Prototype-guided CT foundation model adaptation predicts lung cancer survival

A prototype-guided CT foundation model predicted lung cancer progression-free survival with AUC 0.765 (507 patients) and overall survival with AUC 0.822 (420 patients), outperforming clinical and deep-learning baselines.

• ProtoSurv was pretrained on 409,261 chest CT sequences from 104,783 patients using a CT-report vision-language model, then adapted to limited labeled prognosis data via prototype-guided pseudolabeling, confidence-based filtering, and feature calibration.
• For progression-free survival prediction in 507 patients receiving targeted therapy, ProtoSurv achieved AUC 0.765 and concordance index 0.684.
• For overall survival prediction in 420 patients receiving (chemo-)radiotherapy, ProtoSurv achieved AUC 0.822 and concordance index 0.727, outperforming conventional clinical models and representative deep learning baselines.

Automated summary

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