Abstract
1175
Introduction: Segmenting the human striatum to assess functionality has proven to be critical in studying neurological and neuropsychiatric diseases. Studies in nonhuman primates have established a clear functional organization of projections to and from complex striatal subregions1. . Imaging researchers, particularly in positron emission tomography (PET), have translated this functional organization into operationally defined substructures that can be manually traced on T1-weighted MR images2,3. Nevertheless, this manual segmentation is time-consuming. More rapid atlas-based methods have been developed such as the Imperial College London Clinical Imaging Center (CIC). Although easy to implement, atlas-based methods typically nonlinear warp the input volume which may affect smaller ROI’s such as striatal subregions4. We have developed a deep learning technique using Multi-Task Learning (MTL) that jointly learns how to segment striatal subregions consisting of pre-commissural putamen (prePU), pre-commissural caudate (preCA), post-commissural putamen (postPU), post-commissural caudate (postCA), and ventral striatum (VST) using a convolutional neural network (CNN). In this work, we employ a 3-D U-Net architecture with single MRI input and multiple segmentation outputs and compare it to CIC segmentations using PET and fMRI objective assessment.
Methods: 68 3D T1-weighted images and their corresponding manually drawn striatal subregion segmentations were used to train the network. The datasets consisted of patients with schizophrenia and matched controls. As shown in Figure 1, this MTL network consisted of a 3D U-Net architecture. This model was trained to minimize the sparse softmax cross entropy. An independent dataset of 19 volunteers imaged with [11C]raclopride PET and MRI were used for the testing set. For additional comparison, striatal substructure ROIs from the CIC atlas5 were retrieved from the MIAKAT (Imanova, Ltd; London, UK) software package, and applied to the test data. Dice Similarity Coefficients (DSC) were utilized to initially evaluate the performance of the automated methods. Multi-modal assessment consisted of PET and fMRI analysis. 60 min of dynamic emission data were acquired on an Siemens mCT scanner, following a bolus injection of 349 +/- 109 MBq of [11C]raclopride. Data were reconstructed by FBP with CT used for attenuation correction. Binding Potential (BPND) was derived in each ROI using simplified reference tissue model (SRTM)6 with cerebellum as reference tissue. Linear Regression was performed with BPND obtained from manual and MTL-generated ROIs and R2 was reported. fMRI test dataset consisted of multiband blood oxygen level dependent (BOLD) MR sequences. Raw timeseries and whole-brain RSFC were extracted from both hand-drawn and automated ROIs following preprocessing. Finally, the correlation across voxels in RSFC was calculated between hand-drawn and automated ROIs. Results: Striatal segmentations from manual and automated methods can be seen in Figure 2. When comparing DSC, MTL-generated segmentations were more comparable to manual segmentations than CIC across all ROI’s (Table 1). When comparing PET quantification, R-square show that MTL is more comparable to manual segmentations than CIC across all ROI’s (Table 2). In terms of fMRI connectivity, MTL segmentations also had closer correlation to manually drawn ROI’s (rMTL > 0.87 ; rCIC > 0.76) across all ROI's. Conclusion: The developed MTL framework for striatal subregion segmentation shown here provides reliable segmentations with more comparable PET and fMRI results that more closely match those obtained with manually drawn ROIs than atlas-based segmentations.
Mean Dice Similarity Coefficient (DSC) with standard deviations for all ROIs in both CNN and CIC ana
. Regression coefficients from correlative analysis between BPND calculated from manually drawn ROIs
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.