Interactive Segmentation of 3D MRI Images

Interactive Segmentation of 3D MRI Images

Stať ve sborníku v databázi WoS či Scopus

The development of segmentation methods suitable for 3D image processing is a very important step in 3D data analysis, object detection and recognition/classification. Such methods are being developed especially in medical imaging and also in robotics where 3D data is available. In this project, we aimed to implement an existing 3D segmentation method using graph cuts, a method which has drawn much attention of the imaging community due to its interactivity, ability to find a global optimum of segmentation, performance of the existing algorithms and the possibilities of extending this method. In the paper we will introduce the state of art in graph-cuts theory, showing the relation to the segmentation problem and stating the segmentation criteria. Then we will look at the existing max-flow algorithm implementations and describe the version which was implemented in this project. Then we will focus on the evaluation of the implemented method, showing the performance on both artificial and real objects. With artificial objects, the characteristics of the data are preset thus we can clearly evaluate the algorithm and prove the correct functioning. With real 3D images obtained using magnetic resonance scanners, there are ambiguities in the solution and we therefore relied on a ground truth which has been assessed in a cooperation of a technician and an expert. Due to under-segmentation problems during the evaluation, we have also explored ways of introducing the regional properties in graph cuts, using intensity characteristics of the analyzed objects. The evaluation of the performance using these regional properties will be shown and finally, we will present the future outlooks of the research.

The development of segmentation methods suitable for 3D image processing is a very important step in 3D data analysis, object detection and recognition/classification. Such methods are being developed especially in medical imaging and also in robotics where 3D data is available. In this project, we aimed to implement an existing 3D segmentation method using graph cuts, a method which has drawn much attention of the imaging community due to its interactivity, ability to find a global optimum of segmentation, performance of the existing algorithms and the possibilities of extending this method. In the paper we will introduce the state of art in graph-cuts theory, showing the relation to the segmentation problem and stating the segmentation criteria. Then we will look at the existing max-flow algorithm implementations and describe the version which was implemented in this project. Then we will focus on the evaluation of the implemented method, showing the performance on both artificial and real objects. With artificial objects, the characteristics of the data are preset thus we can clearly evaluate the algorithm and prove the correct functioning. With real 3D images obtained using magnetic resonance scanners, there are ambiguities in the solution and we therefore relied on a ground truth which has been assessed in a cooperation of a technician and an expert. Due to under-segmentation problems during the evaluation, we have also explored ways of introducing the regional properties in graph cuts, using intensity characteristics of the analyzed objects. The evaluation of the performance using these regional properties will be shown and finally, we will present the future outlooks of the research.

graph cuts, interactive segmentation, magnetic resonance imaging

graph cuts, interactive segmentation, magnetic resonance imaging

JEŽ, O.; GRAHAM, J.

01.01.2007

Košíce, Slovakia

Proceedings of the 8th INTERNATIONAL CARPATHIAN CONTROL CONFERENCE

1

4