Multimodal vViT transformer predicts renal tumor WHO/ISUP grade from CT and radiomics

A variable Vision Transformer (vViT) using CT images, radiomics, and clinical data achieved an AUC of 0.856 for distinguishing low- from high-grade renal tumors, with radiomic features contributing most to the prediction.

• Retrospective study trained on 111 patients (1398 images), internally validated on 15 patients, and tested on 30 patients.
• vViT accuracy on the test set was 0.811, significantly outperforming baseline ViT and ResNeXt models (p<0.05), but not significantly better than ConvNeXt.
• Single-center design with a small test cohort (n=30) limits generalizability; no external validation was performed.

RadPigeon summaries are original and for information only. They are not clinical advice.