Automatic Mesh-Based Segmentation of Multiple Organs in MR Images

ABSTRACT: Segmentation of multiple anatomical structures in MR images is often required for biomedical engineering applications such as clinical simulation, image-guided surgery, treatment planning, etc. Moreover, there is a growing need for automatic segmentation of multiple organs and complex structures from this medical imaging modality. Many successful multi-object segmentation attempts were introduced for CT images. However in the case of MR images it is a more challenging task due to intensity inhomogeneity and variability of anatomy appearance. Therefore, state-of-the-art in multi-object MR segmentation is very inferior to that of CT images. In literature dealing with MR image segmentation, the region-based approaches are sensitive to noise and non-uniformity in the input image. The edge-based approaches are challenging to group the edge information into a coherent closed contour. The atlas-based techniques can be problematic for complicated structures with anatomical variability. Deformable models are among the most popular methods for automatic detection of different organs in MR images. However they still have an important limitation which is that they are sensitive to initial position and shape of the model. An unsuitable initialization may provide failure to capture the true boundaries of the objects. On the other hand, a useful aim for an automatic multi-object MR segmentation is to provide a model which promotes understanding of the structural features of the distinct objects within the MR images. The current automatic initialization methods which have used different descriptors are not completely successful in extracting multiple objects from MR images and we need to find richer information that is available from edges. In this regard, anisotropic adaptive meshes seem to be a potential solution to the aforesaid limitation. Anisotropic adaptive meshes constructed from MR images contain higher level, abstract information about the anatomical structures of the organs within the image retained as the elements shape and orientation. Existing methods for constructing adaptive meshes based on image features have a practical limitation where manifest itself in inadequate mesh elements alignment to inclined edges in the image. Therefore, we also have to enhance mesh adaptation process to provide a better mesh-based representation. In this Ph.D. project, considering the highlighted limitations we are going to present a novel method for automatic segmentation of multiple organs in MR images by incorporating mesh adaptation techniques. In our progress, first, we improve an anisotropic adaptation process for the meshes that are constructed from MR images where the mesh elements align adequately to the image content and improve mesh anisotropy along edges in all directions. Then the resulting adaptive meshes are used for initialization of multiple active models which leads to extract initial object boundaries close to the true boundaries of multiple objects simultaneously. Finally, the Vector Field Convolution method is utilized to guide curve evolution towards the object boundaries to obtain the final segmentation results and present a better performance in terms of speed and accuracy.

Abstract

Segmentation of multiple anatomical structures in MR images is often required for biomedical engineering applications such as clinical simulation, image-guided surgery, treatment planning, etc. Moreover, there is a growing need for automatic segmentation of multiple organs and complex structures from this medical imaging modality. Many successful multi-object segmentation attempts were introduced for CT images. However in the case of MR images it is a more challenging task due to intensity inhomogeneity and variability of anatomy appearance. Therefore, state-of-the-art in multi-object MR segmentation is very inferior to that of CT images. In literature dealing with MR image segmentation, the region-based approaches are sensitive to noise and non-uniformity in the input image. The edge-based approaches are challenging to group the edge information into a coherent closed contour. The atlas-based techniques can be problematic for complicated structures with anatomical variability. Deformable models are among the most popular methods for automatic detection of different organs in MR images. However they still have an important limitation which is that they are sensitive to initial position and shape of the model. An unsuitable initialization may provide failure to capture the true boundaries of the objects. On the other hand, a useful aim for an automatic multi-object MR segmentation is to provide a model which promotes understanding of the structural features of the distinct objects within the MR images. The current automatic initialization methods which have used different descriptors are not completely successful in extracting multiple objects from MR images and we need to find richer information that is available from edges. In this regard, anisotropic adaptive meshes seem to be a potential solution to the aforesaid limitation. Anisotropic adaptive meshes constructed from MR images contain higher level, abstract information about the anatomical structures of the organs within the image retained as the elements shape and orientation. Existing methods for constructing adaptive meshes based on image features have a practical limitation where manifest itself in inadequate mesh elements alignment to inclined edges in the image. Therefore, we also have to enhance mesh adaptation process to provide a better mesh-based representation. In this Ph.D. project, considering the highlighted limitations we are going to present a novel method for automatic segmentation of multiple organs in MR images by incorporating mesh adaptation techniques. In our progress, first, we improve an anisotropic adaptation process for the meshes that are constructed from MR images where the mesh elements align adequately to the image content and improve mesh anisotropy along edges in all directions. Then the resulting adaptive meshes are used for initialization of multiple active models which leads to extract initial object boundaries close to the true boundaries of multiple objects simultaneously. Finally, the Vector Field Convolution method is utilized to guide curve evolution towards the object boundaries to obtain the final segmentation results and present a better performance in terms of speed and accuracy.