Comparison of classification performance achieved by our image features and that by the voxel features used in prior work. The performance on the six tasks are compared, and nine data sets are used for each task. For each task on a specific data set, the entire data set is randomly partitioned so that 2/3 of the data is in the training set and the rest 1/3 is in the test set. A total of 30 random partitions are generated, and the average performance is reported. The numbers of genes used for different tasks are given in Table 1.