huangyjhust/pure-mura

A pre-operative subset division of MURA dataset for musculoskeletal abnormalty classification

Pure-MURA

A pre-operative sub-set of MURA dataset(https://stanfordmlgroup.github.io/competitions/mura/).

The dataset is avaliable at https://drive.google.com/file/d/140YHuHvjGxUJfvu2vsaKY4H4MR6tJDd_/view?usp=sharing

I noticed that, ~50% of the positive samples are with fracture indicators(steel plates, manual marks or in plaster) that give away their fracture labels. These indicators misguide CNN's understanding of fractures, thus CNNs easily identify fractures from seeing these indicators. However, a clinically valuable fracture recognition system should be capable of recognizing fractures from pre-operative samples rather than post-operative samples.

Therefore, I mannually divided MURA dataset into two subsets, the subset with fracture indicators or pre-operative subset without indicators. By training models with the pre-operative subset, classification should be based on visual clues of fractures, not the indicators. Some post-operative samples may remain in the pre-operative dataset due to mistakes of mannual labour, but they are minority after all.

Below are a few fracture samples from Pure-MURA/train_AllPlate/

Below are a few fracture samples from Pure-MURA/train_NoPlate/