Improved Cluster-Wise Inference in Functional Magnetic Resonance Imaging via Geometric Redefinition of Cluster Size

Abstract

Accurate identification of significant activations in functional Magnetic Resonance Imaging (fMRI) is essential for reliable neuroimaging research. Cluster-wise inference based on Random Field Theory (RFT) has been widely adopted for over two decades due to its superior sensitivity compared with voxel-wise corrections. However, recent studies have revealed that conventional RFT-based approaches may yield inflated false-positive rates. This study proposes an improved cluster-wise inference method by redefining cluster size from a geometric perspective. Specifically, we calculate cluster size using intrinsic volume rather than voxel count, and we investigate the influence of expected cluster size estimation and voxel connectivity. Analyses use publicly available first-level images from 198 subjects and matched simulations; factors cross Cluster-Defining Threshold (CDT) p = 0.001/0.01, one- vs two-sample t tests (two group sizes), and Euler Characteristic (EC) implementations (Statistical Parametric Mapping (SPM) default vs corrected), under 6/18-connectivity. On Gaussian-null simulations (32 configurations), the geometric definition achieved FWER at or near 0.05 across all settings and was consistently lower than voxel-counting. On real data (128 configurations spanning four paradigms), it reduced FWER relative to voxel-counting. Sensitivity was assessed only on simulations with five planted clusters over SNR = 1−4, where detection rates were comparable to the voxel-count approach. These findings highlight the value of geometric definitions in enhancing statistical inference for multi-voxel activation patterns in fMRI. Overall, redefining cluster extent by intrinsic volume improved family-wise error rate control while maintaining sensitivity, providing a simple drop-in change to cluster-wise RFT workflows.