Mutual Information-Based Feature Facet Clustering with Ensemble SVM for fMRI Visual Object Decoding

Abstract

Deciphering visual stimuli from fMRI data presents a significant challenge in computational neuroscience. This paper introduces a novel, optimized ensemble learning framework for high-accuracy visual object recognition. Our method employs a mutual information-based hierarchical clustering technique to automatically segment the high-dimensional voxel space into independent feature facets. An ensemble of Support Vector Machine (SVM) classifiers is then trained on these facets. Crucially, the entire framework—including the number of facets, the fusion operator, and SVM parameters (C, gamma)—is globally optimized using the Simulated Annealing algorithm to ensure peak performance. We rigorously evaluated our approach on three public fMRI datasets: DS105 (8 visual objects), DS107 (4 semantic categories), and DS116 (2 visual oddball stimuli). The proposed model demonstrated exceptional performance, achieving mean recognition accuracies above 95% across all three datasets, with peak subject-level accuracy reaching 100%. Specifically, our Ensemble-965 model (using the detailed Talairach Atlas) attained accuracies of 98.6% on DS105, 97.5% on DS107, and 99.4% on DS116, surpassing current state-of-the-art brain decoding methods under comparable validation conditions. These results indicate that our method provides a robust, accurate, and highly effective solution for visual brain decoding.