Simulating functional brain images in Alzheimer's disease

The Small Sample Size problem is a common issue in neuroimaging, where the number of features frequently surpass the number of samples, making generalization a difficult task and leading to inconsistent results. To overcome this problem, we have developed a new algorithm intended to generate a new set of functional images with similar characteristics to an original one, in our case, a subset of the PET ADNI database. This way, a bigger dataset can be used to evaluate our methodology, providing better generalizable results. The procedure involves modeling the original database with PCA in order to project the images to the eigenbrain space, and then modelling the coordinate set of the original images in the eigenbrain space using KDE. Once these parameters are obtained, new coordinates can be generated using the density functions of the original database and a random number generator. The generated dataset is then evaluated to show that the features of the new images match the original ones while reducing the variance.