Atlas-based automatic segmentation of Brown Adipose Tissue in

Atlas-based automatic segmentation of Brown Adipose Tissue in
upper torso of lean human subjects in CT
Merisaari H1,2, U Din M1, Raiko J1, Teräs M1, Virtanen K1
1Turku
2Department
PET Centre, Turku University Hospital, 20521, Turku, Finland
of Information Technology, Turku University, 20014, Turku, Finland
OBJECTIVES
Brown adipose tissue (BAT) has a role in human non-shivering thermogenesis.
Physiological mechanisms governing BAT activity are frequently studied by
segmenting Brown Adipose Tissue (BAT) areas on CT images in hybrid PET-CT
studies. In addition, inclusion of potential, smaller BAT regions for image analysis
in whole body is challenging by free hand. For reproducibility of obtained
measurements there is a need to develop an automated technique for BAT
segmentation. Our aim was to develop fully automated BAT segmentation method
by using anatomical and tissue radiodensity information in CT images.
In the validation of the method, mean HU values in BAT were measured from
manual segmentations and with atlas-based method in 10 subjects containing
both conditions, with and without cold exposure condition. The Ground Truth
values from manual segmentations were compared to the automatic method.
Fig. 2 Sample image with
deformed
BAT
atlas
overlaid on CT in coronal
(A), sagittal (B) and
transaxial slices (C). The
approximate atlas region is
used to help automatic
localization of the BAT that
uses HU value thresholds
to exclude other than BAT
tissue .
METHODS
Segmentation method was developed utilizing expert knowledge about general
BAT localization, and Hounsfield Unit (HU) intensity values for creating Regions of
Interest (ROI). The ROIs are generated to CT image in a fully automatic manner
after ROIs are drawn once on the template image.
See Fig. 1 for schematic view on the delineation procedure. (A) All of the CT
images were co-registered to the first CT with multi-level non-linear co-registration
in open source registration package FAIR [1]. (B) Then, the CT template was
created by averaging the intensity values of the co-registered images together.
(C) For BAT atlas segments in left and right side of the body, two approximate
regions were delineated on the CT template into locations where BAT is expected
to be found. (D) Finally, ROI delineation is applied to individual subjects by coregistering the template to the subject CT, thus creating transformation from BAT
atlas to the individual. (E) The transformed BAT delineation is then used with HU
intensity thresholding for determining regions that are BAT. See Fig. 2 for sample
image of transferred atlas ROI.
For incorporating localization information by medical expert into our method (step
(C)), 22 CT images were randomly selected from a larger set of images, and a CT
template was created from them.
RESULTS
The mean HU measurements of 20 ROIs from 10 subjects were in agreement to the
measurements executed manually. When comparing the mean HU values in slices,
Bland-Altman bias was 1.1 and 2.7 HU and 95% limits of agreement in (-34.6, 36.9) and
(-21.9, 27.4) HU for left and right side of upper torso, correspondingly, (see Fig. 3).
(L)
(R)
CT1
CT1
…
CTN
CT1 co-registered
Non-rigid
co-registration
(A)
Averaging
(B)
CTN co-registered
Fig. 3 Comparison of BAT HU measurements with between manual and automatic
delineation method. (L) Left side of upper torso (R) Right side of upper torso.
Non-rigid
co-registration
CT with automatic
ROIs
DISCUSSION & CONCLUSION
Template
CT1 | ROIs
CTN | ROIs
(E)
Draw ROIs
to template
The proposed CT based automated BAT segmentation method provides a quick
and more reproducible way for BAT measurements in lean human subjects. The
method needs to be further investigated with obese subjects and larger patient
population to see the robustness in more varying CT data.
ACKNOWLEDGEMENTS
ROIs
(C)
Deformations
(D)
ROIs in CTi … CTN domain
Fig. 1 Schematic view of CT atlas creation followed by automatic ROI delineation. Atlas
is created from CT image data (A-B), followed by manual delineation of BAT into the
template (C). Finally, the template ROIs are utilized in automatic delineations (D-E).
This work was funded by Finnish Academy grant for strategic Japanese-Finnish
research Cooperative program on “Application of medical ICT devices”. Image
data collected in Turku PET Centre is highly appreciated.
REFERENCES
[1] Modersitzki, Jan. FAIR: flexible algorithms for image registration. Vol. 6. SIAM,
2009.