A robust segmentation method with triple-factor non-negative matrix factorization for myocardial blood flow quantification from dynamic ⁸² Rb positron emission tomography

In this work, we proposed a triple-factor non-negative matrix factorization (TNMF) method to semiautomatically segment the regions of interest (ROIs) of the left ventricular (LV) cavity and myocardium to improve the reproducibility of myocardial blood flow (MBF) quantification from dynamic ⁸² Rb positron emission tomography (PET).

The proposed TNMF method was evaluated using NCAT phantom simulation with three noise levels. The segmented ROIs, time-activity curves (TACs), and K₁ derived from the TNMF method were compared with the ground truth simulated. The TNMF method was further evaluated in two patients each undergone both rest and stress ⁸² Rb PET studies. The TNMF and manual segmentations were implemented by two different observers, and the interoperator variations of MBF and myocardial flow reserve (MFR) were compared between the two methods.

Our simulation results showed that the TNMF method for dynamic PET image segmentation was robust as evidenced by the high Dice similarity coefficient, regardless of high or low count level. The relative bias in K₁ estimation was less than 1%. Our patient results also showed that reasonable ROIs for the LV cavity and myocardium could be obtained precisely for patients with and without myocardial perfusion defects. The TACs derived from the TNMF method were highly correlated with those obtained with the manual method (R²  ≥ 0.964). The interoperator variations of MBF and MFR were markedly reduced using the TNMF method.

In conclusion, the TNMF method is highly feasible for semiautomatic segmentation of the LV cavity and myocardium, with the potential to improve the precision of MBF quantification by improving segmentation.