Medical image segmentation is a challenging problem and a necessary first step in many image analysis and quantification methods. The segmentation of anatomical objects presents a particular problem especially for surgery and radiotherapy planning. Our research was motivated by the need of detailed virtual 3D hip models during the computer assisted 3D planning of orthopaedic interventions (Handels et al. 2001, Ehrhardt et al. 2003).
In the majority of cases, the segmentation work is done using manual or semiautomatic segmentation methods like region-/volume growing, thresholding, or snake methods. For many medical applications the live-wire segmentation has proved to be a robust and user-friendly semiautomatic method for the extraction of structure outlines. But for the segmentation of image volumes the user has to create a contour in each slice. Hence the method is very time-consuming and tedious.
The live-wire method has been extended to a semi-automatic segmentation of volume data in different approaches (Falcao, Udupa 2000, Schenk et al. 2001). These methods use orthogonal slice segmentation or interpolation techniques respectively to segment the three-dimensional structures on the basis of some initial contours.
User interaction is still necessary for creating adequate initial contours and for the interactive correction of the segmentation result.
Fig 1: Automatic segmentation of the os ischii in a patient dataset by transferring the atlas contour. Left: atlas image. Right: results of automatic segmentation.
Since in most clinical applications automatically created segmentation results have to be revised by experts and segmentation failures may still occur we developed an intuitive adjustment interface that facilitates the correction and optimisation of automatically created contours (Fig2).
Fig2 : Semiautomatic segmentation of the acetabulum. Moving one seed point corrects the contour. Left: acetabulum in the atlas image. Center: results of automatic segmentation. Right: results after manual adjustment.
Thus the method combines the advantage of automation with the flexibility of a manual segmentation.
The results showed that the tool is applicable for the segmentation of anatomical hip structures because 51% of interaction time could be saved while the segmentation quality is similar to the quality of manually created segmentations.
Furthermore the method can be applied to other medical problems.
A first evaluation on liver CT data showed promising results concerning the saving of time and the segmentation quality.
In a third application, the contour transfer method has been adapted for the segmentation of pathological structures. In this adaptation a manually created initial contour is transferred iteratively to the neighboured slices. This method shows a remarkable advantage in interaction time compared to manual methods during the segmentation of brain tumours in three-dimensional MR datasets (Fig3).
Fig. 3. Semiautomatic segmentation of a glioblastom. The manually created contour (3rd slice) has been transferred iteratively.
