Fig 1.
Summary of the learning pipeline.
1) Each image in the datasets is preprocessed (see section Preprocessing and Fig 2), reducing the dimensions from about 100,000,000 (79 × 95 × 68 × 200) to about 500,000. 2) The MHPC system then extracts the 3D-HOG features of each image reducing the number of dimensions to about 100,000; see section Histogram of oriented gradients (HOG) features. 3) The last step tries to select the best learner (from the initial set of base learners) and feature set, based on running 5-fold cross validation over the training set, using different combinations of the number of features and base learners. This step reduces the number of dimensions to a number under 1000; see section Results. HOG feature extraction, minimum redundancy maximum relevance (MRMR) feature selection and base learner selection are all parts of the MHPC algorithm (shown in the red box above). See Algorithm 1 for details. This figure is best viewed in color.
Table 1.
Details of Training Dataset participants, ADHD-200.
Table 2.
Details of Holdout Dataset participants, ADHD-200.
Table 3.
Details of Training Dataset participants, ABIDE.
Table 4.
Details of Holdout Dataset participants, ABIDE.
Fig 2.
The preprocessing pipeline for functional and structural magnetic resonance images is summarized in the figure. Orange shapes in the image show the steps of preprocessing necessary for both fMRI and structural MRI scans. Green shapes show the preprocessing steps only needed for fMRI scans. This figure is best viewed in color.
Fig 3.
Left and right panels show HOG bins in 2D and 3D space, respectively.
Fig 4.
Gradient vector of a sample pixel.
For illustration, we describe the 2D HOG feature computation process. Here, we consider a single pixel, the one shown in red, whose neighbors have intensities 56, 93, 94, and 55. The blue arrow is the sample gradient, computed as described below. (This figure is best viewed in color.)
Fig 5.
Input and output of 2D HOG on a brain image.
The left panel shows an axial slice of a structural MR image of a brain. The right panel shows the HOG features of the same slice. Here, we represent the HOG features by an 8-sided “star”, where the length of each arm is the size of the histogram in that direction. This representation is generated using VLFeat [32].
Fig 6.
5-fold cross validation accuracies on the training set.
The accuracies are obtained using RBF SVM (with various sigma values), on the training portion of the ADHD-200 dataset using functional images plus personal characteristic data. This figure is best viewed in color.
Table 5.
ADHD-200, functional images.
Table 6.
ABIDE, functional images.
Table 7.
ADHD-200, structural images.
Table 8.
ABIDE, structural images.
Table 9.
ADHD-200, personal characteristic data.
Table 10.
ABIDE, personal characteristic data.
Fig 7.
Summary of ADHD-200 dataset classification results.
The black horizontal dotted line shows the baseline chance accuracy of the test set. Each vertical bar shows the mean and range of the cross validation results for the selected base learner (L) and feature set (FS*(L)) on the training set, as produced with MHPC (Algorithm 1). The blue asterisks * show the accuracy of each classifier on the hold-out set. The classifiers on the x-axis are ordered by the types of features they used, including various combinations of structural MRI, functional MRI, and personal characteristic data. The legend also identifies the actual classifier used. This figure is best viewed in color.
Table 11.
ADHD-200, personal characteristic data with functional images.
Table 12.
ADHD-200, personal characteristic data with structural images.
Table 13.
ABIDE, personal characteristic data with functional images.
Fig 8.
Summary of ABIDE dataset classification results.
Conventions are the same as for Fig 7. This figure is best viewed in color.
Table 14.
ABIDE, personal characteristic data with structural images.
Table 15.
ADHD-200 results summary.
Table 16.
ABIDE results summary.
Fig 9.
Extracted block regions from ADHD-200 structural MRI dataset.
See text for details.
Fig 10.
Extracted block regions from ABIDE fMRI dataset.
See text for details.