Using interpretable rule-learning artificial intelligence to optimally differentiate adrenal pheochromocytomas from adenomas with CT radiomics.

Abstract

Purpose: To identify interpretable CT-based radiomics features that can differentiate adrenal pheochromocytomas from adenomas.

Methods: An institutional database was used to identify patients with pathologically proven adrenal pheochromocytomas 5/1/05-5/1/23. To be included, patients needed to have a contrast-enhanced abdominal CT with an adrenal mass within 12 months of pathology (n = 95). For comparison, 57 adenomas were identified from a set of consecutive CT examinations. The final dataset included 152 adrenal masses (95 pheochromocytomas; 57 adenomas) with 121 used in the development set and 31 in the test set. Following confirmation of accurate automated segmentation, 463 radiomic features were evaluated and used to create an interpretable artificial intelligence (AI) rule-learning model. Model performance was reported using F1 score.

Results: The study included 146 patients (age 59 years +/- 21; 89 females). A three-feature rule, High Gray Level Zone Emphasis > 184, Roundness > 0.35, and Boundary Low Gray Level Emphasis < 0.021 produced an F1 score of 0.97 on the train set (95% confidence interval [CI]: 0.94, 0.99) and 0.96 on the test set (95% CI: 0.89, 1.00). The rule-learning model determined that the rule most predictive of pheochromocytoma was Maximum Pixel Attenuation > 125 HU resulting in an F1 score of 0.89 (95% CI: 0.83, 0.94) on the training set and 0.93 (95% CI: 0.83, 0.99) on the test set.

Conclusion: A rule-learning AI model identified the smallest optimal set of interpretable CT radiomics features, sufficient to achieve 96% accuracy in differentiating adrenal pheochromocytomas from adenomas on contrast enhanced CT.