Improving Radiology Report Generation with Semantic Understanding.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Natural Language Processing

Background:

• Radiology report generation (RRG) is crucial for clinical decision-making.
• Current RRG models often require substantial computational resources for domain adaptation.
• Efficiently learning the medical domain for RRG remains a challenge.

Purpose of the Study:

• To propose RRG-LLM, a model that enhances RRG by learning the medical domain with minimal computational resources.
• To improve the accuracy and efficiency of radiology report generation.

Main Methods:

• Leveraging Large Language Models (LLMs) finetuned with Low-Rank Adaptation (LoRA) for efficient medical domain adaptation.
• Fine-tuning only the linear projection layer for projecting image information onto the text dimension.
• Utilizing a novel approach to extract key information from radiology images.

Main Results:

• The RRG-LLM model demonstrated significant improvements in report generation quality.
• Achieved a 0.096 (51.7%) increase in ROUGE-L score compared to the baseline.
• Achieved a 0.046 (42.85%) increase in METEOR score compared to the baseline.

Conclusions:

• RRG-LLM offers an effective and computationally efficient method for medical domain learning in RRG.
• The proposed finetuning strategy for the projection layer enhances information extraction from radiology images.
• This approach shows promise for advancing automated radiology report generation.