A Constrained Deep Neural Network for Adaptive Smoothing of Task fMRI Data
Poster No:
1559
Submission Type:
Abstract Submission
Authors:
Zhengshi Yang1, Xiaowei Zhuang2, Mark Lowe3, Dietmar Cordes1
Institutions:
1Cleveland Clinic, Las Vegas, NV, 2cleveland clinic, Las Vegas, NV, 3The Cleveland Clinic, Cleveland, OH
First Author:
Co-Author(s):
Introduction:
Gaussian smoothing of task fMRI data was frequently used to improve the signal-to-noise ratio and facilitate the detection of cluster-like active regions [1]. However, it comes with the cost of spatial blurring artifact, namely inactive voxels near active voxels were mistakenly identified as active. Multiples techniques were previously proposed for adaptively smoothing task fMRI data. In this study, we proposed a constrained deep neural network (DNN) model, which reserves the advantages of previous adaptive smoothing approaches and addresses their limitations.
Methods:
Spatial smoothing could be treated as a weighted summation of original time series (Y) from a center voxel and its neighboring voxels and then assign the time series to the center voxel (y), formulated as y≡Yα. The spatial weight coefficient α is a constant vector across all voxels in Gaussian smoothing. The task stimuli were modelled in the design matrix X and then the activation of each voxel was the coefficient vector β optimized by running general linear model (GLM) to maximize the correlation r between y and Xβ. In contrast, the coefficient α is uniquely optimized for each voxel in the proposed DNN model. The model consists of a set of 3D sum-constrained convolutional layers followed by nonnegative-constrained fully-connected layers, with original time series as input and optimally smoothed time series as output. Stacking multiple convolutional layers could allow voxels to incorporating information from distant voxels beyond the nearest voxels, without substantially increasing computational burden. A customized cost function was defined to consolidate the activation in active regions and avoid exaggerated activation in inactive regions. 3T MRI imaging data from 88 subjects in the Human Connectome Project were used in this study. The structural T1 images were acquired with a resolution of 260×311×260 to yield 0.7mm×0.7mm×0.7mm isotropic voxel size. The working memory task fMRI data were acquired with 405 time points from a gradient-echo fast EPI sequence with parameters: multiband factor 8, TR/TE=720/33.1 ms; flip angle=52 degrees; 72 slices; spatial resolution=2mm×2mm×2mm and imaging matrix=104×90. A set of 20 simulated task fMRI sessions was generated to mimic real fMRI data for evaluating the performance of different techniques, including conventional canonical correlation analysis (CCA) and sum-constrained CCA [2]. Finally, all analysis methods were applied on the wavelet-resampled fMRI data to determine the null distribution of r, and the correlation values at 99.9 percentile, defined as r_p, were computed for each method, which indicate the cut-off value at the significance level p = 0.001.
Results:
The proposed DNN method has comparable correlation magnitude with GLM in the inactive regions, but it provides better discrepancy between inactive regions and active regions. The mean AUC values in the false positive rate <=0.1 range (max value 0.1) are 0.035, 0.064, 0.066, and 0.070 for GLM, CCA, sumCCA, and DNN respectively. The 2D histograms of the correlation map were shown in Figure 1, with the color representing the number of voxels in each bin. The proposed DNN model restricted voxels of low correlations in GLM from achieving higher correlation, indicating that DNN does not lead to noticeable over-estimated activation for inactive voxels, in contrast to Gaussian smoothing and CCA variants. The DNN model did not show exaggerated correlation for inactive voxels based on r_p values. By overlaying the gray matter tissue mask (blue) and activation map above the cut-off value r_p, DNN alleviated blurring artifact compared to traditional spatial smoothing approaches (Figure 2).
·2D histograms of correlations obtained from GLM (x-axis) and other methods (y-axis).
·Active map with the cut-off value r_p from a single subject. Gray matter mask is overlaid in the image as marked in blue.
Conclusions:
The proposed method showed its advantage of adaptively smoothing task fMRI data for more robust brain activation detection, which might be beneficial for basic research (e.g. layer fMRI data without manually delineating cortical layers) and clinical application (e.g. task fMRI for surgical planning [3]).
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 2
Methods Development 1
Keywords:
Data analysis
FUNCTIONAL MRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Zhuang, X., Z. Yang, and D. Cordes, A technical review of canonical correlation analysis for neuroscience applications. Human Brain Mapping, 2020. 41(13): p. 3807-3833.
3. Silva, M.A., et al., Challenges and techniques for presurgical brain mapping with functional MRI. NeuroImage: Clinical, 2018. 17: p. 794-803.