Research On Three Dimensional Segmentation Of Volumetric Medical Magnetic Resonance Images

Magnetic resonance imaging(MRI)has become the primary auxiliary means for human body function,pathological and anatomical research.Accurate segmentation of MR images is of great importance for analysis of pathophysiologies,acquiring of key biomedical indicators,modeling of tissue biophysics.However,due to the nonuniform magnetic fields,variations across soft tissues and partial voluming effects,the MR images are often corrupted severely that appears as an intensity inhomogeneity in these images and largely overlapping of intensity profiles across tissue structures,which complicates the segmentation of the detailed structures in MR images.Furthermore,low image contrast,considerable variability in soft tissue structures across subjects,and weak object boundary also cause considerable difficulties in medical image segmentation.All of these reasons make automated and accurate segmentation of MR images ramains a challenging problem.Many existing MR image segmentation methods suffer the problems of lacking of accuracy and robustness,etc.To address these issues,in this paper,we deeply research the three dimensional segmentation of the volumetric medical MR images.We first propose a segmentation model based on active contour model to deal with the complex image conditions,such as intensity inhomogeneity,noise and weak object boundary.The proposed model is then integrated into a voxel-wise classification combining active contour evolution framework to automatically and accuratly segment tissue structures in presence of intensity profile overlapping.Finally,we redesign the voxel classification architecture and the information integration scheme of the energy functional in the segmentation framework,and extend it into the multiphase segmentation of class imbalanced data with promising results.The main contributions in this work are four fold:Firstly,in order to cope with the difficulties caused by complex image conditions,such as intensity inhomogeneity,noise and weak object boundary,in the segmentation of volumetric MR images,this paper persents a robust statistics driven volume-scalable active contour model for segmentation of volumetric MR images.The proposed method simplifies the initialization of active contour by drawing several seeds in the object and/or background,and further defines an new energy functional by utilizing volume-scalable local information that are derived from the initial seeds and the neighbor region of the active contour.Due to the local robust statistics in constructing the feature vectors and the volume-scalable localization scheme in the energy fitting term,the proposed method can effectively cope with intensity inhomogeneity,noise and weak boundary,which thereby guarantees the accurate segmentation volumetric MR images with complex conditions.The accuracy of segmentation,robustness to complex conditions and contour initialization were validated on the basis of extensive synthetic and real datasets.Secondly,a combined concatenated random forests and volume-scalable active contour model approach is proposed for fully automatic segmentation of the volumetric MR images with low tissue contrast.Specifically,the proposed model formulates the segmentation problem as a hubrid problem with tissue classification and tissue boundary contour evolution.The voxle-wise classification results inferred from the random forests provide the active contour model with elaborate initialization and shape constraints,while the contour evolution of the active contour model refines the voxel classification results.In this way,these two models can complement with each other.The proposed method for integrating the local and contextual volumetric data information can drive the random forests and active contour model more effectively.In contrast to previous random forest based segmentation methods,the proposed concatenated scheme can iteratively refine the segmentation results and fuse the multiple concatenated branches.Compared to standard supervised learning schemes,the training process of the proposed concatenated random forests can flexibly select training samples for each concatenated branch while avoiding the retraining of the whole model that caused by the changing of the training samples,which improves the scalablity of the model.Furthermore,the proposed segmentation framework combines the independent voxel classification with the contour evolution,the segmentation results of the random forests are therefore refined to achieve smooth and geometrically contrainted boundary.In addition,compared to other active contour model based methos,the active contour model in the proposed segmentation framework is integrated with the initial contour and shape constraint,and therefore can perform segmentation automatically.Due to the initial contour is close to the final contour,the iteration number of the contour evolution is reduced.Finally,with the shape prior constraint,the model can achieve promising regimentation results for tissre structure with intensity profile overlapping.Thirdly,a random forests based concatenated and connected segmentation model is proposed.The proposed model can integrated the multi-scale and task-adapted information from multimodal MR volumes,and achieve automated,accurate and robust segmentation of tissue structures.Specifically,the random foreasts are employed as the imaging modality specific feature learning kernels to perform representation learning on graphs directly from multimodal MR volumes.In this way,the task-adapted feature representations learning scheme is implemented,which can achieve specific object boundaries.The concatenated and connected architecture in the proposed model guarantees the segmentation accuracy,while the two-phase training strategy can effectively deal with the class imblanced data.The regularization scheme in the training process of the peoposed model can effectively improve the ability of generalization.Comparative experiments demonstrate the effectiveness of the proposed segmentation framework.Finally,a combined concatenated and connected random forests and multi-scale patch driven active contour model approach is proposed for segmentation of volumetric MR images.The preliminary segmentation is first performed by using the concatenated and connected random forests,the preliminary segmentation results are then fed into a multiphase active contour model as an initial contour as well as a shap constraint by a novel prior information integration method.Furthermore,the multi-scale and multimodal spatial constraint information from the MR volumes are integrated into the energy functional of the active contour model by using the patch based sparse representation technique.Finally,the energy functional is minimized to achieve the accurate multiphase segmentation of the MR volumes by the coupled level sets.The proposed MR volume segmentation model achieves promising results on public available data sets,and the comparative experiments demonstrate the significant advantages of the proposed method over other state-of-the-art automated segmentation methods.