Program Information
Robustness and Reproducibility of Radiomic Features in 99mTc-Sestamibi SPECT Imaging of Myocardial Perfusion
S Ashrafinia1*, P Ghazi2 , C Marcus3 , M Taghipour4 , R Yan5 , I Valenta6 , M Pomper7 , T Schindler8 , A Rahmim9 , (1) ,Baltimore, MD, (2) INTEGRIS Baptist Medical Center, Oklahoma City, OK, (3) Johns Hopkins University, School of Medicine, Baltimore, MD, (4) Harvard Medical School, Brigham and Women's Hospital, Boston, MA, (5) Johns Hopkins University, School of Medicine, Baltimore, MD, (6) Johns Hopkins University, School of Medicine, Baltimore, MD, (7) Johns Hopkins University, School of Medicine, Baltimore, MD, (8) Johns Hopkins University, School of Medicine, Baltimore, MD, (9) Johns Hopkins University, Baltimore, MD
Presentations
SU-E-708-5 (Sunday, July 30, 2017) 1:00 PM - 1:55 PM Room: 708
Purpose: 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. We aimed to assess the reproducibility and reliability of textural (radiomic) features in ⁹⁹ᵐTc-Sestamibi Myocardial-Perfusion SPECT (MPS) images, and to derive robust features for correlation with clinical outcomes.
Methods: 94 patients were selected with normal (non-ischemic) stress MPS scans (injected with 8-30mCi ⁹⁹ᵐTc-sestamibi). Images were iteratively reconstructed (attenuation-corrected, isotropic-cubic-voxels), and verified by nuclear medicine physician to be free from two common MPS artifacts: image overcorrection, and liver spillover. Semi-automatic segmentation and polar map generation was performed (MIM Software™) under radiologist supervision, to generate: A) total-myocardium, B) three-region vascular (LAD-RCA-LCX), and C) 17 polar segments. Images were uniformly quantized with various gray-levels (4,8,16,32,64,128,256,512). 85 radiomic features were generated. For A-B-C segmentation scales, Spearman’s rank-correlation was calculated between 512 gray-levels and each of 7 other gray-levels to select relatively consistent quantization levels. The intra-class correlation (ICC) between remaining gray-levels across all patients was used to adopt robust features across each segmentation model.
Results: Maximum count in images varied substantially (3,428-42,761). Morphological and moment-invariant features should be left out due to standardized segmentation. Consistent Spearman-correlation≥0.6 was observed for gray-levels≥128 for B-scale and ≥64 for C-scale. At A-scale (myocardium), dissimilarly showed a relatively consistent correlation across all gray-levels. ICC, averaged over all segments within A,B,C, was ≥0.8 for 19,19,16 features and ≥0.9 for 13,11,7 features, respectively. Shared features with ICC≥0.9 include: skewness, kurtosis(hist), entropy,variance,correlation,IDMN(GLCM), RLN(GLRLM). ICC≥0.8 further includes entropy(hist), differenceentropy,IDN(GLCM), LGRE,SRLGE(GLRLM) and ZSN,LGZE(GLSZM) to this set.
Conclusion: Quantization by 128 or 256 gray-levels is suggested for capturing heterogeneity information. A set of 14 reproducible and robust features were identified, some of which have demonstrated reproducibility in other modalities, and are recommended for further investigation of predictive or prognostic values.
Contact Email: