Artificial neural network in classification of traumatic brain injury using rs-fMRI and PET imaging
Poster No:
1384
Submission Type:
Abstract Submission
Authors:
Faezeh Vedaei1, Najmeh Mashhadi2, Mahdi Alizadeh3, George Zabrecky3, Daniel A. Monti3, Nancy Wintering3, Andrew B. Newberg3, Feroze B. Mohamed3
Institutions:
1University of Pennsylvania, Philadelphia, PA, 2University of California-Santa Cruz, Santa Cruz, CA, 3Thomas Jefferson University, Philadelphia, PA
First Author:
Co-Author(s):
Introduction:
Mild traumatic brain injury (mTBI) is a public health concern that may adversely affects person's quality of life, thinking, memory and behavior. Most of the complications of mTBI are brain function alterations that impact on cognitive performance and in a long-term may lead to neurodegenerative diseases.1 Functional brain imaging including resting-state functional magnetic resonance imaging (rs-fMRI) and positron emission tomography (PET) have been used to detect brain function abnormalities in brain disorders.2,3 While statistical group level analysis can provide spatial patterns of brain function at population level, these methods are not able to provide imaging signature of brain disorders that can be applied at single individual level. Machine learning (ML) algorithms including artificial neural network (ANN) have recently gained more popularity in medical image analysis due to their capability to learn complex representations of data.4 We aimed to develop automatic classification model using ANN and rs-fMRI and PET imaging to distinguish between patients at chronic stage of mTBI from healthy controls (HCs).
Methods:
rs-fMRI and PET were acquired from 83 patients with mTBI and 40 HCs. Voxel-wise brain maps of rs-fMRI metrics including fractional amplitude of low frequency fluctuation (fALFF), degree centrality (DC), regional homogeneity (ReHo), functional connectivity strength (FCS), and voxel-mirrored homotopic connectivity (VMHC) generated for each subject using DPARSF_V5.0 after preprocessing steps. Also, PET data was processed using PETPVE12 toolbox running on MATLAB. ANN architecture was developed using autoencoder, several hidden layers of rectified linear unit (ReLU) function and the last layer of sigmoid function. Input features were generated applying Automated Anatomical Labeling (AAL) atlas and extracting the mean values of each measurement from 116 region of interest (ROI). The classification model was developed for single modality and multimodality measurements.5 The performance of the models was estimated via 5-fold cross validation (CV) using the receiver operator characteristic (ROC) curve analysis. Also, top 10 ROIs with the most contribution to the model prediction were extracted for each metric.6,7
Results:
Classification performance analysis showed high scores for each single modality. However, the accuracy of classification improved using multimodality model that achieved the highest using multimodality of rs-fMRI and PET measurements (Table 1). The selected top 10 important features for all the metrics were among several main brain functional network including default mode network (DMN), sensorimotor network, visual cortex, cerebellum, and limbic system (Figure 1).
Conclusions:
In the need to robust biomarkers discriminating patients with mTBI from HCs, our study for the first developed automatic classification model using deep leering (DL)-based architectures. We employed autoencoder to extract the extract hidden representations using unsupervised approach. We showed that each single modality can capture specific information from neurophysiology of mTBI providing comprehensive approach to study of brain function alteration in brain disorders including mTBI. Also, considering the complicated pathologic process of mTBI, multiple neuroimaging metrics may provide complementary information and can be investigated via multimodality classification models.
Cognitive dysfunction that is accompanied with mTBI is linked with brain function alteration in specific brain areas. We have shown that brain regions located in the frontal, temporal, and parietal lobes, as well as brain regions located in the limbic system and visual cortex as well as cerebellum are among the most important features have the most contribution in model prediction.8-10 These results suggest that DL-based classifiers might be extended to quantitative imaging biomarker providing a new avenue for prediction of individual patients in the clinical settings.
Cognitive dysfunction that is accompanied with mTBI is linked with brain function alteration in specific brain areas. We have shown that brain regions located in the frontal, temporal, and parietal lobes, as well as brain regions located in the limbic system and visual cortex as well as cerebellum are among the most important features have the most contribution in model prediction.8-10 These results suggest that DL-based classifiers might be extended to quantitative imaging biomarker providing a new avenue for prediction of individual patients in the clinical settings.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural) 2
PET Modeling and Analysis
Keywords:
Other - Artificial neural network, resting functional MRI, positron emission tomography, mild traumatic brain injury, classification
1|2Indicates the priority used for review
·Table1, Results
·Figure1, Results
Provide references using author date format
2- Byrnes, K. R. et al. FDG-PET imaging in mild traumatic brain injury: A critical review. Front. Neuroenergetics 6, (2014).
3- Maldjian, J. A. Applications of rs-fMRI to Traumatic Brain Injury. 27, 685–696 (2018).
4- Yin, W., Li, L. & Wu, F. X. Deep learning for brain disorder diagnosis based on fMRI images. Neurocomputing 469, 332–345 (2022).
5- Zhou, T., Thung, K. H., Zhu, X. & Shen, D. Effective feature learning and fusion of multimodality data using stage-wise deep neural network for dementia diagnosis. Hum. Brain Mapp. 40, 1001–1016 (2019).
6- Hao, X. et al. Multi-modal neuroimaging feature selection with consistent metric constraint for diagnosis of Alzheimer’s disease. Med. Image Anal. 60, 101625 (2020).
7- Zhang, F. et al. Multi-modal deep learning model for auxiliary diagnosis of Alzheimer’s disease. Neurocomputing 361, 185–195 (2019).
8- Konstantinou, N., Pettemeridou, E., Stamatakis, E. A., Seimenis, I. & Constantinidou, F. Altered resting functional connectivity is related to cognitive outcome in males with moderate-severe traumatic brain injury. Front. Neurol. 9, (2018).
9- Schwedt, T. J., Chong, C. D., Peplinski, J., Ross, K. & Berisha, V. Persistent post-traumatic headache vs. migraine: an MRI study demonstrating differences in brain structure. J. Headache Pain 18, 1–8 (2017).
10- Stevens, M. C. et al. Multiple resting state network functional connectivity abnormalities in mild traumatic brain injury. Brain Imaging Behav. 6, 293–318 (2012).