Usage of Different Brain Atlases on Comprehensive Training Procedure Using Basic MLP Structure
Poster No:
1456
Submission Type:
Abstract Submission
Authors:
Jo-Hsin Shih1, Po-Hsien Lee2, Zhitong John Wang3, Changwei Wu3, Ai-Ling Hsu4
Institutions:
1Department of Artificial Intelligence, Chang Gung University, Taoyuan, Taiwan, 2School of Medicine, Taipei Medical University, Taipei, Taiwan, 3Graduate Institute of Mind, Brain and Consciousness, Taipei Medical University, New Taipei, Taiwan, 4Department of Psychiatry, Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan
First Author:
Co-Author(s):
Zhitong John Wang
Graduate Institute of Mind, Brain and Consciousness, Taipei Medical University
New Taipei, Taiwan
Graduate Institute of Mind, Brain and Consciousness, Taipei Medical University
New Taipei, Taiwan
Changwei Wu
Graduate Institute of Mind, Brain and Consciousness, Taipei Medical University
New Taipei, Taiwan
Graduate Institute of Mind, Brain and Consciousness, Taipei Medical University
New Taipei, Taiwan
Introduction:
Major depressive disorder (MDD) is a prevalent and severe mental illness, posing a high risk of disability worldwide (Whiteford et al., 2013). Characterized by impairments in mood and cognitive functioning, MDD has been associated with malfunctions of brain networks, as measured by resting-state functional magnetic resonance imaging (rs-fMRI) (Kaiser et al., 2015). The growing availability of large-scale rs-fMRI datasets focusing on MDD has enabled the development of various deep-learning models in differentiating MDD patients from controls (Qin et al., 2022; Venkatapathy et al., 2023). Among these models, the mandatory selection of a parcellation atlas for feature extraction may affect the resulting accuracy; however, there is a lack of studies investigating its impact on model performances. In this work, we tested the hypothesis that models trained with the rs-fMRI parcellated atlas would achieve the highest performance as compared to those trained with other atlases generated from structural or task-based origins.
Methods:
The rs-fMRI scans of 498 subjects (256 MDD patients vs. 242 controls) from the Rest-meta-MDD consortium (Chen, X. et al., 2022) were analyzed to assess the impact of using commonly adopted whole-brain atlases on model accuracy. Given the inhomogeneous confounding factors in the consortium collected from 25 sites, data from a single site (site 20) of 533 subjects, all acquired on a 3T scanner (Siemens Tim Trio), were specifically chosen. Additional inclusion criteria included (1) preprocessed data parcellated using four atlases, i.e., AAL (Tzourio-Mazoyer et al., 2002), Dosenbach (Dosenbach et al., 2010), Craddock (Craddock et al., 2012), and Power (Power et al., 2011) atlases; and (2) subjects aged between 21 to 70 years. Next, the functional connectivities of each atlas were calculated using Pearson correlation.
Considering that the complexity of deep-learning architectures often determines its learning capability, models were trained using the simple two-layer architecture of multi-layer perceptron with 32 and 16 neurons in each layer. To prevent information leakage that potentially inflates model accuracy, the models were trained on 80% of the data using BRAPH (Mite et. al 2017). The parameters were validated on 10% of the data, and the accuracy was tested on the remaining 10% of the data. Additionally, the settings for model training included the use of Adam optimizer, L2 regularization, a dropout rate of zero, a batch size of 32, 20 epochs, and a loss function of cross entropy. Moreover, models were repeatedly trained using a 30-time resample approach to statistically assess the impact of atlas utilization on model accuracy. This assessment was conducted using an ANOVA and post-hoc tests with a p-value < 0.05 being considered statistically significant.
Considering that the complexity of deep-learning architectures often determines its learning capability, models were trained using the simple two-layer architecture of multi-layer perceptron with 32 and 16 neurons in each layer. To prevent information leakage that potentially inflates model accuracy, the models were trained on 80% of the data using BRAPH (Mite et. al 2017). The parameters were validated on 10% of the data, and the accuracy was tested on the remaining 10% of the data. Additionally, the settings for model training included the use of Adam optimizer, L2 regularization, a dropout rate of zero, a batch size of 32, 20 epochs, and a loss function of cross entropy. Moreover, models were repeatedly trained using a 30-time resample approach to statistically assess the impact of atlas utilization on model accuracy. This assessment was conducted using an ANOVA and post-hoc tests with a p-value < 0.05 being considered statistically significant.
Results:
During training, every model achieved an accuracy over 80% (in the training dataset) within 5 epochs and reached 100% in 20 epochs. Figure 1 shows the model accuracies on the testing dataset, presenting the mean and standard error of the mean across 30 repetitions. Moreover, an ANOVA test showed a significant difference in accuracy among models trained with data parcellated by 4 atlases at F(3, 116) = 2.93, p < 0.036. Post-hoc Tukey HSD tests were conducted for pairwise comparisons. Notably, the accuracy between the model trained by Craddock and Power atlases was found to be significantly different (p < 0.05).
·Distributions of classification accuracy across four different atlases
Conclusions:
This study investigated the effect of data parcellation using four different atlases on the accuracies of deep-learning models. By employing a 30-time resampling method on the dataset, we demonstrated that the data parcellated using rs-fMRI-based atlas (Craddock) resulted in a higher model accuracy compared to data parcellated using task-fMRI-based atlas (Power).
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Keywords:
FUNCTIONAL MRI
Psychiatric Disorders
Other - Deep Learning, Major depressive disorder, Classification
1|2Indicates the priority used for review
Provide references using author date format
Kaiser, R. H. et al. (2015), 'Large-scale network dysfunction in major depressive disorder: a meta-analysis of resting-state functional connectivity', vol. 72, pp. 603-611.
Qin, K. et al. (2022), ' Using graph convolutional network to characterize individuals with major depressive disorder across multiple imaging sites ', EBioMedicine, vol. 78
Venkatapathy, S. et al. (2023), 'Ensemble graph neural network model for classification of major depressive disorder using whole-brain functional connectivity', Frontiers in Psychiatry, vol. 14
Chen, X. et al. (2002), 'The DIRECT consortium and the REST-meta-MDD project: towards neuroimaging biomarkers of major depressive disorder', Psychoradiology, vol. 2, pp. 32-42.
Tzourio-Mazoyer. et al. (2002), 'Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain', Neuroimage, vol. 15, pp. 273-289.
Craddock, R. C. et al. (2012), 'A whole brain fMRI atlas generated via spatially constrained spectral clustering', Human brain mapping, vol. 33, pp.1914-1928.
Dosenbach, N. U. et al. (2010), 'Prediction of individual brain maturity using fMRI', Science, vol. 329, pp. 1358-1361.
Power, J. D. et al. (2011), 'Functional network organization of the human brain', Neuron, vol. 72, pp. 665-678.
Mijalkov, M. et al. (2017), 'BRAPH: a graph theory software for the analysis of brain connectivity', PloS one, vol. 12