fidoptRHPE hand radiographs
Radiographs of both hands of two children of the Fundación Santa Fe de Bogotá, with 17 keypoints on the fingers and wrist of the left-hand side of the image placed by trained annotators on each (Radiological Hand Pose Estimation, test split).
Read before comparing (7)
- Pool. Hand radiographs of children and adolescents at the Fundación Santa Fe de Bogotá, Colombia; the 79 test images are of 38 boys and 41 girls aged 2 to 16 years (29 to 201 months). The provider gives one image an examination and no patient identifier; its post-release curation deleted nine images, none of those used.
- Images. The provider's 8-bit grey PNGs, 1,796 to 3,480 by 1,101 to 3,480 pixels on the test images, both hands side by side as the film shows them, with the side markers and tags. No pixel size is given, so the unit is the pixel. No mask.
- Truth. 17 keypoints on the hand on the image's left, the one the provider takes as non-dominant: three along each finger and two at the wrist. Each is the mean of at least four placements by 96 trained students, after rejecting placements over two standard deviations from the mean (Escobar et al. 2019). No radiologist checked them and no agreement between readers is published; the other hand carries no truth.
- Alignment. Each image keeps its own grid at the origin; the moving image's header carries the translation that puts the two images' centres together. elastix mutual information gave a mean TRE of 263 pixels and one pair 1,267 pixels apart, against 240 and at most 650 from the centres.
- Baseline. With no registration the 39 test pairs start at a TRE of 240 pixels (84 to 650), rTRE 0.070, a 95th percentile of 348 pixels on average.
- Coverage. The points lie on the fingers and the wrist of one hand: nothing scores the carpal bones between them, the forearm or the other hand.
- Overlap. No other dataset of the corpus holds RHPE images: DigitalHandAtlas is a Los Angeles atlas of other children, and the RSNA bone age images RHPE was split like are not in the corpus. RHPE gives no patient identifier, so two images of one child, in the test and validation splits, cannot be ruled out.
No method has a result on every pair of this dataset yet.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- The provider's 79 test images, the whole test split, in the order of sha256("2026:RHPE:<id>") and paired two by two, the first fixed: 39 test pairs, one image left over. 30 development pairs drawn with seed 2026 from the 356 pairs made the same way of the 713 validation images; the training images are not used.
- Truth
- landmarks by experts the dataset's 17 keypoints on each image, three on each finger and two at the wrist, of the hand on the image's left: each the mean of at least four placements by trained students, after outlier rejection (Escobar et al. 2019)
- Source
- bcv-uniandes.github.io/baar-wp
doi.org/10.1007/978-3-030-32226-7_59
github.com/BCV-Uniandes/Bonet - Licence
- Universidad de los Andes and Fundación Santa Fe de Bogotá terms: non-commercial research and educational use only; access passed on only to colleagues bound by the same terms
30 pairs for tuning, never tested
06079_06167, 05618_05815, 05905_05917, 06040_06043, 05587_05599, 05898_05825, 06173_05976, 05856_06011, 05902_06076, 05730_05818, 06084_05652, 05717_05573, 05722_05548, 05651_05791, 05925_05726, 05759_06024, 05792_05794, 05700_05826, 06102_05982, 05992_06179, 05749_05507, 05508_05882, 06153_05938, 06014_05796, 06086_05866, 05855_06033, 05904_06202, 05768_05850, 05536_06164, 05640_05936