Concurrent segmentation of the prostate on MRI and CT via linked statistical shape models for radiotherapy planning
(a) MR image with prostate gland boundary delineated in 2D. (b) 3D rendering of the prostate shape as seen on MRI. (c) 2D CT image of the prostate, and (d) 2D CT image with delineation of the prostate gland overlaid. (e) 3D rendering of the prostate as seen on CT.
Examples of (a) small FOV diagnostic MRI, (b) large FOV planning MRI, and (c) planning CT images of the prostate. Note that unlike in Figs. 2(a) and 2(b), the prostate is not discernible on the CT [Fig. 2(c)]. The planning MRI serves as a conduit, by overcoming the differences in FOV between CT and diagnostic MRI, to register the diagnostic MRI with the corresponding planning CT.
Summary of the fLSSM strategy for prostate segmentation on CT and MRI. The first step (module 1) shows the fusion of MR and CT imagery. The diagnostic and planning MRI are first affinely registered (shown via a checker-board form) followed by an elastic registration of planning MRI to corresponding CT. The diagnostic MRI can then be transformed to the planning CT (shown as overlayed images). The second step (module 2) shows the construction of the fLSSM where PCA is performed on aligned training samples of prostates from both CT and MRI such that their shape variations are then linked together. Thus, for any standard deviation from the mean shape, two different but corresponding prostate shapes can be generated on MRI and CT. The last step (module 3) illustrates the fLSSM-based MRI-CT segmentation process, where we extract different kinds of features from each image to train a random forest classifier which is combined with a probabilistic atlas to identify voxels that are likely to be inside the prostate. The fLSSM, which is concurrently initialized on the classification result on both MRI and CT, is subsequently evolved to yield concurrent segmentations on both modalities.
Representative gradient feature images (a) and (b) on MRI, (c) and (d) on CT.
Summary of the different types of ground truth used for training and testing the various SSMs considered in this study. Expert, trainee and MRI-CT registration based ground truth are shown on the left side of the figure (respectively in red, green and blue). The right side of the figure shows the automatic segmentations via the different SSMs compared against the expert ground truth, the automated segmentations are shown in yellow.
3D renderings of MRI and CT segmentation results for a patient study, where the surface is colored to reflect the surface errors (in mm) between prostate surface segmentation and the associated . Segmentation of prostate on (a) CT using the ctSSM evaluated agianst , (b) CT using the xLSSM evaluated agianst , (c) CT using fLSSM evaluated against , (d) CT using tLSSM(1) evaluated against , (e) CT using tLSSM(2) evaluated against , and (f) MRI using any LSSM evaluated against .
Leave-one-out reconstruction accuracy obtained for each patient study using both MRI and CT shapes. The bars on the left for a patient study represent the DSC obtained on MRI, while the bars on the right represent the DSC obtained on CT.
The average DSC obtained over ten runs when evolving the fLSSM on each patient study. For each of the ten runs, a random subset of voxels was sampled for the random forest classifier and a random perturbation was added to the centroid initialization. The bars on the left for a patient study represent the DSC obtained on MRI, while the bars on the right represent the DSC obtained on CT. The error bars represent the standard deviation in DSC for the ten runs on each patient study.
Prostate CT segmentation results for each of the 20 patient studies used in this study. Error bars representing the standard deviation in DSC over ten runs are also shown.
Table of commonly used notation.
Summary of the prostate image data sets acquired for each of 20 patients considered in this study.
Mean DSC and MAD for prostate MRI to CT registration, along with their associated standard deviations, calculated over all 20 patient studies.
Mean DSC and MAD for prostate MRI segmentation using the different LSSMs.
Mean DSC and MAD for prostate CT segmentation using the different SSMs.
p-values from two-tailed paired t-tests to identify whether there are statistically significant differences between CT segmentation results (in terms of DSC) of different pairs of SSMs.
Article metrics loading...
Full text loading...