Research On 3D Segmentation Of MR Brain Image Based On Geometric Deformable Model

As an important research direction of medical image segmentation, MR Brain image segmentation plays a crucial role in 3D visualization for anatomical brain structure, radiotherapy planning and surgery operation planning. Due to the complexity of MR brain image no technology could get satisfying segmentation results at present. Geometric Deformable Model is a new segmentation method that appeared in recent ten years. To a great extent the Model solves the existent problems in traditional segmentation technologies and becomes research hotspots in brain image segmentation domain. Based on Geometric Deformable Model the dissertation explores and studies brain image segmentation technology. Two fast Geometric Deformable Model methods: Fast Marching and Narrow Band were improved according to MR brain image's characteristics.Aiming at Narrow Band Method's low efficiency in great data quantity or 3D segmentation condition, the dissertation proposed a hybrid algorithm of Region Growing and Narrow Band. The new algorithm uses defectivesegmentation results of Narrow Band Method to guide region growing process. Accordingly the algorithm can finish segmentation task quickly. In 2D segmentation experiment, the Improved Narrow Band Method is double faster than traditional Narrow Band Method. In 3D segmentation experiment, the efficiency advantage of Improved Narrow Band Method is much more obvious. Aiming at relative low accuracy rate and leakage problems of Fast Marching Method, integrating region information the dissertation proposed Improved Fast Marching Method according to T1 weighted MR brain image's property. The experiment results indicate that the Improved Fast Marching Method better resolves the leakage problem of blurred image boundary.After each algorithm's parameter with a kind of parameter optimizing strategy was tuned, the segmentation results of Narrow Band, Improved Narrow Band, Fast Marching and Improved Fast Marching were evaluated。The evaluation results indicate that the average True Positive rate and average Overlap rate of Improved Narrow Band Method's segmentation outcomes are 3.5% and 2.1% higher than those of traditional Narrow Band Method's segmentation outcomes. In 2D situation, the average True Positive rate and average Overlap rate of Improved Fast Marching Method are 11.0% and 4.5% higher than those of traditional Fast Marching Method. In 3D situation, the two kinds of rate increase 16.3% and 8.4%.The proposed algorithms are not only applicable to normal tissues (such as lateral ventricle) segmentation but also suitable for pathological tissues (such as a tumor) segmentation in MR brain images. They play an important theoretical and practical value in the in-depth clinical application of MR image.