Preliminary evaluation of texture analysis on myocardial perfusion SPECT: a digital cardiac phantom study

Background: Texture analysis (TA) for evaluating the tumor spatial heterogeneity has been used to provide diagnostic and prognostic values of oncologic positron emission tomography, X-ray CT, and magnetic resonance images. The feasibility of TA for myocardial perfusion single-photon emission computed tomography (SPECT) study has remained to be clarified. The purpose of this study was to evaluate TA in myocardial perfusion SPECT (MPS) using a digital cardiac phantom.

Methods: A four-dimensional extended cardiac torso phantom (Duke University, Durham, USA) was used. One normal cardiac phantom and 16 phantoms with various sizes of defects at the anterior wall were generated. We visually classified ^99mTc myocardial perfusion images into 6 patterns from normal to complete defects. All cardiac phantoms were electrocardiographic- and respiratory-gated models, and the left ventricular ejection fraction and end-systolic volume were constantly 61% and 132 mL, respectively. The LIFEx freeware (version 4.09; www.lifexsoft.org) was used to perform TA with 31 textural features. A cardiac volume of interest (VOI) was generated with the normal perfusion model, and a VOI size was set using the thresholds of 30, 35, 40, 45, 50, 55, and 60% of the maximum count. SPECT counts were normalized and discretized into 5 grey levels (GL): 4, 8, 16, 32, and 64. The skewness and kurtosis of histogram were calculated with these GL.

Results: Cardiac VOI sizes changed from 101 mL for 60% threshold to 320 mL for 30% threshold. The mean skewness and kurtosis values were 0.45 ± 0.036 and 2.51 ± 0.15, respectively. The dendrograms for cluster analysis using the textural features automatically classified 6 perfusion patterns. Agreements between classifications obtained with the cluster analysis and visual assessment changed in the various combinations of VOI thresholds and GL. Kappa values between the cluster analysis and visual assessment in 4, 8, 16, 32, and 64 GL were 0.691−0.849, 0.687−0.924, 0.623−1.000, 0.464−0.778, and 0.419−0.773, respectively. Perfect agreement (kappa = 1.00) was determined in the 40 % VOI threshold and 16 GL.

Conclusions: TA is feasible in MPS and showed an excellent agreement with a visual assessment by using suitable VOI size and GL. The selection of VOI size and GL were important factors to perform TA in MPS study.