Staging Prostate Cancer with AI: A Comparative Study of Large Language Models and Expert Interpretation on PSMA...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prostate cancer staging is crucial for treatment decisions.
• Unstructured PSMA PET-CT reports limit large-scale data analysis.
• Large language models (LLMs) show potential for extracting structured data from clinical reports.

Purpose of the Study:

• To assess the feasibility of using LLMs for multi-step clinical reasoning in prostate cancer staging.
• To evaluate the performance of Gemini 2.5 Pro and ChatGPT 4o in staging prostate cancer from PSMA PET-CT reports.

Main Methods:

• 80 Turkish PSMA PET-CT reports were analyzed by Gemini 2.5 Pro and ChatGPT 4o.
• LLMs used structured prompts with embedded knowledge (AJCC/CHAARTED criteria) and few-shot examples.
• Classifications for T, N, M stages, and disease volume were benchmarked against expert staging.

Main Results:

• Gemini 2.5 Pro achieved 93.8% accuracy and 0.910 kappa for composite staging.
• ChatGPT 4o achieved 91.3% accuracy and 0.874 kappa for composite staging.
• Both LLMs exceeded 95% accuracy for N and M staging, demonstrating near-perfect agreement.

Conclusions:

• Expert-guided LLMs can accurately stage prostate cancer from free-text PSMA PET-CT reports.
• LLMs show promise as assistive tools for automating data extraction and accelerating research.
• This approach can enhance quality assurance in oncological data management.