Home < Projects:TopologicalSegmentation

Segmentation

Catheter ablation has been proposed for treatment of atrial ﬁbrillation arrhythmia. MRI data, obtained at University of Utah, are used to explore lesion ablation and scariﬁcation locations. In addition, MRI analysis may help to predict if the ablation procedure will help a patient or not. Many of these image analysis tasks are largely based on segmentation of left atrial wall, which is done manually or semi-automatically. Automatic segmentation uses moving contours or surfaces (interfaces) to segment image data by minimizing a predeﬁned energy function. These moving interfaces are highly aﬀected by image data, which can be thought as a force ﬁeld pushing the interface to features of choice. Thus, the choice of interface attracting image features is critical.

Our Approach

We are exploring different image features that are appropriate for such automatic segmentation tasks. Our objective is to develop and improve topological extrema ranking algorithms (e.g., maximum persistence, and perceptual ridge importance) to provide the most appropriate features, and to compare the results to expert’s manual segmentation.

Based on the obtained edge or ridge features, we are using two different approaches to create a contour. The first uses a novel boundary linking algorithm where all the detected ridges converted to 8-point extrema vector, and then the connections between the ridges are made using these vectors. The second approach uses active contours to provide smoother results.

Results

The example of extracted ridges is shown in the Figure 1. The connected boundary is shown in the Figure 2. For the tested volumes, when comparing the obtained boundary with expert's manual segmentation, the average distance error is about 2 pixels. The results of GVF active contour approach are shown in the Figures 3-5. The bright gray contour is the expert's segmentation, and yellow contour is the result of our algorithm. The average distance error there is less than 2 pixels.