Differentiating Fat-Poor Angiomyolipoma From Renal Cell Carcinoma Using Contrast-Enhanced CT: Development and External Validation of an AI-Assisted Diagnostic Model

Background Fat-poor angiomyolipoma (fp-AML) is a common benign renal mass that often mimics renal cell carcinoma (RCC) on computed tomography (CT), leading to unnecessary surgery. We aimed to develop and externally validate an AI-assisted radiomics model based on preoperative triphasic contrast-enhanced CT to differentiate fp-AML from RCC. Methods In this multicenter retrospective study, patients with surgically treated, non-metastatic sporadic solid renal lesions were included if preoperative triphasic CT (non-contrast, arterial, and venous phases) and postoperative pathology (RCC of any subtype or fp-AML without visible fat) were available. Tumors were segmented using an nnU-Net-assisted workflow on arterial-phase images with radiologist refinement, and other phases were registered to the arterial phase. Radiomics features were extracted and reduced using LASSO; the top 25 features were used to train a random forest classifier. The optimal cutoff was determined using the Youden index in the internal test set. Model performance was assessed by AUC, accuracy, sensitivity, specificity, and F1-score in the training, internal test, and independent external validation cohorts. Results A total of 655 patients were included (training: n = 364; internal test: n = 156; external test: n = 135) with an fp-AML:RCC ratio of approximately 1:4. The model achieved an AUC of 0.868 (95% CI: 0.748–0.940) in the training set and 0.868 (95% CI: 0.712–0.948) in the internal test set. In the external validation cohort, the model demonstrated stable generalizability with an AUC of 0.803 (95% CI: 0.741–0.865), accuracy of 73.3%, sensitivity of 72.2%, specificity of 100%, and F1-score of 0.90. In the subgroup of small renal masses (≤ 4 cm), discrimination remained consistent in test cohorts (internal AUC 0.808; external AUC 0.799). Conclusions We developed and externally validated a triphasic CT-based radiomics model that reliably differentiates fp-AML from RCC across centers, with particularly high specificity. This AI-assisted tool may support standardized preoperative decision-making and help reduce unnecessary surgery for benign fp-AML, especially in small renal masses.