=0.0001) and (4.23, 9.01, p=0.02), respectively. The Wilcoxon test revealed no significant dose differences between the dosimetric parameters obtained from the DLAS and reference contours (Table 1). A Spearman correlation between geometric metrics and ?D revealed a correlation between DSC and ?D_0.03cc (r=-0.73, p=0.03) and MDA and ?D_0.03cc (r=0.73, p=0.03) for DLAS_1 (Table 1).

Conclusion: Highly precise manual editing of duodenal contours may not be required for larger volume dose results used to evaluate pancreas SBRT plans during MRgOART. However, in regions of steep dose gradients, inaccurate contours may affect D_0.03cc results. Abstract 3353 – Table 1