OpenMedQ: Broad Open Pretraining for Medical Vision-Language Models

OpenMedQ is a newly released medical vision-language model pretrained on a comprehensive dataset of 3.35 million samples from 14 medical datasets, covering areas such as pathology and radiology. It achieves state-of-the-art BLEU-1 scores of 75.9 on PathVQA, outperforming larger models like Med-PaLM M, and achieves the highest average macro-F1 score of 0.757 on eight unseen medical classification benchmarks, surpassing other models like BiomedCLIP. This model and its accompanying code provide a reproducible baseline for practitioners in the medical AI domain, enhancing capabilities in medical image understanding and clinical question answering.