gms | German Medical Science

Objective: Precise image registration of intraoperative ultrasound (iUS) and preoperative MRI (preMRI) as well as the image fusion based on these is still an unsolved problem in image-guided surgery. Artifacts within 3D US reconstructions hinder direct fusion of preMRI and 3D iUS. A 2D-based segmentation of dural folds in iUS images, which can act as a guiding frame in 3D space, might improve the registration process. In this context we present our results of an automatic segmentation.

Method: Cases of parieto-occipital gliomas extending to the occipital horn of the lateral ventricle were chosen from our imaging database. The latter includes preMRI and corresponding sets of navigated iUS, recorded at various stages of the operation.

We developed a fast and easy to use automatic segmentation algorithm in MeVisLab that is able to segment the falx, the tentorium and adjacent gyri. All of these are anatomical structures with a line-type character in 2D iUS, represented as prevalent strong edges. The strategy is to initially smoothen homogenous regions of the iUS image for speckle reduction, while preserving and enhancing prominent edges of the structures of interest (SOI). The resulting image is processed by a hessian filter, enabling an analysis of the strength and direction of lines and edges. The region with the strongest/longest edges is detected as the region of interest (ROI). The ROI is used to filter the prominent lines and edges of the SOI from the result of the hessian filter. Finally, an object-based image analysis is applied to the segmented structures to improve the method by additional falx, gyri-, and tentorium-specific image features.

Results: The segmentation was tested on 100 iUS scans in which the angle, size, direction and contrast of the displayed SOI varied considerably. The results were compared to reference segmentations of an expert. The good quality of the segmentation results can be verified by the following quality yardsticks: On the average, our method achieved a Dice coefficient of 0.75, a Hausdorff distance of 3.7 mm, and a Jaccard index of 0.6 with a general processing time of < 1 second per slice.

Conclusions: We present a robust and automatic segmentation of the falx, the tentorium, and adjacent gyri, which is a challenging task due to shape and contrast variations of these structures. With our method, tedious manual segmentations can be omitted and it supports future rigid and deformable registrations, leading to an improved registration of pre- and intraoperative images.