External specificity of the RCC classification model

Determine the specificity of the radiomics–foundation fusion model for clear cell versus non-clear cell renal cell carcinoma classification on adequately sized, balanced external cohorts, because specificity cannot be established from the available external evaluation.

Background

The study evaluates radiomics, conventional convolutional neural network features, MedicalNet-pretrained features, MedVAE representations, and fusion models for binary clear cell renal cell carcinoma versus non-clear cell renal cell carcinoma classification from contrast-enhanced CT. Although the best-performing 3D MedVAE gated-fusion model achieved promising external results, the external TCGA cohort contained only two non-clear cell cases. This extreme class imbalance prevents reliable estimation of specificity and leaves the model’s external diagnostic specificity unresolved. Larger, balanced external cohorts are therefore needed to determine whether the observed transfer performance generalizes beyond the exploratory evaluation.