Comparing Images Using the Hausdorff Distance
IEEE Transactions on Pattern Analysis and Machine Intelligence
| Hi-index | 0.00 |
Background: Endorectal MRI provides detailed images of the prostate anatomy and is useful for radiation treatment planning. Here we describe a Demons field-initialized B-spline deformable registration of prostate MRI. Material and methods: T2-weighted endorectal MRIs of five patients were used. The prostate and the tumor of each patient were manually contoured. The planning MRIs and their segmentations were simulated by warping the corresponding endorectal MRIs using thin plate spline (TPS). Deformable registration was initialized using the deformation field generated using Demons algorithm to map the deformed prostate MRI to the non-deformed one. The solution was refined with B-Spline registration. Volume overlap similarity was used to assess the accuracy of registration and to suggest a minimum margin to account for the registration errors. Results: Initialization using Demons algorithm took about 15min on a computer with 2.8GHz Intel, 1.3GB RAM. Refinement B-spline registration (200 iterations) took less than 5min. Using the synthetic images as the ground truth, at zero margin, the average (S.D.) 98 (+/-0.4)% for prostate coverage was 97 (+/-1)% for tumor. The average (+/-S.D.) treatment margin required to cover the entire prostate was 1.5 (+/-0.2)mm. The average (+/- S.D.) treatment margin required to cover the tumor was 0.7 (+/-0.1)mm. We also demonstrated the challenges in registering an in vivo deformed MRI to an in vivo non-deformed MRI. Discussion: We here present a deformable registration scheme that can overcome large deformation. This platform is expected to be useful for prostate cancer radiation treatment planning.