I was not able to reproduce the results promised in the paper in either pytorch or tensorflow. I wrote the authors several months ago and did not hear back.
Unforunately, it looks like the authors have also since deleted a link to the previously released data without explanation. Several researchers have reached out to me for the data so I have re-uploaded it here. Hopefully another will be able to reproduce the results.
Network | Accuracy (encounter) | Precision (encounter) | Recall (encounter) | F1 (encounter) | Kappa (encounter) |
---|---|---|---|---|---|
DenseNet169 (baseline) | .83 (.84) | .82 (.82) | .87 (.90) | .84 (.86) | .65 (.65) |
MobileNet | .81 (.83) | .80 (.82) | .85 (.89) | .82 (.85) | .62 (.62) |
NASNetMobile | .82 (.83) | .78 (.80) | .89 (.92) | .83 (.86) | .63 (.63) |
Also, ResNet50 in pytorch which achieved equivalent results.
The Mura Dataset
@misc{1712.06957,
Author = {Pranav Rajpurkar and Jeremy Irvin and Aarti Bagul and Daisy Ding and Tony Duan and Hershel Mehta and Brandon Yang and Kaylie Zhu and Dillon Laird and Robyn L. Ball and Curtis Langlotz and Katie Shpanskaya and Matthew P. Lungren and Andrew Ng},
Title = {MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs},
Year = {2017},
Eprint = {arXiv:1712.06957},}
Study | Normal | Abnormal | Total |
---|---|---|---|
Elbow | 1,203 | 768 | 1,971 |
Finger | 1,389 | 753 | 2,142 |
Forearm | 677 | 380 | 1,057 |
Hand | 1,613 | 602 | 2,215 |
Humerus | 411 | 367 | 778 |
Shoulder | 1,479 | 1,594 | 3,073 |
Wrist | 2,295 | 1,451 | 3,746 |
Total | 9,067 | 5,915 | 14,982 |
- Each study contains 1-N views (images)
- 40,895 multi-view radiographic images
Radiologists (95% CI) | Model (95% CI) | |
---|---|---|
Elbow | 0.858 (0.707, 0.959) | 0.848 (0.691, 0.955) |
Finger | 0.781 (0.638, 0.871) | 0.792 (0.588, 0.933) |
Forearm | 0.899 (0.804, 0.960) | 0.814 (0.633, 0.942) |
Hand | 0.854 (0.676, 0.958) | 0.858 (0.658, 0.978) |
Humerus | 0.895 (0.774, 0.976) | 0.862 (0.709, 0.968) |
Shoulder | 0.925 (0.811, 0.989) | 0.857 (0.667, 0.974) |
Wrist | 0.958 (0.908, 0.988) | 0.968 (0.889, 1.000) |
Aggregate F1 | 0.884 (0.843, 0.918) | 0.859 (0.804, 0.905) |