Reconstruction of digit images from human brain fMRI activity through connectivity informed Bayesian networks

Newly emerged developments in decoding of stimulus images from fMRI measurements have shown promising results. Decoding-classification has been the main concern of decoding studies, whereas the matter of reconstruction (decoding) of stimulus images from fMRI data, especially natural images, lacks adequate examination and it requires plenty of efforts to improve.

The present study employs Bayesian networks for decoding-reconstruction which is a novel application of this tool. Moreover, as a novel approach, we exploit the brain connectivity information in decoding-reconstruction procedure through Bayesian networks.

The proposed method was applied to reconstruct 100 images of digits 6 and 9 from the fMRI measurements obtained when showing some handwritten images of 6 and 9 to the subject. The information of only 10 brain voxels were exploited and an average (standard deviation) city-block distance error of 0.1071(0.0134) was obtained for all stimuli's reconstruction. In comparison with current common methods: The results reveal that Bayesian networks are successful in decoding-reconstruction of handwritten digits and inclusion of brain connectivity information makes them perform even more efficiently and improves decoding-reconstruction as well (reducing average error by almost 5%).

In the task of decoding-reconstruction, the models including brain connectivity appear significantly superior to other existing models.