Abstract
OBJECTIVES: to assess the ability of high-performing open-weight large language models (LLMs) in extracting key radiological features from prostate MRI reports. METHODS: Five LLMs (Llama3.3, DeepSeek-R1-Llama3.3, Phi4, Gemma-2, and Qwen2.5-14B) were used to analyze free-text MRI reports retrieved from clinical practice. Each LLM processed reports three times using specialized prompts to extract (1) dimensions, (2) volume and PSA density, and (3) lesion characteristics. An experienced radiologist manually annotated the dataset, defining entities (Exam) and sub-entities (Lesion, Dimension). Feature- and physician-level performance were then assessed. RESULTS: 250 MRI exams reported by 7 radiologists were analyzed by the LLMs. Feature-level performances showed that DeepSeek-R1-Llama3.3 exhibited the highest average score (98.6% ± 2.1%), followed by Phi4 (98.1% ± 2.2%), Llama3.3 (98.0% ± 3.0%), Qwen2.5 (97.5% ± 3.9%), and Gemma2 (96.0% ± 3.4%). All models excelled in extracting PSA density (100%) and volume (≥98.4%), while lesions' extraction showed greater variability (88.4-94.0%). LLMs' performance varied among radiologists: Physician B's reports yielded the highest mean score (99.9% ± 0.2%), while Physician C's resulted in the lowest (94.4% ± 2.3%). CONCLUSIONS: LLMs showed promising results in automated feature-extraction from radiology reports, with DeepSeek-R1-Llama3.3 achieving the highest overall score. These models can improve clinical workflows by structuring unstructured medical text. However, a preliminary analysis of reporting styles is necessary to identify potential challenges and optimize prompt design to better align with individual physician reporting styles. This approach can further enhance the robustness and adaptability of LLM-driven clinical data extraction.