Diagnosis of Cardiac Amyloidosis Using a Radiomics Approach Applied to Late Gadolinium-Enhanced Cardiac Magnetic Resonance Images: A Retrospective, Multicohort, Diagnostic Study

OBJECTIVES: To assess the potential of a radiomics approach of late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) in the diagnosis of cardiac amyloidosis (CA). MATERIALS AND METHODS: This retrospective study included 200 patients with biopsy-proven light-chain (AL) amyloidosis. CA was diagnosed on the basis of systemic amyloidosis confirmed with evidence of cardiac involvement by imaging and clinical biomarkers. A total of 139 patients [54 ± 8 years, 75 (54%) men] in our institution were divided into training cohort [n = 97, mean age of 53 ± 8 years, 54 (56%) men] and internal validation cohort [n = 42, mean age: 56 ± 8 years, 21 (50%) men] with a ratio of 7:3, while 61 patients [mean age: 60 ± 9 years, 42 (69%) men] from the other two institutions were enrolled for external validation. Radiomics features were extracted from global (all short-axis images from base-to-apex) left ventricular (LV) myocardium and three different segments (basal, midventricular, and apex) on short-axis LGE images using the phase-sensitive reconstruction (PSIR) sequence. The Boruta algorithm was used to select the radiomics features. This model was built using the XGBoost algorithm. The two readers performed qualitative and semiquantitative assessment of the LGE images based on the visual LGE patterns, while the quantitative assessment was measured using a dedicated semi-automatic CMR software. The diagnostic performance of the radiomics and other qualitative and quantitative parameters were compared by a receiver operating characteristic (ROC) curve analysis. A correlation between radiomics and the degree of myocardial involvement by amyloidosis was tested. RESULTS: A total of 1,906 radiomics features were extracted for each LV section. No statistical significance was indicated between any two slices for diagnosing CA, and the highest area under the curve (AUC) was found in basal section {0.92 [95% confidence interval (CI), 0.86–0.97] in the LGE images in the training set, 0.89 (95% CI, 0.79–1.00) in the internal validation set, and 0.92 (95% CI, 0.85–0.99) in the external validation set}, which was superior to the visual assessment and quantitative LGE parameters. Moderate correlations between global or basal radiomics scores (Rad-scores) and Mayo stage in all patients were reported (Spearman’s Rho = 0.61, 0.62; all p < 0.01). CONCLUSION: A radiomics analysis of the LGE images provides incremental information compared with the visual assessment and quantitative parameters on CMR to diagnose CA. Radiomics was moderately correlated with the severity of CA. Further studies are needed to assess the prognostic significance of radiomics in patients with CA.