Integrated deep learning with interactive segmentation and tri-phase CT improves cystic renal lesion risk stratification in multicenter study

An interactive deep learning model combining tri-phase CT for cystic renal lesions had AUC 0.969 (0.911-1.000) vs Bosniak v2019 AUC 0.884 (p=0.016) on external test, sensitivity 94.4%, specificity 96.7%.

• The TPC-Perceiver significantly outperformed Bosniak v2019 classification (AUC 0.969 vs 0.884, p=0.016), with specificity 96.7% vs 76.9%, suggesting reduced false positives.
• Decision curve analysis indicated greater net benefit for the deep learning framework across a broad range of threshold probabilities.
• The model was externally validated on a cohort from four institutions (n=139).

Automated summary

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