Biomedical Engineering Reference
In-Depth Information
the range of inter- and intraobserver variability for both ventricles and com-
pared favorably to similar studies performed by other groups using RT3D ul-
trasound for quantification of cardiac function. Manual tracing measures were
significantly less reliable with large standard deviation of errors and low cor-
relation coefficients. Finally, the 3D level set deformable model achieved the
highest degree of accuracy, which can be explained by a more accurate segmen-
tation of small and distorted ventricular shapes when integrating the third spatial
dimension.
2.5
Conclusion
Level set methods for segmentation and registration of medical images have
been the focus of intense research for the past decade producing very promising
results. Major advantages of the method include its robustness to noisy condi-
tions, its aptitude in extracting curved objects with complex topology and its
clean numerical framework of multidimensional implementation. Despite their
success, these methods still need to be refined to address two limitations:
1. Computation time needs to be further reduced, for viability of the method
in clinical application where interactivity (and therefore close to real-time
computation) is critical. This optimization will have to handle the con-
stant increase in data size observed in medical imaging applications with
improvements of spatial resolution, temporal resolution and now the in-
troduction of combo scanners such as PET/CT machines.
2. Robustness to variation in image quality and organ anatomy needs to be
studied. Unfortunately, the methods described in this chapter were only
rarely validated in clinical studies. On the other hand, it is well known
that these methods require tuning of their parameters to adapt to the na-
ture of the image data to segment. In that optic, it is therefore critical to
evaluate robustness of the performance on a set of data that covers the
range of quality encountered in clinical practice for a particular exami-
nation. For methods based on shape models, it is also critical to test the
method on a variety of abnormal (e.g., disease) cases that differ from the
average anatomy that they typically represent. Such validation for medical
application should always clearly specify the context of the problem at
