CMR-Analysis (including machine learning)
Avanti Gulhane, MD
Acting Clinical Instructor
University of Washington
Seattle, Washington, United States
Avanti Gulhane, MD
Acting Clinical Instructor
University of Washington
Seattle, Washington, United States
Peter Muzi, BEng
Research Scientist
University of Washington, United States
Mladen Zecevic, BEng
Research Scientist
University of Washington, United States
Murat Sadic, MD, PhD
Resident
University of Washington
Seattle, Washington, United States
Mohamed Abdelmotleb, MD
Cardiothoracic Imaging Fellow
University of Washington, United States
Mehrzad Shafiei, MD
Post Doctoral Scholar
University of Washington, United States
Hamid Chalian, MD
Associate Professor of Radiology
University of Washington, United States
Matthew Nyflot, PhD
Associate Professor
University of Washington, United States
Hubert Vesselle, MD, PhD
Professor of Nuclear Medicine
University of Washington, United States
Ganesh Raghu, MD
Professor of of Pulmonary, Critical Care and Sleep Medicine
University of Washington, United States
Karen Ordovas, MD, MAS
Professor of Radiology, Cardiothoracic Imaging Section Chief, Associate Vice-Chair for Academic Affairs
University of Washington
Kirkland, Washington, United States
Radiomics feature extraction is an attractive option to potentially aid quantitative information on CMR to detect cardiac abnormalities not easily visualized by the naked eye. Currently, PET/CT is considered the gold standard for detection of myocardial inflammation on cardiac sarcoidosis (CS). Late gadolinium enhancement (LGE) can detect cardiac involvement in CS but is suboptimal for active inflammation. We aim to identify radiomics features derived from LGE CMR in patients with CS that could predict FDG uptake on PET/CT.
Methods:
We retrospectively identified consecutive patients with confirmed or probable CS referred for both CMR and F18-FDG PET/CT within 90 days from each other in the past 5 years. MIM encore software ® (Cleveland, OH, version 7.1.3.) was used to extract radiomics features. A single experienced radiologist blinded to PET information marked two fixed region of interest (ROI) of 2ml on one short axis slice per patient, excluding the blood pool and epicardial fat. Preprocessing was performed using the pyradiomics extension launched from within MIM and steps included image resampling to 1x1x1 mm using linear interpolation and discretization to a fixed bin width of 25 units.
F18-FDG PET/CT studies were reviewed for presence of FDG uptake above blood pool and were defined as positive or negative for inflammation. CMR segments were reviewed for presence or absence of LGE separately by 2 experienced board-certified radiologists. Extracted radiomics features for each segment were compared with FDG findings on PET/CT using parametric and non-parametric test of hypothesis according to variable distribution.
Results:
In the 23 patients (mean age 54.7 years, 61% male) with CS analyzed, 18/46 (39.1%) segments had FDG uptake on PET/CT and 12 segments (26%) showed abnormality on LGE images. Significant quantitative differences (median [IQR] or mean [std]) for LGE derived radiomics features were noted for Entropy (4.8[1.17] vs 5.7[0.5], p= 0.048), Kurtosis (3.33 [1.23] vs 2.55 [1.62], p=0.02), and Grey Level Run Length Entropy (GLRLE) (5.09 [0.75] vs 5.66 [0.93], p=0.02) when compared to PET findings of absence or presence of FDG uptake. Additional features extracted showed non-significant differences between segments with and without FDG uptake, including Skewness, Gray Level Cooccurrence Matrix Autocorrelation (GLCMA), Gray Level Cooccurrence Contrast (GLCC) and Grey Level Run Length Non-Uniformity (GLRLNU) (Table 1).
Significant difference was noted for Entropy, Kurtosis and Skewness, GLRLE and GLRLNU between segments with LGE+/FDG+ (active disease) and segments with LGE+/FDG- (scar) (Table 2).
Conclusion: Quantitative radiomics features on LGE CMR have the potential to predict inflammation in CS. Larger datasets are needed to explore these novel quantitative metrics that may be further exploited in artificial intelligence efforts in CMR.