BioMed-VITAL: A Clinician-Aligned AI Framework for Biomedical Visual Instruction Tuning

BioMed-VITAL Framework: Practical Solutions and Value Enhancing Biomedical Visual Instruction Tuning Recent AI advancements like GPT-4V excel in various tasks but need specialized datasets for fields like biomedicine. BioMed-VITAL incorporates clinician preferences to create top-notch data for these models. Improving Model Performance BioMed-VITAL significantly enhances model performance, achieving an 18.5% boost in open visual chat and an 81.73% win rate in biomedical visual question answering. This framework merges expert knowledge with model training for superior results. Data-Centric Approach Through data generation, selection, and instruction tuning, BioMed-VITAL fine-tunes models effectively for specialized biomedical tasks. By involving clinician expertise, it boosts model performance and accuracy. Alignment with Clinician Preferences BioMed-VITAL creates high-quality datasets by aligning closely with clinician preferences. This alignment ensures the model learns from relevant, expert-approved data, leading to better outcomes in medical visual chat and question answering tasks.