Neural Decoding Technique to Reconstruct Stimulus from the Evoked fMRI Voxel Responses

In cognitive neuroscience, neural decoding is a technique to reconstruct the stimulus that evoked the neural activity in the brain. The stimulus includes graphical characters, face images, handwritten characters and natural images. The functional Magnetic Resonance Imaging (fMRI) detects the neural activity caused by the varying Blood-Oxygenated-Level-Dependent (BOLD) signal. The proposed technique aims to develop a novel framework for visual image reconstruction of natural images. The exact reconstruction of natural images is challenging due to its complexity. To achieve this accuracy the voxels describing the structural information and semantic information are considered from the lower and higher brain areas respectively. The voxels having the significant information of pixels are then selected. The reconstruction is achieved using both encoding and decoding models. The encoding model is designed using the image prior (dataset consisting of natural images). The model describes how the stimulus features are encoded in the brain activity. The image prior is clustered using multiclass Support Vector Machine classifier by extracting the Gabor features of the images followed by the k-means clustering. Using the Euclidean distance calculation, the maximum probable category is chosen using which the encoding model is designed for further reconstruction. At the decoding end, a hybrid Bayesian framework is formed by gating the clustered image priors and the selected voxels to reconstruct the stimulus from the evoked fMRI voxel responses. The proposed work focuses on developing an accurate, less complex and automatic software technique for visual image reconstruction of natural images.