A lung CT vision foundation model facilitating disease diagnosis and medical imaging.

Summary

LCTfound is a large-scale lung CT foundation model trained on 105,184 scans that jointly encodes imaging and clinical data to support eight clinical tasks, outperforming strong baselines across centers. It provides a unified, deployable framework for diagnosis, prognosis, reconstruction, and navigation, potentially standardizing AI-assisted thoracic imaging.

Key Findings

• Trained on 105,184 multicenter lung CT scans with diffusion-based pretraining and joint imaging–clinical encoding.
• Supports eight tasks (enhancement, virtual CTA, sparse-view reconstruction, segmentation, diagnosis, prognosis, response prediction, 3D navigation).
• Consistently outperformed leading baselines across multiple centers, providing a unified deployable framework.

Clinical Implications

If prospectively validated and integrated, this model could improve consistency and speed of thoracic imaging interpretation, enhance lesion detection/segmentation, support treatment planning, and enable advanced reconstruction in low-dose or sparse-view settings.

Limitations

• Lack of prospective, randomized clinical deployment studies demonstrating impact on patient outcomes.
• Generalizability and fairness across diverse scanners, populations, and institutions require further auditing.

Future Directions

Prospective clinical trials, regulatory-grade validation, fairness auditing across demographics and scanners, and integration studies assessing workflow efficiency, safety, and outcome benefits.