Virtual-Eyes Improves Foundation Model Performance for Lung Cancer Risk Prediction
Published:Dec 30, 2025 15:34
•1 min read
•ArXiv
Analysis
This paper investigates the impact of a quality control pipeline, Virtual-Eyes, on deep learning models for lung cancer risk prediction using low-dose CT scans. The study is significant because it quantifies the effect of preprocessing on different types of models, including generalist foundation models and specialist models. The findings highlight that anatomically targeted quality control can improve the performance of generalist models while potentially disrupting specialist models. This has implications for the design and deployment of AI-powered diagnostic tools in clinical settings.
Key Takeaways
- •Virtual-Eyes, a CT quality-control pipeline, improves the performance of generalist foundation models (e.g., RAD-DINO) for lung cancer risk prediction.
- •Specialist models (e.g., Sybil, ResNet-18) may be negatively impacted by Virtual-Eyes, suggesting context dependence and shortcut learning.
- •The study highlights the importance of preprocessing and its differential impact on various model types in medical imaging AI.
Reference
“Virtual-Eyes improves RAD-DINO slice-level AUC from 0.576 to 0.610 and patient-level AUC from 0.646 to 0.683 (mean pooling) and from 0.619 to 0.735 (max pooling), with improved calibration (Brier score 0.188 to 0.112).”