Differentiating between adrenocortical carcinoma and pheochromocytoma by radiomics features at CT: A multi-institutional retrospective study.

4603 Background: Adrenocortical carcinoma (ACC), a highly malignant and rare tumor of the adrenal gland, shares similar CT characteristics with the less malignant pheochromocytoma (PHEO), resulting in low CT diagnostic accuracy. The standard therapeutic strategy for ACC differs from that of PHEO, making accurate diagnosis of ACC challenging but important. Radiomics offers an opportunity to classify adrenal tumors. However, previous research has primarily focused on differentiating malignant adrenal lesions from benign cases. In addition, the rarity of ACC makes it difficult to initiate large-scale studies, thus hindering further applications of radiomics. This study aimed to differentiate between ACC and PHEO using radiomics features based on contrast-enhanced CT. Methods: A total of 158 patients (median age, 47 years; inter-quartile range, 32-57 years; 76 males) pathologically diagnosed with ACC or PHEO between 2011 and 2023 were enrolled from three independent institutions. Radiomics features were extracted from different phases of contrast-enhanced CT images, and then selected by a two-step procedure. The radiomics model was developded and trained in the development cohort of 109 patients from Institution 1, then the model performance was tested in the test cohort of 49 patients from Institution 2 and 3. The area under the receiver operating characteristic curve (AUC) of the radiomics model was compared with two radiologists using the DeLong test. The SHAP method was used to improve the interpretability of the radiomics model. Results: We developed and tested a radiomics model that consisting of 10 selected radiomics features. In the test cohort, the radiomics model exhibited a significant improvement in diagnostic performance over 2 radiologists (AUC 0.92 vs. 0.79, 0.63) and achieved high accuracy (86%), sensitivity (81%) and specificity (88%) in differentiating between ACC and PHEO. Furthermore, the diagnostic process of this radiomics model was visualized using the SHAP method. Conclusions: To our knowledge, this is the first multi-institutional study to develop an interpretable radiomics model for preoperative differentiation between ACC and PHEO based on contrast-enhanced CT. The diagnostic performance of the radiomics model surpassed that of experienced radiologists, which may aid clinical decision making and improve treatment outcomes for ACC.