Prediction of cardiac remodeling and myocardial fibrosis in athletes based on IVIM-DWI images

Figures

• 

  Figure 1. Importance of variables in GBM model by MIP The top six predictors ranked from high to low MIP score were as follows: ADCslow values of AHA 3 (basal inferoseptal), AHA 16 (apical lateral), AHA 12 (mid anterolateral), AHA 14 (apical septal), 5 (basal inferolateral) and AHA 2 (basal anteroseptal). The smaller the MIP score, the more important the feature. AHA, American Heart Association; ADCslow, the slow apparent diffusion coefficient; GBM, Gradient boosting machine; MIP, Modified Informative Position.

• 

  Figure 2. The evaluation index of the GBM model (A) The AUC of 0.947 (95% CI:0.924–0.971) in the training set were obtained. (B) The AUC of 0.841 (95% CI:0.795–0.887) in the test set were obtained. (C) The Rad-score boxplot of the training set and the validation set showed the calculated score to effectively distinguish athletes with CR and/or MF from the athletes without CR and/or MF. (D and E) The calibration curves of the training (D) and validation (E) sets indicated good consistency between the model predictions and the actual outcomes. The closer the curve is to the straight line, the better the model is. (F) Decision curve maps further revealed that the model achieved substantial clinical net benefit across a wide range of risk thresholds.

• 

  Figure 3. Spearman’,s correlation after applying PCA Before and after PCA, the initial rankings of components from the original features were compared across methods by using Spearman’,s rank correlation to assess the inter-method agreement. The features correlation was very low, with a correlation coefficient of 0 and a p-value less than 1e-15 for all pairs. PCA, Principal Component Analysis.

• 

  Example of IVIM raw images in the basal-ventricular short axis slices for athletes (A–F) The images with different b values for 0, 20, 50, 100, 200, 350 s/mm2 were shown as A-F, respectively. (G–I) The calculated parametric maps of diffusion ADCslow (G), perfusion fraction f (H), and ADCfast (I) were showed respectively. (J–L) The 16 regions of interest in left ventricular myocardium were measured sequentially by drawing an oval area as large as possible in these calculated parametric maps of ADCslow (J), f (K) and ADCfast (L), avoiding the intracardiac blood flow, papillary muscles, edge pixels and small local artifacts area.