Mario Serrano-Sosa1, Jared Van Snellenberg2, Jiayan Meng2, Jacob Luceno2, Karl Spuhler3, Jodi Weinstein2, Anissa Abi-Dargham2, Mark Slifstein2, and Chuan Huang2,4
1Biomedical Engineering, Stony Brook University, Stony Brook, NY, United States, 2Psychiatry, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States, 3Radiation Oncology, NYU Langone, New York, NY, United States, 4Radiology, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States
Synopsis
Segmenting striatal subregions can be
difficult; wherein atlas-based approaches have been shown to be less reliable
in patient populations and have problems segmenting smaller striatal ROI’s. We
developed a Multi-Task Learning model to segment multiple 3D striatal
subregions using a Convolutional Neural Network and compared it to the Clinical
Imaging Center atlas (CIC). Dice Score Coefficient and multi-modal objective
assessment (PET and fMRI) were conducted to evaluate the reliability of
MTL-generated segmentations compared to atlas-based. Overall, MTL-generated segmentations were more
comparable to manual than CIC across all ROI’s and analyses. Thus, we show MTL
method provides reliable striatal subregion segmentations.
Introduction
Striatal
pathology or dysfunction is thought to play a critical role in neurological and
neuropsychiatric diseases such as schizophrenia, Parkinson’s, and Huntington’s
disease1-4. Widely available
atlas-based approaches, such as Freesurfer and the Imperial College London
Clinical Imaging Center (CIC), can
robustly identify the whole striatum and its major structures (the caudate and
putamen), but have been less reliable in segmenting the striatal subregions,
particularly in-patient populations5. Atlas-based
approaches, although readily available to implement, require nonlinear warping
of the input volume to the atlas. At times, this may create deformation fields
that will affect fine details, especially for smaller ROI’s such as striatal
subregions6; indicating that a more optimal method for striatal
subregion segmentation is needed that is comparable to manually drawn ROIs.
Multi-Task Learning (MTL) is a
technique used in deep learning that allows representations between related tasks
to be shared that can generalize the model to predict the original task, and
many more, with better accuracy7. We propose to utilize MTL to segment subregions of the
striatum 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). Analysis of MTL-generated segmentations was also compared to
CIC-generated segmentations. Objective assessment was applied to the automated
segmentations to analyze PET binding potential data, as well as resting state
functional connectivity (RSFC) fMRI data, compared to manually drawn ROIs.Methods
A
total of 68 datasets consisting of patients with schizophrenia and matched
controls were used to train the network. The 3D T1-weighted image were
used as input for the MTL network and six output tasks were simultaneously
trained on manually drawn ROI’s; where five of the six were ROI masks and the
last was the masked background. 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 between output and ground truth. For an additional
comparison, striatal substructure ROIs from the CIC atlas8 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. Afterwards, task-based assessment was evaluated using
automated segmentations to measure PET and fMRI quantification.
An independent dataset of 19 volunteers imaged
with [11C]raclopride PET and MRI
were used for the testing set. 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)9 with cerebellum as reference tissue. Linear Regression was performed
with BPND obtained from manual and CNN-generated ROIs. R2
and percent differences were reported.
A subset of 16 participants from the PET test
dataset described above also underwent 30 minutes of RSFC scanning and made up
the fMRI test dataset. Multiband blood oxygen level dependent (BOLD) MR
sequences were acquired using a multiband acceleration factor of 6, no in-plane
acceleration or parallel imaging, with 66 slices, 192 mm field-of-view, 2 mm
isotropic voxel size, 60° flip angle, 850 ms TR, and 25 ms TE. Raw timeseries
were then extracted from both hand-drawn and automated ROIs following
preprocessing, and the correlation between the hand-drawn and CNN timeseries
data was calculated for each fMRI run. Also, timeseries correlations between
each voxel were calculated for each ROI and averaged across runs to obtain
whole-brain RSFC correlations. Finally, the correlation across voxels in RSFC
was calculated between hand-drawn and automated ROIs.Results
Figure 2 shows an
example of manual and automated segmentations. Visually, MTL-generated
segmentations are similar to manually segmented ROIs. As shown in Figure 2,
MTL-generated segmentation seems to smooth VST where manual segmentation may
have sharp edges. When comparing DSC, MTL-generated segmentations were more
comparable to manual segmentations than CIC across all ROI’s (Table 1). When
comparing PET quantification, Figure 3 and Table 2 reflect linear regression
analysis, R-square and percent difference of MTL segmentations were more
comparable to manual segmentations than CIC across all ROI’s. In terms of fMRI
quantification, MTL segmentations also had closer correlation to manually drawn
ROI’s (Table 3, Figure 4).Conclusion
In this study, we developed a deep MTL framework
for striatal subregion segmentation and assessed the network outputs with PET
and fMRI-derived quantitative outcome measures. We show here that our MTL
method provides reliable striatal subregion segmentations for the purpose of
PET quantification, fMRI timeseries extraction, and whole-brain RSFC
correlations, with more comparable PET and fMRI results that more closely match
those obtained with manually drawn ROIs than atlas-based segmentations.Acknowledgements
The authors would like to thank Xiaoyan Xu PhD
for performing the manual segmentation on some of the data sets. Datasets were funded by P50 MH 086404 (AAD), K01 MH 107763 (JXVS), R21 MH 099509
(AAD). JJW was supported by K23 MH 115291, JXVS was supported by K01 MH 107763,
and CH was supported by NARSAD 24971.References
1. Horga G, Cassidy CM, Xu X, et al.
Dopamine-Related Disruption of Functional Topography of Striatal Connections in
Unmedicated Patients With Schizophrenia. JAMA
Psychiatry. 2016;73(8):862-870.
2. Laruelle M, Abi-Dargham A, Gil R,
Kegeles L, Innis R. Increased dopamine transmission in schizophrenia:
relationship to illness phases. Biological
psychiatry. 1999;46(1):56-72.
3. Howes OD, Kambeitz J, Kim E, et al. The
nature of dopamine dysfunction in schizophrenia and what this means for
treatment: meta-analysis of imaging studies. Archives of general psychiatry. 2012;69(8):776-786.
4. Teichmann M, Gaura V, Démonet J-F, et
al. Language processing within the striatum: evidence from a PET correlation
study in Huntington's disease. Brain : a
journal of neurology. 2008;131(4):1046-1056.
5. Makowski C, Beland S, Kostopoulos P, et
al. Evaluating accuracy of striatal, pallidal, and thalamic segmentation
methods: Comparing automated approaches to manual delineation. Neuroimage. 2018;170:182-198.
6. Gouttard S, Styner M, Joshi S, Smith
RG, Hazlett HC, Gerig G. Subcortical structure segmentation using probabilistic
atlas priors. Paper presented at: Medical Imaging 2007: Image Processing2007.
7. Ruder S. An overview of multi-task
learning in deep neural networks. arXiv
preprint arXiv:170605098. 2017.
8. Tziortzi AC, Searle GE, Tzimopoulou S,
et al. Imaging dopamine receptors in humans with [11C]-(+)-PHNO: dissection of
D3 signal and anatomy. Neuroimage. 2011;54(1):264-277.
9. Lammertsma AA, Hume SP. Simplified
reference tissue model for PET receptor studies. NeuroImage. 1996;4(3 Pt 1):153-158.