Excess height versus cluster extent in subtraction SPECT

Subtraction of ictal and interictal SPECT perfusion images of the brain has the potential of locating the epileptogenic region. This region generally shows large differences between both images. However, differences can also be induced by noise in the projection data. We hypothesized that the extent, besides the intensity, of observed clusters of voxels in thresholded subtraction images, is an important parameter in the classification of clusters into real perfusion differences and noise induced differences. To test this hypothesis, we performed a number of simulation experiments. Using a Monte Carlo approach, we constructed cumulative distribution functions (CDF) of the excess height (i.e. the largest difference in a cluster) and the cluster extent under the condition of no perfusion change (i.e. only noise induced clusters). The reproducibility of the CDF curves was shown using measured patient data. Furthermore, a 3D brain software phantom experiment was used to examine the detection and classification of an induced region of hyperperfusion. We assigned to every observed cluster a probability, derived from the CDF curves, for excess height and extent. For different probability thresholds, sensitivity and specificity of the detection of the induced hyperperfusion based on its probability for excess height and cluster extent was measured. These measurements were combined in receiver operating characteristic (ROC) curves. These ROC curves showed a better performance when using classification based on cluster extent. We conclude that the cluster extent is an important parameter in the characterization of clusters in thresholded subtraction of perfusion SPECT images of the brain.