Comparison of Three Commercial Model Based Segmentation Software and Atlas Based Segmentation Software in Contouring of Prostate Cancer and Brain Cancer

Abstract

Delineating contours manually is tedious and time consuming. Recently there are many commercial auto contouring software as increasing the importance of automatic segmentation approaches to reduce the contouring load. Therefore, we would select software to reduce the actual working time by comparing manual contouring with auto contouring. Materials/Methods: It is multi-center study. 20 patients, divided into two groups (each group has ten patients) with prostate cancer and brain cancer were enrolled in this study. We used three model based segmentation (MBS) software solutions (Pinnacle 9.8 spice, Raystation 4.5, and Multiplan 5.1 by Accuray) and one atlas based segmentation (ABS) software solution (MIM 6.4. by MIMVista) to generate the automatic contouring on the planning CT. All MBS software just make contours according to default option without any preparation. We conducted with three method e Pinnacle (P), Raystation(R) and Multiplan(M) for MBS. However, ABS software should have subjects (who are already registered for ABS to work auto contouring and also they are not the patients involved in this study). Therefore, we made two groups of atlas, 10 subjects of atlas and 60 subjects of atlas. We used two matching techniques, Single-best matched atlas-based segmentation method (SBM) and Multi-atlas technique in ABS. We analyzed auto contouring with 4 classified group for ABS e SBM for 10 subjects (A1), Multi-atlas for 10 subjects (B1), SBM for 60 subjects (A6), and Multi-atlas for 60 subjects (B6). Especially, Multiplan required T1 MR images to delineate contours for brain cancer. In addition, we obtained brain, brain stem, both eye ball contours for brain case and bladder, rectum, both femoral head contours for prostate case. Average Dice Similarity Coefficients (DSC) was calculated for each structure to compare against manually defined "gold" standard contours of 7 groups respectively. Values closer to 1 indicate higher accuracy. Overall percent improvement was calculated as the proportion of the error corrected by the method, or % difference on 1-DSC. Results: B6 was significantly more accurate than other group (p < 0.03) with average DSC of 0.904 AE 0.057 compared to 0.851 AE 0.106, 0.878 AE 0.082, 0.871 AE 0.081, 0.770 AE 0.13, 0.822 AE 0.118 and 0.823 AE 0.119 respectively for A1, B1, A6, M, P and Individual contours were improved by average 15% in B6 (p < 0.05) for the both femoral head, rectum. Overall, B6 showed the greatest improvement over all groups. Conclusion: Most of auto contoured contours should be modified. Autocontouring reduces contouring time than contouring manually. Among 7 groups, DSC of B6 which has the most subjects is the highest. ABS software takes more time and effort to use in the first place. However, ABS would have much better auto contouring accuracy compared with MBS software with continuous addition of subjects' data for each organ on atlas database.