Machine learning-based differentiation of benign and malignant adrenal lesions using 18F-FDG PET/CT: a two-stage classification and SHAP interpretation study

Figures

• 

  Fig. 1. The research flow chart

• 

  Fig. 2. t-SNE projection of the dataset based on selected features. Each point represents an individual sample, with colors indicating the binary classification labels (yellow = negative, purple = positive). The t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm was used to reduce the high-dimensional feature space to two dimensions for visualization

• 

  Fig. 3. LASSO feature selection results. A LASSO path plot showing the coefficient shrinkage process as a function of log-transformed regularization strength (lambda). B 10-fold cross-validation curve indicating the optimal lambda value that minimizes binomial deviance. C Final coefficients of selected features based on the optimal lambda. T/L SUVmax was assigned the largest negative weight, while adrenal gland lesions SUVmax and tumor plain scan CT value had positive contributions to the model

• 

  Fig. 4. Distribution of clinical and imaging features by binary target outcome (0 = Negative, 1 = Positive). Panels A–G show comparisons of each variable using boxplots: A Adrenal gland lesions SUVmax, B Age, C Liver SUVmean, D T/L SUVmax ratio, E Tumor diameter, F Tumor plain scan CT value, and G Gender

• 

  Fig. 5. Feature correlation heatmap. The plot illustrates pairwise Pearson correlation coefficients among all selected variables and the binary target. Color gradients range from − 1.0 (strong negative correlation, blue) to + 1.0 (strong positive correlation, red), with neutral values shown in white

• 

  Fig. 6. Model performance comparison using ROC and evaluation metrics. A Receiver Operating Characteristic (ROC) curves for seven classification models: Logistic Regression, Random Forest, AdaBoost, Support Vector Machine (SVM), XGBoost, Bagging, and Artificial Neural Network (ANN). All models demonstrate high predictive ability, with AUC values ranging from 0.98 to 1.00. B Bar plot comparison of Accuracy, F1 Score, and AUC across the seven models, showing that ensemble methods (Random Forest, XGBoost, Bagging) consistently achieve top performance across all metrics

• 

  Fig. 7. Decision curve analysis (DCA) comparing the net clinical benefit of different classification models across a range of threshold probabilities. The y-axis represents the net benefit, and the x-axis represents the threshold probability at which a patient would opt for treatment. Models shown include Logistic Regression, Random Forest, AdaBoost, Support Vector Machine (SVM), XGBoost, Bagging, and Artificial Neural Network (ANN). The “Treat All” and “Treat None” strategies are included as reference baselines (dashed lines)

• 

  Fig. 8. A The bar chart shows the mean absolute SHAP values for each feature, representing their average contribution to the model’s prediction of adrenal lesion malignancy. Features with higher SHAP values are more influential. B The SHAP summary plot visualizes the distribution and direction of each feature’s impact. Each point represents an individual patient sample. The x-axis indicates whether the feature pushes the prediction toward benign (left) or malignant (right). The color scale reflects the actual feature value: red for high values and blue for low values. For example, lower adrenal SUVmax (blue) tends to be associated with malignancy in our dataset, while higher CT values (red) also contribute to malignancy prediction. This plot helps visualize nonlinear effects and variable interactions on the model output.

• 

  Fig. 9. SHAP dependence plots showing nonlinear and interaction effects among key predictors.Each subplot (A–G) illustrates the relationship between a specific feature’s value (x-axis) and its SHAP value (y-axis), which indicates the feature’s contribution to predicting malignancy. A higher SHAP value means the feature is pushing the model toward predicting malignancy.The color of each point represents the value of a second interacting feature, helping to reveal interaction effects. For example, in (A), lower adrenal SUVmax is associated with higher SHAP values (malignancy), and the liver SUVmean (color gradient) modulates this relationship. These plots help explain the nonlinear behavior of features and their combined impact on prediction, making the model more interpretable to clinicians

• 

  Fig. 10. LASSO-based feature selection for subclassification of malignant adrenal lesions (lung cancer vs. lymphoma). A LASSO coefficient path illustrating the shrinkage of coefficients as the regularization parameter (log λ) increases. B Cross-validation curve showing binomial deviance across different λ values. The optimal λ is selected based on the minimum deviance with the one-standard-error rule. C Coefficients of the features retained by LASSO at the optimal λ. Eight variables were selected, including T/LSUVmax, Adrenal gland SUVmax, SUVpeak, Tumor diameter, Gender, Tumor CT value, Age, and Liver SUV metrics. Notably, T/LSUVmax and Adrenal SUV metrics showed the strongest positive associations with lymphoma classification

• 

  Fig. 11. Boxplot Comparison of Feature Distributions Between Lung Cancer and Lymphoma Subgroups. Panels A–I display the distribution of nine key imaging and demographic variables across two diagnostic groups: lung cancer (Target = 0) and lymphoma (Target = 1). A–B Adrenal gland lesions SUVmax and SUVpeak. C–D Age and Gender. E–F Liver SUVmax and SUVmean (G) T/LSUVmax ratio. H Tumor diameter. I Tumor plain scan CT value

• 

  Fig. 12. Performance comparison of machine learning models for classifying lung cancer vs. lymphoma. A Receiver operating characteristic (ROC) curves for all eight models evaluated. B Bar plots showing Accuracy, F1-score, and AUC for each model