Embedded prompt tuning: Towards enhanced calibration of pretrained models for medical images

Wenqiang Zu, Shenghao Xie, Qing Zhao, Guoqi Li, Lei Ma

Abstract: Highlights•Identifies the domain gap in adapting foundation models from natural to medical images.•Enhances the existing paradigm of Parameter-Efficient Fine-Tuning (PEFT) by introducing Embedded Prompt Tuning (EPT).•Demonstrates through rigorous benchmarking that EPT outperforms state-of-the-art methods in few-shot learning scenarios for medical image tasks.•Proposes and validates the concept of EPT as a powerful distribution calibrator, effectively addressing anomalies in feature space distribution and improving the interpretability of foundation model behaviour.