Smoothing effect on calculated dose
We hypothesize that the radiation beams are less attenuated when a CT image has such
smoothing effect. This would subsequently result in overestimation of the calculated doses. To
test this hypothesis, we performed a smoothing on a real CT image via Gaussian convolution
with full width at half maximum of 6 mm and recalculated the dose. Dosimetric parameters
evaluated were found to be higher than the original values; for example, 1.1% of the prescription
dose in D95 to the PTV, 0.6% of the original value in maximum dose to the chiasm, and
3.1%/5.6% of the original values in mean dose to the left/right cochleae.
FIG S1. Age distribution of the test patient
(denoted by “Test”) and the two atlas groups. The
dots in this graph were “jittered” to avoid overlap.
Note that the group of 6 atlases is a subset of the
group of 12 atlases. The selections of the three
groups were made such that age and gender
distributions were not statistically different
between the test patient group and either of the
atlas groups (Student t-test on age, P>0.5;
Fisher’s exact test on gender, P>0.1).
FIG S2. Correlation between gross tumor
volume and the accuracy of constructed
pseudo CT evaluated by RMSD from real CT
(R=0.55, P=0.040, N=14).
FIG S3. Correlation between age difference
(between atlas and patient) and the accuracy
of constructed pseudo CT evaluated by RMSD
from real CT (R=0.18, P=0.02, N=168).
FIG S4. Demonstration of the case (patient #14) with problematic image deformation. The gray
scale of CT images was adjusted such that ventricular structure is displayed. The images are all
at the same slice location. The pseudo CTs constructed by different schemes (b-e) are
compared with the real CT (a) and the corresponding MR image (f). The yellow arrows on
pseudo CTs indicate a region where ventricle area in the patient image is occupied by brain
tissue in the deformed single atlas image due to mis-registration. The ventricular structure in this
region is not clear in MEAN6 but it is improved by PRGP6 and PRGP12.
FIG S5. Variation of cranial bone density and the relation with accuracy in deformation of atlas
images. (a) The distribution of bone density in test and atlas groups. The bone density was
estimated by CT numbers in cranial bone structure segmented by Otsu’s method. The mean ±
standard deviation values were: 1924±80 (N=14) for the test patients (denoted by “Test”) and
1927±75 (N=6) and 1899±79 and (N=12) for the two atlas groups. No difference between test
patient and either of the atlas groups was found (Student t-test, P>0.44). (b) Scatter plot
between ages of 26 test/atlas patients (please note that the 6 atlas group is a subset of the 12
atlas group) and CT numbers. No significant correlation was found (R=0.31, P=0.13). (c) When
counting only voxels within cranial bone, the age difference between patient and atlas images
was not correlated with the deformation accuracy any more (R=0.03, P=0.70, cf. Fig. S3). This
must have been contributed by the individual variation of bone density which is not strongly
associated with age.
(a) MR images of an example patient
(c) Atlases acquired by different
sets of parameters before
histogram matching
(d) Atlases in (c) after histogram matching
(denoted by “HM1” – “HM5”)
HM1
(b) Atlas acquired by the same
parameters as the patient images
(denoted by “SAP”)
HM2
HM3
Weighting
(e)
0.3
HM4
0.2
0.1
0
HM5
-0.1
SAP HM1 HM2 HM3 HM4 HM5
FIG S6. Evaluation of the histogram matching algorithm. MR images of an example patient in the test group (a) were compared
with images from an atlas sharing the same acquisition parameters with the patient (b, denoted by “SAP”) and the other atlases
acquired by different sets of parameters (c) but processed with the histogram matching algorithm (d, denoted by “HM1”-”HM5”).
The weighting values assigned to the atlases (‫ܓ‬Cିଵ in Eq. A3) during PRGP were evaluated. The boxplots (e) did not show any
apparent bias (i.e., higher weighing values) towards SAP against HM1 - HM5. Please note that all atlas images in this figure have
been deformed to the patient images.