Segmentation of 3D Tubular Tree Structures in Medical Images ...

128 Chapter 8. Conclusion and Outlook
the seminal ideas of Beichel et al. [10, 113], only the works of Szymszack et al. [138],
Florin et al. [39], and Graham et al. [50] presented applications using similar ideas. However,
none of these methods has been evaluated on larger sets of clinical datasets or
compared to other methods.
As seen from the applications summarized above, the objectives and problems that
have to be addressed in such application areas for segmentation of tubular tree structures
are very diverse and challenging. But with the proposed general approach for segmentation
of branched tubular networks (Section 1.2) and the set of developed methods most of the
problems occurring in such kinds of datasets can be addressed and the developed methods
proofed to be suitable for different application domains.
8.2 Directions for Future Work
The general framework and the methods we have developed in this thesis provide a flexible
basis for future work in the area of vessel/airway segmentation. It can be adapted to other
applications or requirements by replacing or improving individual parts (of our three-step
approach). Besides the application and evaluation of the already developed methods in
other application areas, we see primarily two further directions for possible future work:
Improvement of the robustness of the tube extraction approach: The extraction
of tubular structures is the first and most critical part of the overall approach. For
larger tubular structures, the presented methods perform well and the developed methods
also proofed sufficient for identification of thinner low-contrast vessels in liver and
coronary CT datasets (Sections 5 and 6). However, in case of airways still about 1/3 of
the airways remain undetected (Section 7). The reasons for this behavior are the complex
background in lung CT datasets and the fact that even smaller local disturbances
can lead to problems in combination with the used height-ridge traversal procedure and
the hysteresis-thresholding (Section 2.4). One may consider to enhance the robustness of
the tube extraction method for smaller airways either by appropriate pre-processing of
the datasets, post-processing TDF results, or by incorporating additional airway-specific
prior knowledge into the airway detection. Therefor, pre-processing techniques for noise
suppression specifically developed for tubular structures such as “vessel enhancing diffusion”
[95] or “flux-based anisotropic diffusion” [66] could be used. Possible TDF result
post-processing methods could combine local evidence about candidate airways by taking
their elongated structure into account. This could be based on “improved structure ten-