Advancing In-Context Learning for Efficient and Stable Medical Report Generation.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computer Vision

Background:

• Vision-language models (VLMs) excel at multimodal tasks but struggle with medical report generation (MRG) due to limited data and high annotation costs.
• Standard in-context learning (ICL) for VLMs in MRG is inefficient and produces inconsistent, clinically inaccurate reports.

Purpose of the Study:

• To introduce Principal In-Context Vectors (PCVs) as a novel, training-free framework for enhancing VLM performance in MRG.
• To address the limitations of existing ICL methods by providing a computationally efficient and accurate approach.

Main Methods:

• Developed PCVs by extracting hidden states from auto-regressive VLMs and applying principal component analysis (PCA) to distill demonstrations into stable semantic representations.
• Injected PCVs into new queries to guide generation without requiring any model tuning.

Main Results:

• PCVs significantly improved zero-shot and fully supervised MRG quality across four benchmark datasets.
• The approach demonstrated effectiveness in diverse scenarios, including cross-center, cross-disease, and longitudinal data.

Conclusions:

• PCVs offer a lightweight and scalable solution for adapting pre-trained VLMs for practical clinical deployment in MRG.
• This method enhances the accuracy and clinical meaningfulness of generated medical reports efficiently.