Multi-model radiomics and machine learning for differentiating lipid-poor adrenal adenomas from metastases using...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Radiomics using automatic CT segmentation shows promise for distinguishing adrenal adenomas from metastases.
• Preoperative diagnostic value of these radiomics methodologies requires further evaluation.

Purpose of the Study:

• To evaluate the diagnostic value of radiomics based on automatic segmentation for differentiating adrenal adenomas from metastases.
• Utilize retrospective clinical and contrast-enhanced CT (CECT) data for analysis.

Main Methods:

• Retrospective analysis of 416 patients with adrenal masses (larger than 10 mm) using CECT data.
• Automatic segmentation of adrenal lesions, radiomic feature extraction (PyRadiomics), and feature selection (MI, MRMR, LASSO).
• XGBoost machine learning model incorporating clinical and imaging features, evaluated using AUC, accuracy, and SHAP analysis.

Main Results:

• The XGBoost model achieved AUC values of 0.81 (arterial phase), 0.81 (venous phase), 0.88 (combined phases), and 0.92 (combined phases and clinical indicators).
• Five-fold cross-validation demonstrated strong performance with average scores around 0.87-0.90.
• SHAP analysis provided interpretability by identifying influential features for model predictions.

Conclusions:

• A machine learning model combining multimodal CT radiomics and automatic segmentation improves differentiation between adrenal adenomas and metastases.
• This approach facilitates machine-based clinical feature extraction, offering a reliable basis for diagnosis and treatment planning.