GPT-Driven Radiology Report Generation with Fine-Tuned Llama 3

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing in Healthcare
• Radiology Reporting Automation

Background:

• Deep learning integration in radiology faces adoption challenges.
• Radiology reporting efficiency can be improved through AI assistance.
• Automating report conclusions is a key area for AI development in medical imaging.

Purpose of the Study:

• To develop and evaluate a fine-tuned Llama 3-8B large language model (LLM).
• To automate the generation of accurate and concise conclusions for MRI and CT radiology reports.
• To assess the AI-generated conclusions' performance against human radiologists.

Main Methods:

• Fine-tuning the Llama 3-8B model on 15,000 radiology reports using transfer learning, 4-bit quantization, and LoRA.
• Training the model on an NVIDIA RTX 3090 GPU over five epochs.
• Quantitative evaluation using BERTScore, ROUGE, BLEU, and METEOR metrics.
• Qualitative assessment via a Turing-like test with 13 independent radiologists.

Main Results:

• The fine-tuned LLM achieved strong quantitative scores (e.g., BERTScore F1: 0.8054, ROUGE-L F1: 0.4628).
• AI-generated conclusions were preferred over human-written ones in 21.8% of cases during human evaluation.
• The model received an average rating of 3.65/5 from radiologists, demonstrating competitive performance and consistency.

Conclusions:

• The fine-tuned Llama 3-8B model effectively generates accurate and coherent conclusions for radiology reports.
• This AI-driven automation can assist radiologists, reduce workload, and enhance report consistency.
• Further validation and integration into clinical workflows are recommended, addressing limitations like dataset bias and computational needs.