Explainable 3D AI framework with counterfactuals and uncertainty detects and segments renal cysts on contrast-enhanced CT

A multi-center 3D AI framework for renal cyst segmentation on contrast-enhanced CT achieved a Dice score of 0.82 and sensitivities of 0.95–0.97 on external test sets. Uncertainty-based flagging reduced poor segmentations from 17% to 5% when 20% of cases were reviewed.

• Counterfactual explanations and radiomics identified image features associated with segmentation performance.
• External validation included 568 patients and 1486 cysts from three centers and an open-source dataset.
• VAE-GAN reconstruction did not significantly alter Dice scores.

Automated summary

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