A generalist vision-language foundation model for diverse biomedical tasks.

Area of Science:

• Biomedical Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Traditional AI models in biomedicine are often task-specific, limiting their real-world application and holistic data utilization.
• Generalist AI offers versatility but existing solutions are typically resource-intensive and proprietary.
• There is a need for accessible, adaptable AI tools in the biomedical field.

Purpose of the Study:

• To introduce BiomedGPT, an open-source, lightweight, generalist vision-language foundation model for diverse biomedical tasks.
• To evaluate BiomedGPT's performance in radiology, report generation, and summarization.
• To demonstrate the efficacy of diverse data training for practical biomedical AI.

Main Methods:

• Development of BiomedGPT, an open-source, lightweight vision-language foundation model.
• Training with diverse biomedical data to enable generalist capabilities.
• Experimental validation across 25 tasks, including radiology visual question answering, report generation, and summarization.
• Human evaluations for assessing performance in clinical applications.

Main Results:

• BiomedGPT achieved state-of-the-art results in 16 out of 25 experiments.
• Demonstrated a low error rate (3.8%) in radiology visual question answering.
• Showed satisfactory performance (8.3% error rate) in generating complex radiology reports.
• Achieved competitive summarization capabilities with high human expert preference.

Conclusions:

• BiomedGPT represents a significant advancement as the first open-source, lightweight generalist biomedical AI.
• Effective training with diverse data yields practical AI solutions for improving medical diagnosis and operational efficiency.
• The model's performance highlights the potential of generalist AI in transforming biomedical applications.