Abstract
512
Objectives: Radiomic analysis has witnessed significant activity especially in oncologic MRI, CT and PET, but remains to be thoroughly assessed in SPECT and/or cardiac imaging. Myocardial perfusion SPECT (MPS) is established for diagnosis of patients suspected with coronary artery disease (CAD). Meanwhile, coronary artery calcification (CAC) scoring is shown to offer added value in patients with negative MPS for identifying patients with significant CAD. Nonetheless, CAC scoring is not readily available in the community setting and is not currently reimbursed by CMS. We aimed to assess whether quantitation of heterogeneity of MPS scans via radiomics analysis enables the prediction of CAC scores obtained from CT. Methods 372 patients were selected with normal (non-ischemic) stress MPS scans (injected with 8-30mCi 99mTc-Sestamibi, consensus reading). Cardiac risk-factors, including BMI, smoking, diabetes, hypertension, hyperlipidemia and family history of CAD were recorded. Images were iteratively reconstructed (attenuation-corrected, isotropic-cubic-voxels), and verified by a nuclear-medicine expert to be free from fixed defect, and two common MPS artifacts: severe overcorrection and liver/diaphragmic spillovers. Semi-automatic segmentation was performed under radiologist supervision to generate 7 regions-of-interests: myocardium, 3 vascular segments from the vascular map (LADv-LCXv-RCAv), 3 vascular segments from the bull’s eye 17-segment polar plot (LADp-LCXp-RCAp). These 7 segments were then uniformly discretized into grey-level (GL) bins (8 different discretization GLs: 22,⋯,29), and 188 3D radiomic features were subsequently evaluated, which were standardized based on the Image Biomarker Standardization Initiative [1]. We then performed univariate analysis (Spearman correlation) with correction for multiple testing (Benjamini-Hutchberg false discovery rate (FDR), α=0.05). Subsequently, we performed multivariate machine learning analysis (step-wise linear regression, 60%/20%/20% for training/cross-validation/test, sum of squared errors criterion) on each segmentation to assess the predictability of CAC scores for a given segment from its MPS radiomics. Results To reduce sensitivity to outliers, we thresholded CAC scores over 400 by 400+log2(CAC). In univariate analysis, the consistently significant features (FDR q-value<0.05) observed were: intensity skewness and GLCM cluster shade for RCA, and intensity at 90% volume histogram for LCX. Multivariate analysis was performed A) without and B) with patient risk factors data. In A, CAC in LADv was predicted significantly well (p-value<0.001). CAC in LCXv and LCXp were also predicted with p-values of 0.0164 and 0.0181, respectively. In B, in addition to hyperlipidemia which consistently appeared in the analysis, radiomic features depicted the best predictability of CAC for entire myocardium with p-value <0.001. LCXv and LCXp had p-values of 0.0157 and 0.0181, and the p-value for prediction of CAC in LADv was 0.0154.
Conclusions: Our results demonstrate the ability of radiomics analysis to capture valuable information from MPS scans, enabling significant correlation of perfusion heterogeneity to CAC scores. These results suggest that radiomic analysis has the potential to add diagnostic and prognostic value to standard MPS for wide clinical usage. Acknowledgement This work was in part supported by the 2017 Bradley-Alavi fellowship (Saeed Ashrafinia) from SNMMI.
