Automatic Identification of ADHD Subjects using functional Magnetic Resonance Imaging(fMRI) Data
| Paper: | Exploiting the brain’s network structure in identifying ADHD subjects |
| Contact: | Soumyabrata Dey, A. Ravishankar Rao, and Mubarak Shah |
Problem Definition and Overview
Attention Deficit Hyperactive Disorder (ADHD)is a common brain disorder found among children. In spite of all the efforts made for the studies of ADHD, the root cause of this problem is still unknown to the community. Till date no well-known biological measure exists that can diagnose it. The goal of this project is to propose novel frameworks for automatic detection of ADHD subjects using their functional Magnetic Resonance Imaging(fMRI) data. Also, we identified brain regions which contributed most in segregating ADHD subjects from normal subjects. We believe that the findings will be helpful to the community for better understanding of the problem. We have used the data set, released for the ADHD-200 competion, for the experimental validation of our proposed methods. The summery of the data set is described in Table 1.
Our algorithm exploits the topological differences between the functional networks of the ADHD and controlled brains. Different steps of our approach are described in the below figure. The input to our algorithm is brain fMRI sequences of the subjects. fMRI data can be viewed as a 4-D video such that the 3-D volume of the brain is divided into small voxels and imaged for a certain duration . The data can also be viewed as a time series of intensity values for each of the voxels. The correlation of these intensity time-series can be an indication of how synchronous the activities of two voxels are, and higher correlation values suggest that two voxels are working in synchronization. A functional network structure is generated for the brain of each of the subjects under study by computing the correlations for all possible pairs of voxels and establishing a connections between any pairs of voxels if their correlation value is sufficiently high. Different network features, such as degree maps, cycle maps and weight maps are computed from the network to capture topological differences between ADHD and control subjects. A brain mask is computed that includes only the regions with useful information to classify ADHD and control subjects. We refer this mask as ’useful region mask’. Finally, the network features from the voxels selected in the useful region mask are extracted to train a PCA-LDA based classifier. Figure 1 provides the flowchart of our algorithm.
The experiments are perfomed by training a PCA-LDA based classifier on the whole training set and then verify the classifier’s performance on the test sets of 6 data centers. Training and testing are performed using useful region mask and without using the mask. The results shows that the use of the mask produces better classification accuracies. The results are shown in the below table.
The below figure showing useful region mask detected using the subjects of KKI training data set. The useful regions are marked in orange on different slices of the brain.
The below figure showing the regions with average degree differences between ADHD and control subject groups. The figure is constructed using the subjects of KKI training data set. The dark red to white color map is used to represent higher degree of control subjects and blue to green color map is used to show the opposite.
Soumyabrata Dey, Ravishankar Rao and Mubarak Shah, Exploiting the Brain’s Network Structure in Identifying ADHD Subjects, Frontiers in System Neuroscience, November 2012.
Berkan Solmaz, Soumyabrata Dey, Ravishankar Rao and Mubarak Shah, ADHD Classification Using Bag of Words Approach on Network Features, SPIE Medical Imaging Conference, February 2012.
ADHD-200 data set can be found here. [DATA]