Due to some data errors and missing labels on the VQA-RAD dataset, we further constructed an enhanced dataset, VQA-RADPh, to raise data quality.
We invited medical imaging experts to check the dataset one by one and found that the question-answer pairs on the VQA-RAD dataset were not completely correct. In this case, the questionable data are filtered out from all the question-answer pairs. We discussed and revised these data with experts in multiple medical fields to correct the dataset. And it was found that the 48 question-answer pairs in the test set were not labeled. To improve data quality, we repaired the anomaly data and expanded the candidate answers from 486 to 521.