Room: Exhibit Hall | Forum 2
Purpose: To evaluate different automated atlas-based segmentation engines for prostate radiotherapy.
Methods: Ten diverse prostate cases were selected to create an atlas using MD approved contours for two software systems, Eclipse Smart Segmentation (ESS) and MIM. Auto-segmentation was created for the organs at risk (OARs): bladder, rectum, left femoral head and neck, right femoral head and neck, prostate, seminal vesicles, and pelvic lymph nodes on five test subjects that were not included in the atlas group. Performance of auto-segmentation was assessed using DICE coefficient and volume ratios by comparing the contours from ESS and MIM with those MD approved contours. The DICE coefficient quantified the spatial overlap between the auto-segmented and MD contoured volumes while the volume ratio was calculated from the auto-segmented volume divided by the MD contoured volume.
Results: Average DICE coefficients of ESS were 0.94, 0.7, 0.66, 0.58, 0.77, 0.78, and 0.68 while the coefficients of MIM were 0.73, 0.76, 0.73, 0.32, 0.89, 0.9, and 0.67 for bladder, rectum, prostate, seminal vesicle, left femoral head/neck, right femoral head/neck and pelvic lymph nodes, respectively. The average volume ratios for the OARs were between 0.81 -1.41 and 0.78-2.05 for ESS and MIM, respectively. ESS performed better in bladder and seminal vesicle while MIM performed better in rectum, prostate and femoral heads/necks. Due to the inherent nature of MIM, several small volumes averaged together caused the resultant contours on test subjects to be larger in volumes compared to MD and ESS volumes.
Conclusion: Our study found that auto-segmentation could potentially reduce clinical workload in contouring. However, due to the characteristic of segmentation methods, special considerations may need to be taken into account in order to generate good segmentation results. Future work will increase the number of subjects in the atlas, and also compare segmentation performances in different body sites.