Background: Clear cell renal cell carcinoma (ccRCC) shows marked clinical heterogeneity, and precise grading is critical for risk stratification and treatment planning. Conventional grading relies on invasive tissue sampling, which imposes patient burden and may miss intratumoral heterogeneity.
Methods: We introduce the RCC-Align framework, which leverages paired whole‑slide histopathology images (WSIs) and computed tomography (CT) scans during training to perform cross‑modal contrastive learning. A contrastive loss aligns microscopic tissue morphology with macroscopic radiologic representations, transferring grade‑discriminative cues to CT features. The model is trained and evaluated on paired TCGA and CPTAC cohorts using patient‑level five‑fold cross‑validation. Performance metrics include AUC, AUPRC, and statistical significance for low‑ versus high‑grade prediction (p‑value). Alignment quality is quantified by cosine similarity between WSI and CT embeddings.
Results: RCC-Align achieved an AUC of 0.601 (95% CI: 0.524‑0.673) and an AUPRC of 0.599 (95% CI: 0.541‑0.676), outperforming the CT‑only baseline DINOv2‑Finetuned (AUC 0.545, AUPRC 0.543) with a significant improvement in low‑grade prediction (p = 0.004). Cosine similarity analysis showed stronger paired WSI‑CT embedding alignment compared with baselines. The WSI‑based reference model GigaPath‑Finetuned reached an AUC of 0.719.
Conclusion: Pathology‑guided contrastive learning enhances CT‑based ccRCC grading while requiring only CT at inference, offering a complementary tool when biopsy is unsafe, infeasible, or limited by heterogeneity. Larger, multi‑institutional external validation is required before clinical translation.
Review