filter blurs the edges of the cartilage, thus making it
harder to locate the correct boundary.
Figure 2a displays a common problem that ASM
exhibits. Landmarks are attracted to regions with
strong edges, therefore in slices where the tibial and
femoral cartilage touch, there is no edge information
for the upper boundary of the tibial cartilage and ASM
locates the upper boundary of the femoral cartilage
instead. The resulting shape estimate from ASM also
tends to be uneven, compared to PAAM and AFM.
This is because when new landmark locations are
found, the pose and shape parameters are updated to
best match the shape model to these landmark points.
This step can potentially shift the shape model away
from the true landmark positions.
All methods perform very poorly on a few par-
ticular slices, namely slices located at the very ends
of the cartilage in the middle of the knee. On these
slices, the cartilage areas are very small and also dis-
play very poor contrast and high shape variability.
Even the manual operators exhibit some discrepancy
in their results for these slices.
4 CONCLUSION
In this paper, we have compared four different model-
based segmentation methods (ASM, AAM, PAAM
and AFM) for the purpose of tibial cartilage segmen-
tation. All four methods have been fine tuned so that
only the optimal settings for cartilage segmentation
were compared. In conclusion, methods that work on
local texture perform better for cartilage segmentation
due to a restriction to more relevantinformation being
used for regression and fitting. In addition, the use of
landmark connectivity for orientation consistency re-
sults in an even more specific description of the tex-
ture.
ACKNOWLEDGEMENTS
The authors would like to thank Fahad Hanna and
Yuanyuan Wang for their time and contribution with
performing the manual cartilage segmentation used in
this paper, and also Ren´e Donner for providing parts
of the AFM implementation. Georg Langs has been
supported by the Austrian Science Fund(FWF) under
the Grant P17083-N04 (AAMIR).
REFERENCES
Cheong, J., Suter, D., and Cicuttini, F. (2005). Development
of semi-automatic segmentation methods for measur-
ing tibial cartilage volume. In Proc. DICTA 2005,
pages 307–314.
Cicuttini, F., Wluka, A. E., Forbes, A., and Wolfe, R.
(2003). Comparison of tibial cartilage volume and ra-
diologic grade of the tibiofemoral joint. Arthritis and
Rheumatism, 48:682–688.
Cootes, T., Edwards, G., and Taylor, C. (2001). Active ap-
pearance models. IEEE TPAMI, 23(6):681–685.
Cootes, T., Taylor, C., Cooper, D., and Graham, J. (1995).
Active shape models - their training and application.
Computer Vision and Image Understanding, 61(1).
Donner, R., Reiter, M., Langs, G., Peloschek, P., and
Bischof, H. (2006). Fast active appearance model
search using canonical correlation analysis. IEEE
TPAMI, 28(10):1690 – 1694.
Faggian, N., Paplinski, A., and Sherrah, J. (2006). Local
texture patches for active appearance models. In Proc.
IVCNZ 2006.
Faggian, N., Romdhani, S., Paplinski, A., and Sherrah, J.
(2005). Color active appearance model analysis using
a 3d morphable model. In Proc. DICTA 2005, pages
407–414.
Freeman, W. T. and Adelson, E. H. (1991). The design and
use of steerable filters. IEEE TPAMI, 13(9):891–906.
Goodall, C. (1991). Procrustes methods in the statistical
analysis of shapes. Journal of the Royal Statistical
Society, 53:285–339.
Langs, G., Peloschek, P., and Bischof, H. (2003). Asm
driven snakes in rheumatoid arthritis assessment. In
Proc. SCIA 2003, pages 454–461.
Langs, G., Peloschek, P., Donner, R., Reiter, M., and
Bischof, H. (2006). Active feature models. In Proc.
ICPR 06,, volume 1, pages 417–420.
Oost, C. R., Lelieveldt, B. P. F.,
¨
Uz¨umc¨u, M., Lamb, H. J.,
Reiber, J. H. C., and Sonka, M. (2003). Multi-view
active appearance models: Application to x-ray lv an-
giography and cardiac mri. In IPMI 2003, pages 234–
245.
Raynauld, J. (2002). Magnetic resonance imaging of ar-
ticular cartilage: toward a redefinition of ”primary”
knee osteoarthritis and its progression. The Journal of
Rheumatology, 29:1809–1810.
Thodberg, H. (2003). Minimum description length shape
and appearance models. In IPMI 2003, pages 51–62.
van Ginneken, B., Frangi, A. F., Staal, J., ter Haar Romeny,
B. M., and Viergever, M. A. (2001). A non-linear
gray-level appearance model improves active shape
model segmentation. In Proc. MMBIA 2001, pages
117–127. IEEE Computer Society Press.
Yan, S., Liu, C., Li, S., Zhang, H., Shum, H.-Y., and Cheng,
Q. (2002). Texture-constrained active shape models.
In Proc. GMBV 2002, pages 107–113.