Automatic Segmentation Of Hippocampal Subfields Using Sparse Representation And Discriminative Dictionary Learning

Hippocampus is a tiny and complex anatomic structure located in both sides of the brain’s temporal lobe, which plays a critical role in human’ s spatial location ability and memory. Clinical research indicates that the volume’s changes of hippocampus and its subfields are closely associated with many neurodegenerative diseases, such as Alzheimer’s disease, temporal lobe epilepsy and elderly depression. Manual segmentation method can achieve high segmentation accuracy. However, this manual approach is too laborious to deal with large-scale clinical data. It’s of great clinical value to study automated segmentation techniques, which can reduce human intervention and improve segmentation accuracy.Multi-atlas based segmentation methods explore the prior of atlases and achieve segmentation of hippocampus by atlas image registration and label fusion. Accuracy of registration has to be very high in the traditional multi-atlas methods, while the label fusion strategies fail to balance the weights for each atlas effectively. To solve the above problems, this thesis regards hippocampal subfields segmentation as a multi-classification problem. A hippocampal subfields segmentation methods based on sparse representation and discriminative dictionary learning is proposed in this thesis. The following three aspects are studied in detail:(1) Image classification methods based on sparse representation and dictionary learning are studied. We first make an intensive study of theory of sparse representation and their application in the field of image classification. Then we discuss the different models for dictionary learning, by which two type of dictionaries can be obtained:dictionary with representative ability and dictionary with discriminative ability. We describe the basis about the model of discriminative dictionary learning and disadvantages of this kind of models. Finally, we propose to introduce this model from the field of face recognition to the field of image segmentation.(2) A model for discriminative dictionary learning is prosed to achieve segmentation of hippocampal subfields. On the basis of the model of dictionary learning used for face recognition, we add a label error term to the object function, according to LBP feature of atlases’ label image. A discriminative dictionary can be obtained by dictionary learning for segmentation of nine hippocampal subfields. The comparison with other segmentation methods shows that the proposed method achieves an average DiceSimilarity Coefficient(DSC) of 0.890 for the larger hippocampal subfields, outperforms typical approaches based on multi-atlas.(3) Patch-based technique is proposed to improve segmentation accuracy. We can remedy the registration error between the image to be segment and atlas images by using all the patches in searching windows for sparse representation and dictionary learning. Besides, computing complexity and segmentation accuracy can be balanced well by setting the appropriate sizes of patches and searching windows.