1 · Pick an image
Test images from VQA-Med 2019 that the model never saw during training, or upload your own.
2 · Ask a question
Answer
–
Research demo only. Not for diagnosis or any clinical use.
How well it answers
Exact-match accuracy on the official VQA-Med 2019 test set (500 questions, 125 per category), next to the challenge's best team.
How it works
A compact version of the MediQNet design (originally BioBERT + Swin Transformer), sized to run on a phone: the image and the question are encoded separately, fused, and classified over every answer seen in training. The question's category, detected from its wording, limits which answers are allowed.