PP09 Presentation Time: 10:50 AM

Purpose

To quantify the 2 sources of human interobserver variability (IoV) introduced during prostate MRI-assisted radiosurgery postimplant dosimetry, and compare this human variability to automatically generated postimplant dosimetry.

Materials and Methods

Twenty-five patients underwent LDR prostate brachytherapy (LDRPBT) and were imaged with fully balanced SSFP for postimplant dosimetry. Four certified medical dosimetrists (CMDs) identified the radioactive seeds in each of the 25 postimplant MRIs. Seven additional clinical observers, including 3 radiation oncologists (ROs), contoured the prostate and 4 organs at risk (OARs) on the MRIs. Consensus segmentation masks from the 3 ROs' contours were computed for each patient using the simultaneous truth and performance level estimation (STAPLE) algorithm. Thirty unique postimplant dosimetry plans were computed for each patient. Twenty-eight of the dosimetry plans were computed from each unique pair of the CMDs' seed plans and observer contours (4 CMDs × 7 contourers). One dosimetry plan was performed automatically for each patient using custom autosegmentation and automatic seed identification software developed previously. A reference dosimetry plan for each patient was computed using the consensus STAPLE segmentations and the seeds identified by the most experienced CMD. IoV of 5 dosimetry parameters was quantified separately for each IoV source: IoV introduced by seed localization, and IoV introduced by anatomy contouring. The coefficient of variation (CoV) of each dosimetry parameter was quantified for each patient across CMDs using the consensus STAPLE segmentations. Similarly, the CoV of each dosimetry parameter was quantified for each patient across contouring observers using the seed locations determined by the most experienced CMD. Additionally, the differences in dosimetry parameters between the reference and automatic dosimetry plans were quantified.

Results/Discussion

CoVs due to variability of anatomy contouring (CoV_contours) were significantly higher than those due to variability of seed localization (CoV_seeds) for all dosimetry parameters (median CoV_contours vs. median CoV_seeds: prostate D90 - 15.116% vs. 0.651%, p<0.001; prostate V100 - 5.359% vs. 0.365%, p<0.001; prostate V150 - 7.017% vs. 1.252%, p<0.001; rectum V100 - 79.230% vs. 8.688%, p<0.001; EUS V200 - 107.740% vs. 21.176%, p<0.001). CoV_contours were lower when the contouring observers were restricted to only the 3 ROs, but they were still significantly higher than CoV_seeds for all dosimetry parameters. Median differences in prostate D90, prostate V100, prostate V150, rectum V100, and EUS V200 between automatically computed and reference dosimetry parameters were 3.156%, 1.630%, -1.071%, -0.001%, and -0.002%, respectively. Most automatically computed dosimetry parameters were within the expected variability of those produced by human observers (Figure 1).

Conclusion

IoV of contouring constitutes the largest source of IoV in postimplant dosimetry plans performed by clinical observers. When using automatic contouring and seed localization tools, review and refinement of computer predictions will be necessary for some patients. However, automatic dosimetry tools can estimate postimplant dosimetry that is close to, or within, the expected variability of dosimetry plans produced by clinical observers.