Purpose: In multi-atlas based segmentation, label fusion, which incorporates labels by weighting each atlas label, is a mandatory process. Therefore, numerous label fusion methods have been proposed to compute more accurate voting weights. Among them,...
Purpose: In multi-atlas based segmentation, label fusion, which incorporates labels by weighting each atlas label, is a mandatory process. Therefore, numerous label fusion methods have been proposed to compute more accurate voting weights. Among them, the joint label fusion (JLF) method has been widely used for organ segmentation due to the advantage of compensating for dependent errors between atlases. However, the optimal JLF algorithm parameter set has not yet been explored, and even guidelines for setting it are insignificant. In this study, we investigate the feasibility of JLF algorithm for the prostate segmentation in MR images and suggest an optimal parameter set for automatic prostate segmentation. Our experimental results also provide a useful reference for setting parameters and atlas numbers in organ segmentation by JLF.
Methods: 20 healthy subjects participated in T2-weighted prostate MR imaging and a series of cross validation was conducted for all JLF parameter combinations. In each case, we rigidly registered the atlases to the target image. Then, each combination of JLF’s parameters ("r" _("p" _"xy" ), "r" _("p" _"z" ), "r" _("s" _xy ), "r" _("s" _"z" ), β) was applied to calculate voting weights for multi-atlas based segmentation. These segmentation results were evaluated by five validation metrics of the Prostate MR Image Segmentation (PROMISE12) challenge.
Results: The optimal parameters of JLF algorithm were "r" _("p" _"xy" )= 10, "r" _("p" _"z" )= 1, "r" _("s" _xy )= 3, "r" _("s" _"z" )= 1, and β = 3, and the segmentation results are as follows; DSC: 0.8495 ± 0.0392, RVD: 15.2353 ± 17.2350, aRVD: 18.8710 ± 13.1546, 95% HD: 7.2366 ± 1.8502 voxels, and ABD: 2.2107 ± 0.4972 voxels. As the number of voxels participating in the voting weight calculation and the number of referenced atlas are increased, the overall segmentation performance is improved. In addition, it was shown that JLF has an average tendency to over-segmentation of the prostate.
Conclusions: We suggest optimal JLF parameter sets from empirical analysis of segmentation results. The experiment shows the feasibility of the JLF for automatic prostate segmentation in MRI. Moreover, it also serves as a reference for setting appropriate parameters and atlas numbers in target organ segmentation.