Multi-Atlas Segmentation of the Prostate: A Zooming Process with Robust Registration and Atlas Selection
Yangming Ou, Jimit J. Doshi, Güray Erus, Christos A. Davatzikos · 2012
As an entry to the MICCAI 2012 Prostate Segmentation Challenge, this paper presents a multi-atlas-based automatic pipeline for segmenting prostate in MR images. Image registration is needed to trans- fer expert-dened ground-truth prostate segmentation from atlases onto target image. However, registration is rendered dicult in this dataset, due to dierent images having dierent eld of view (FOV), dierent imaging protocol (from multiple imaging centers), large anatomical vari- ability around prostate and large pathologic variability (some having enlarged prostate or prostate cancer). To overcome these limitations, we propose a process for multi-atlas-based prostate segmenta- tion. We rst register all atlases onto target image to obtain an initial segmentation of the prostate. This step includes robust registration, atlas selection and majority-based label fusion. Then, we zoom-in: re-run all atlas-to-target registrations, but this time restricting the registration to the vicinity of the prostate, ignoring compounding structures that are far away from the prostate and are largely variable. As a result, we can expect more accurate registrations and hence rened prostate segmen- tation. Our cross-validated results show improvement in segmentation by robust registration and atlas selection, compared to using all atlases. Additional improvement is observed when zooming in and focusing on registration of the prostate vicinity. We report average 0.84 dice overlap with expert-dened prostate segmentation in training subjects. Accuracy in testing datasets will be released by organizers of this challenge.