Differentiation of adrenal metastases and adenomas based on clinical characteristics, deep learning features, and radiomics features derived from ultrasound imaging

Accurate diagnosis of adrenal incidentalomas is crucial in patients with extra-adrenal malignant tumors. This study aims to develop a nomogram integrating clinical features, deep learning-derived imaging features, and ultrasound radiomics characteristics to distinguish adrenal metastases from adrenal adenomas. A retrospective analysis was conducted on 449 cases, including 228 cases of adrenal metastases and 221 cases of adrenal adenomas, divided into training and testing cohorts at a 7:3 ratio. Patient clinical data and ultrasonographic images were collected, with regions of interest (ROIs) delineated on ultrasound images. Feature extraction, selection, and radiomics model (Rad) construction were performed, followed by model evaluation. Multiple deep learning models were employed to identify the optimal architecture for deep feature extraction. These deep features were combined with radiomics features to establish a deep learning radiomics model (DLR) for the differentiation between adrenal metastases and adenomas. The study demonstrates that the Rad, DLR, and combined models exhibit superior diagnostic performance in differentiating adrenal metastases from adrenal adenomas. In the testing cohort, the combined model outperforms the Rad and DLR models. The area under the curve (AUC) in the testing set for Rad, DLR, and combined models were 0.839, 0.839, and 0.850, respectively. The nomogram integrating clinical features, deep learning features, and ultrasound radiomics features demonstrates robust performance in differentiating adrenal metastases from adrenal adenomas and can assist in preliminary diagnostic stratification of indeterminate adrenal nodules in patients with extrarenal tumors.