Clustered pixels analysis for functional MRI activation studies of the human brain

Abstract Conventional t ‐statistics and cross‐correlation coefficients are commonly used for analysis of functional magnetic resonance images. The sensitivity of these statistics is usually low because severe Bonferroni‐type corrections are required for multiple statistical comparisons to minimize the false‐positive error. In the human brain, most functional areas are larger in size than a single image pixel, and coactivation of numerous contiguous pixels is expected. The probability of occurrence of clusters due to random noise is small and can be modeled. Cluster size and intensity thresholding can be used to assess statistical significance. Previous cluster analysis strategies used Gaussian models, working best with low spatial resolution images (e.g., positron emission tomography). We present a new cluster analysis model applicable to data with little or even no covariance between adjacent pixels. Computer simulations and phantom experiments were used to verify this strategy. Our new method is substantially more sensitive than both the conventional intensity‐only thresholding (IOT) method and the previous cluster method for signal change less than 6%, with maximum significant enhancement in sensitivity of 12.8 and 3.8 times, respectively. The results obtained from normal volunteers with visual stimulation further confirm the effectiveness of our new approach and show an average increase in detected activation area of 3.1 times over the IOT method and of 1.6 times over the previous cluster method using the new approach. ©1996 Wiley‐Liss, Inc.