20 likes | 220 Views
Automatic Detection of ADHD subjects using Deep Convolutional Neural Network Arjun Watane , Soumyabrata Dey (arjunwatane@knights.ucf.edu, soumyabrata.dey@gmail.com) University of Central Florida. III. Formulation:. Slices. Problem & Motivation: Automatic detection of ADHD Structural MRI
E N D
Automatic Detection of ADHD subjects using Deep Convolutional Neural Network • ArjunWatane, SoumyabrataDey(arjunwatane@knights.ucf.edu, soumyabrata.dey@gmail.com) University of Central Florida • III. Formulation: Slices • Problem & Motivation: • Automatic detection of ADHD • Structural MRI • Strict 3-D anatomical structure • Lack of biological measures for diagnosis • Subjective to Verbal Test • Inconsistent and over-diagnosis problem • Data set: • NYU data center of ADHD-200 • 203 training, 41 test subjects Slices Slice 1 Convolutional Neural Network Extraction of Feature Layer FC6 and FC7 Slice 2 Weighted Late Fusion Final decision Decision vector Weight vector Combine FC6 and FC7 for each slice Vector Length = 4096x2 = 8192 Slice 1 Decision : Calculated from training data Slice n Slice 2 Decision Slices are ranked based on the score. Slice1 has highest weight SVM Classifier 1 Slice n Decision 0 • V. Visualization of Features : • IV. Image Pre-Processing : • II. Convolutional Neural Network : Brain Segmentation Convolution 1 Convolution 2 Convolution 3 Convolution 4 Convolution 5 Gray Matter • VI. Results : FC6 FC7 2048 2048 FC6 FC7 • Network Configurations • Input Blob – 203 Subjects, 21 slices per subject, 256x256 pixels slice image • Layer – convolution and max-pooling to generate feature map • 5 convolution layers, 4 max pooling layers, 2 fully connected layers (FC) • Extraction of features using pretrainedImagenet model FC6+FC7 White Matter • Late Fusion of FC6 and FC7 features showed the highest accuracy of ADHD classification, at 80.49%. Normalized Accuracy Comparison of Independent Feature vs. Feature Combination