Can chatGPT-4o reliably standardize PSMA PET/CT and PET/MRI reports using PROMISE V2 criteria? - An exploratory...

Area of Science:

• Radiology and Nuclear Medicine
• Artificial Intelligence in Healthcare
• Oncology Imaging

Background:

• Structured reporting enhances communication and decision-making in medical imaging.
• Prostate-specific membrane antigen positron emission tomography (PSMA PET) is crucial for prostate cancer management.
• Despite standardized criteria (PROMISE), free-text reports are common, hindering data extraction.

Purpose of the Study:

• To evaluate ChatGPT-4o's performance in extracting PROMISE V2 classifications from unstructured PSMA PET/CT and PET/MRI reports.
• To assess the potential of large language models (LLMs) for efficient, structured reporting in PSMA PET imaging.

Main Methods:

• Utilized ChatGPT-4o to analyze free-text PSMA-PET/CT and PET/MRI reports.
• Extracted PROMISE V2-based miTNM and PRIMARY score classifications.
• Compared classification accuracy between PET/CT and PET/MRI modalities.
• Assessed the plausibility of ChatGPT-4o's classification rationale.

Main Results:

• PSMA-PET/MRI achieved higher overall miTNM accuracy (91.0%) compared to PSMA-PET/CT (79.8%).
• PET/MRI significantly outperformed PET/CT in T-stage (83.8% vs. 57.7%) and N-stage (100% vs. 85.9%) classification.
• M-stage and PRIMARY score classifications were comparable between modalities.
• ChatGPT-4o's rationale for classifications received high plausibility ratings (min. Likert ≥ 4.1).

Conclusions:

• ChatGPT-4o reliably extracts PROMISE V2 N- and M-stage classifications from PSMA PET reports.
• T-stage classification accuracy remains a limitation for LLM-based extraction.
• This study is a foundational step towards using LLMs for structured and efficient PSMA PET reporting.