Using interpretable rule-learning artificial intelligence to optimally differentiate adrenal pheochromocytomas from...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Adrenal masses require accurate differentiation between benign adenomas and malignant pheochromocytomas.
• CT-based radiomics offers a non-invasive method for characterizing adrenal lesions.

Purpose of the Study:

• To identify interpretable CT-based radiomics features for differentiating adrenal pheochromocytomas from adenomas.
• To develop an AI model for accurate adrenal mass classification.

Main Methods:

• A dataset of 152 adrenal masses (95 pheochromocytomas, 57 adenomas) from contrast-enhanced CT scans was analyzed.
• 463 radiomic features were extracted and evaluated using an interpretable AI rule-learning model.
• Model performance was assessed using the F1 score.

Main Results:

• A three-feature radiomics rule achieved an F1 score of 0.97 on the training set and 0.96 on the test set.
• The most predictive rule for pheochromocytoma involved Maximum Pixel Attenuation > 125 HU, yielding an F1 score of 0.93 on the test set.
• The AI model demonstrated high accuracy in differentiating the two adrenal lesions.

Conclusions:

• An interpretable AI rule-learning model effectively identified radiomic features for differentiating adrenal pheochromocytomas from adenomas.
• The developed model achieved 96% accuracy on contrast-enhanced CT, offering a valuable tool for clinical decision-making.