Fig 1.
Repeated nested cross-validation process for feature selection, hyperparameter selection, and model evaluation.
In each repetition, data is split into 5 folds (1–5). In each fold, 80% are used in an internal CV process for feature selection, hyperparameter optimization, and model training (green squares). The resulting model is tested on the remaining 20% (red square), recording the probability scores and the SHAP values of the random forest model. CV: Cross-validation. SMOTE: Synthetic minority oversampling technique. SHAP: Shapley additive explanations.
Table 1.
Patient characteristics.
Fig 2.
Features selected for the prediction of a) histologic subtype (WHO classification, low-risk vs. high-risk tumors), b) TNM stage (IASLC/ITMIG, early vs. advanced stage) and the presence of c) myasthenia gravis. The bar plots on the left display how often features were selected across folds as an indicator of selection stability. The bar plots in the center show the feature importance measured by the mean absolute SHAP values, representing the impact of a feature on the individual model prediction. The boxplots on the right display the individual standardized feature values grouped by the underlying category. The selected feature values differed significantly for all tested categories (p<0.05).
Fig 3.
Correlogram of all selected radiomic features.
Colors and numbers show Pearson’s correlation coefficient r of the respective feature pair.
Fig 4.
Axial CT slices (a, d, e, h) and corresponding volume renderings (b, c, f, g) of segmented VOIs with calculated values for sphericity (S) and flatness (F). LRT: Low-risk thymic epithelial tumor. HRT: High-risk thymic epithelial tumor.
Fig 5.
Receiver operator characteristic curves of the random forest classifier performance in each category.
Values in square brackets indicate 95% confidence intervals. AUC: Area under the ROC curve.
Fig 6.
Comparison of a low-risk TET (LRT, left image) and high-risk TET (HRT, right image) with corresponding radiomic fingerprints. Both tumors have been captured in early stages. DV: Dependence variance. LoGn: Laplacian of Gaussian filter with σ = n. RLNUN: Run length non-uniformity. SRLGE: Short run low gray emphasis. SRE: Short run emphasis. wlHHH: Wavelet filter with high-pass filters (H) in every spatial direction. LDE: Large dependence emphasis.
Table 2.
Random forest classifier performance.