Automatic extraction of TACs from dynamic mouse FDG microPET images based on improved spectral clustering

Objectives Accurate time activity curves (TACs) of different cardiac tissue components (left, right ventricle and myocardium) play a vital role in the estimation of the input function and cardiac quantification. We presented and validated a method to automatically extract TACs with high accuracy from dynamic mouse FDG microPET images.

Methods Regarding the activities of each frame as the features of the voxels, the initial three segmented components and the corresponding TACs were firstly obtained by using spectral clustering, a method that performed the clustering in a low-dimensional space of the Laplacian matrix derived by the high-dimensional affinity matrix encoding the pairwise similarity of voxels. With intent to eliminate the randomness of the spectral clustering and obtain more localized results, we then iteratively updated the TACs starting with the previously acquired initial TACs. In each iteration, we computed the three Pearson correlations between the TAC of a particular voxel with the corresponding three initial TACs, and the voxel was assigned to the corresponding component with the largest correlation value. The new TACs were obtained by averaging the TACs of the three components, and then were used as the initial values for the next iteration. The iteration was repeated until the segmentation results did not change, and the final TACs were acquired. We validated this method on fourteen C57BL/6 mice data sets randomly chosen from the Mouse Quantitation Program database shared by UCLA.

Results The root mean square errors between TACs extracted by manual delineation and our method are 2.1%±1.7%, 3.5%±2.8% and 2.4%±1.6% for Lv, Rv and Myo respectively. Additionally, compared to classic spectral clustering, the proposed method gave more localized components. And the introduction of the iterative correlation minimized the effect of the noise and reduced the variability.

Conclusions The proposed method can extract the TACs of cardiac tissue components automatically with high accuracy and reproducibility.

Research Support The National Natural Science Foundation of China, The Special Project of National Major Scientific Instrument and Equipment Development under Grant 81227901 the National Basic Research Program of China (973 Program) under Grant 2011CB707700