Skip to Main Content
Dynamic cardiac SPECT imaging can provide quantitative and possibly even absolute measures of physiological parameters. However, a dynamic cardiac SPECT study involves a number of steps to obtain estimates of physiological parameters of interest. One of the key steps involves the selection of regions of interest. In the past, this has been done manually or by using a semi-automatic method. We propose to use cluster analysis to segment the data to obtain improved parameter estimates. The algorithm consists of two steps: a standard k-means approach followed by a fine-tuning procedure using fuzzy k-means. Computer simulations were used to test the algorithm and to compute bias in kinetic rate parameters with and without the use of fine-tuning. This was followed by performing four canine (two teboroxime and two Tl-201) and two patient (one teboroxime and one Tl-201) dynamic cardiac PET-fusion SPECT studies. The short-axis slice image data were used as input for the cluster analysis program as well as for the semi-automatic method. The blood input function was obtained from the cluster analysis algorithm as one of the clusters. The pixel-wise time-activity curves were computed from the fine tuning step as a linear combination of the standard k-means cluster data. All the TACs were fit to a two Compartment model. Parametric images of the wash-in rate parameter were obtained and compared to those obtained with the semiautomatic method. Microsphere derived flows were used as a gold standard in the canine studies for comparison. Our results suggest that cluster analysis provides parameter estimates comparable to the semi-automatic method. Moreover, the clustered curves have less noise and yield reasonable fits where with the semi-automatic method the fitting routine sometimes failed to converge. It was also found that the use of clustering required less manual intervention than the semi-automatic method. The use of clustering may help to bring dynamic cardiac SPECT closer to clinical feasibility.