Abstract
94
Objectives: Despite relatively low resolution of SPECT images, recent studies in brain DAT SPECT [1, 2] have shown promising results in heterogeneity and radiomics analysis of SPECT images. Nonetheless, cardiac SPECT radiomics remain to be explored. Myocardial perfusion SPECT (MPS) scan is an established diagnostic test for patients suspected of coronary artery calcification (CAD). In this study, we aimed to utilize a radiomics workflow including identification of appropriate settings for systematic MPS radiomics. Specifically, we investigate prediction of coronary artery calcification (CAC) obtained from CT. CAC scoring is not readily available in the community centers and is not currently reimbursed by CMS, but shown to offer incremental diagnosis information over MPS for identifying patients with significant CAD and negative MPS results [3]. Methods We identified 372 patients with normal (non-ischemic) stress 99mTc-Sestamibi MPS scans with consensus reading. Iterative reconstruction images (attenuation-corrected, isotropic-cubic-voxels) were verified by a NM physician to be free from fixed perfusion defect and artifactual attenuation. Semi-automatic segmentation was performed by a radiologist and 7 regions-of-interests were generated: myocardium, 3 vascular segments (LAD-LCX-RCA) from the vascular map, and the 17-segment polar plot. Our radiomics framework is in compliance with the Image Biomarker Standardization Initiative [4] to ensure the reproducibility of the study. We performed two feature elimination phases to prevent overfitting: pre- and post-feature calculation. ROIs of MPS images have arbitrary units, fixed shape and isotropic voxels. This helps to eliminate morphological, local intensity (peak), and all higher order 2D features pre-calculation, leaving 188 features. Arbitrary voxel intensities suggest discretizing images via fixed number of bins (FNB) method (as opposed to fixed bin-size). We discretized images into 8 different grey-levels (GLs) (22,⋯,29), and evaluated 188×8 features for 7 ROIs. Multiple correlation analysis with correction for multiple testing (Benjamini-Hutchberg false discovery rate (FDR), α=0.05) was subsequently performed between radiomic features and the CAC score for each 7 segments. Results Variance of each feature-GL-ROI was evaluated to assess the feasibility of features perform post-calculation elimination. 938 features had var<1e-5: (5, 6, 18, 39, 177, 205, 231, 257) feature-ROI pair for each GLs (22 to 29), respectively. Higher GLs often have more bins than the No. of voxels, resulting in very few columns in GLRLM, GLSZM, GLDZM, NGTDM and NGLDM; exhibiting very small variation. CAC data is highly prone to outliers; suggesting rank-correlation analysis. Most features are not reproducible to GL variations, and most significant correlations are seen in 16, 32 and 64 GLs. After FDR correction, the consistently significant features (FDR q-value<0.05) observed include: intensity skewness and GLCM cluster shade for RCA, and intensity at 90% vol. histogram for LCX. This analysis also showed LAD and RCA extracted from vascular plot had more significant correlation than bull’s eye plot (FDR q-value<0.05), while LCX from the latter plot is more significant. In the LAD ROI, apical thinning and slight overcorrection due to breast attenuation appears occasionally (118 cases) resulting in a cold uptake in apico-anterior wall. This generates a heterogeneity irrelevant to CAC, impacting the correlation, suggesting potentially removing the apex from LAD.
Conclusions: FNB discretization by 16 to 64 GLs is suggested for radiomics of MP SPECT images. Correlation analysis shows promising results in attributing radiomic features to CAC score. These results suggest that radiomic analysis has the potential to add diagnostic/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.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.