Abstract
We present a new method for the simultaneous, nearly automatic segmentation of liver contours, vessels, and tumors from abdominal CTA scans. The method repeatedly applies multi-resolution, multi-class smoothed Bayesian classification followed by morphological adjustment and active contours refinement. It uses multi-class and voxel neighborhood information to compute an accurate intensity distribution function for each class. Only one user-defined voxel seed for the liver and additional seeds according to the number of tumors inside the liver are required for initialization. The algorithm do not require manual adjustment of internal parameters. In this work, a retrospective study on a validated clinical dataset totaling 20 tumors from 9 patients CTAs� was performed. An aggregated competition score of 61 was obtained on the test set of this database. In addition we measured the robustness of our algorithm to different seeds initializations. These results suggest that our method is clinically applicable, accurate, efficient, and robust to seed selection compared to manually generated ground truth segmentation and to other semi-automatic segmentation methods.
Keywords
Source Code and Data
No source code files available for this publication.
Reviews
Xiang Deng
Friday 25 July 2008
This paper presents a semi-automatic live tumor segmentation technique.
In this method, the liver tumors are first segmented using a multi-resolution multi-class Bayesian classification
algorithm . Morphological operation and active contour method are then used to refine the segmentation.
Detailed comments:
1) This is a well-organized workshop paper. The presentation is clear.
2) From my point of view, the multi-resolution approach might cause difficulty in segmenting small liver tumors, given the large variability of lesion size. Could the authors briefly address this issue?
3) The liver tumors may appear as either hypo-intensity or hyper-intensity, compared to the surrounding liver tissue.
Is this taken into account in the 5-class intensity model?
4) Could the user manually identified point in the initialization be an arbitrary point in the liver, e.g. a point either in normal liver tissue or in the lesion?
5) In Table 2, the values of the five metrics are good enough to show the reproducibility of the proposed method.
The last column on the right should be deleted, because those scores are not generated by comparision with provided reference segmentation.