New multimodal large language models (LLMs) like OpenAI o3 and Gemini 2.5 Pro demonstrated significant advancements in answering Japanese radiology board exam questions, particularly with image input.
Key Details
- 1Eight LLMs were tested on the Japan Diagnostic Radiology Board Examination (JDRBE).
- 2OpenAI o3 achieved 67% accuracy (text-only) and 72% with image input.
- 3Gemini 2.5 Pro also showed notable accuracy improvements with image data.
- 4Both OpenAI o3 and Gemini 2.5 Pro received higher legitimacy scores from radiologist raters than some competitors.
- 5The test set included 233 questions and 477 images (184 CT, 159 MRI, 15 x-ray, 90 nuclear medicine).
- 6Image input statistically improved diagnostic accuracy for several models.
Why It Matters
This study marks the first demonstration of statistically significant improvement in LLM diagnostic accuracy with image input on a radiology board exam, signaling meaningful progress for AI-assisted radiological training and assessment.
Source
AuntMinnie
Related News
•AuntMinnie
Radiology Receives Declining Share of Industry Research Funding
Radiologists received only 1.1% of industry-funded research payments in 2024, with a continuing downward trend.
•AuntMinnie
GPT-4o AI Matches Radiologists in Follow-Up Imaging Recommendations
GPT-4o matched the performance of experienced radiologists and surpassed residents in recommending follow-up imaging from routine radiology reports.
•Cardiovascular Business
AI Leverages Head CTs for Automated Heart Risk Assessments
AI models can turn routine head CT scans into automated cardiovascular risk assessments, expanding the utility of radiology studies.