Studies On The Automated Segmentation Of Cerebral Gray Matter Nuclei In MRI Quantitative Susceptibility Maps

Cerebral gray matter nuclei have strong correlation with some neural disorders.Thus researches on cerebral gray matter nuclei have crucial impact on improving the clinical therapy as well as the pathological mechanism studies.Whereas the quantitative studies of the function and morphology of these nuclei may benefit from the availability of accurate automated segmentation methods.In this article,we investigated the effectiveness and relevant application of several automated segmentation methods applied for segmenting cerebral gray matter nuclei in the MRI quantitative susceptibility maps(QSM).We discussed the challenge of segmentation in MRI and the benefits of using QSM technique in our study.Six nuclei: caudate nucleus(CN),putamen(PUT),globus pallidus(GP),dentate nucleus(DN),substantia nigra(SN)and red nucleus(RN)were segmented in the study.The atlas method is the most frequently-used way among all automated segmentation methods.With both the single-atlas label propagation method and the multi-atlas label fusion method,the segmentation of the CN,PUT,GP and DN were studied.However,the results of SN and RN is not satisfying.Thus,we proposed a seed-points discontinuity based on level set method that exhibited an improved segmentation performance on the SN and RN.For the test of comparing patients with Parkinson's Diseases(PD)and healthy controls(HC),Statistical indicator of the SN and RN segmentations by the proposed method,atlas method and level set method did not show significant difference between two groups(PD vs HC).Finally,we studied the segmentations of these six nuclei using the convolutional neuro network(CNN).The CNN segmentation results were more accurate than the atlas method's results.The constantly accumulation of data for training and the optimization of the network architecture will endow this method with a great potential in automated segmentation of cerebral gray matter nuclei in medical images.