General-purpose vision-language models as computed tomography feature encoders for identifying advanced renal cell carcinoma

To evaluate whether general-purpose vision-language models can generate numerical features from preoperative noncontrast CT for predicting postoperative pathological AJCC stage groups in renal cell carcinoma. This retrospective study included 260 public development cases and 183 external cases from Shanxi Bethune Hospital. We evaluated 2,160 combinations of five VLMs, image inputs, radiomic retrieval, feature definitions, vector lengths and classifiers. Configuration selection used mean three-fold F1, followed by OOF threshold selection and refitting on the complete development cohort. Comparators included 3D ResNet-18, two radiomic classifiers, a retrieval classifier, MedGemma and physicians. Patient-level stratified bootstrap quantified uncertainty. Physician assistance was assessed in a sequential exploratory design. All three repeated calls produced exactly identical output text and numerical values under fixed conditions. The selected GPT-5.6 Sol workflow used bilateral crops, 21 slices, labeled retrieval, 128 semantic features and TabTransformer. Mean validation-fold F1 was 0.808 ± 0.052. External F1 was 0.698 with a 95% confidence interval of 0.563–0.814, and AUROC was 0.942 with a 95% confidence interval of 0.901–0.975. External F1 was 0.708 for ResNet and 0.688 for physician consensus. The VLM-minus-ResNet F1 difference was − 0.009 with a 95% confidence interval of − 0.133 to 0.106. Assisted consensus F1 was 0.750; its change of 0.063 had a 95% confidence interval of − 0.048 to 0.175. Prompt-generated features from preoperative noncontrast CT supported external prediction of postoperative pathological AJCC stage groups in RCC. Performance reflected the full input and modeling workflow. Assisted second reading had a higher observed F1 in the exploratory sequential assessment.