fidoptRibFrac
A chest-abdomen CT of a patient with rib fractures at Huadong Hospital, Shanghai, with every fracture outlined as a region (RibFrac challenge, training and tuning sets).
30 test casesthorax, abdomenimage CTresearch only
Read before comparing (7)
- Pool. RibFrac's 660 chest-abdomen CTs of patients with traumatic rib fractures (Huadong Hospital, Fudan University, Shanghai) come in a training set of 420, a tuning set of 80 and a test set of 160 whose labels are not public. Of the 500 labelled, 20 tuning CTs list no fracture and are left out, an empty truth giving no Dice.
- Truth. Two junior radiologists (5 years) drew a volume around each fracture with the reports beside them, the broken ends included as much as possible, and a senior radiologist (20 years) confirmed them; a network trained on the first labels then proposed fractures, which the senior radiologist confirmed or not, adding about 20% of the fractures (Jin et al. 2020). Graded semi-automatic: the regions are loose volumes around the break, not outlines of bone, 'noisy' by the challenge's own account, and a fifth of them started as a network's proposal.
- Labels. One label, rib_fracture, for all the label map's fractures; their types, in the deposit's CSV, are not scored. On the test CTs 61% of the fractures are of unknown type (code -1), still fractures. The test CTs hold 2 to 26 fractures (median 9.5, 302 in all), 2.2 to 99 mL of fracture volume (median 24 mL).
- Images. The deposit's NIfTI files byte for byte, in Hounsfield units, slices 1 or 1.25 mm apart, pixels 0.58 to 0.88 mm on the test CTs; half of them read -3024 outside the reconstructed circle.
- Start. A mask of the whole image scores a Dice of 0.0008 at the median of the test CTs, an Otsu mask 0.0014. TotalSegmentator 2.18.0's 24 ribs as one mask score 0.219 on test case RibFrac82, and hold 65% of its fracture voxels: the regions reach past the bone into the soft tissue around the break.
- Overlap. RibSeg and RibSeg_Seg take their CTs from RibFrac: their 118 patients, 91 of them labelled with a fracture, are held out. No other dataset of the corpus names RibFrac or Huadong Hospital.
- Outside numbers. None on these cases: the challenge ranks detection (FROC, a fracture found above an IoU of 0.2) and classification on its hidden test set, not overlap. On FracNet's own test cohort FracNet reached a Dice of 0.715, and two radiologists drawing alone 0.643 and 0.531 (Jin et al. 2020).
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 30 test CTs drawn with seed 2026 among the 359 patients outside development, and 30 development CTs. The pool is the 480 labelled patients whose CSV lists a fracture, less the 91 whose CTs RibSeg pairs. RibFrac's 160 test CTs have no public labels; the draw mixes its training and tuning parts.
- Truth
- contours by a program, verified by people every fracture of the patient's label map under one label, whatever its type (displaced, non-displaced, buckle, segmental, or of unknown type): a region around the broken ends drawn on axial slices with curved planar reformats by two junior radiologists from the diagnostic reports, confirmed by a senior radiologist, and completed by the fractures a first network found that this senior radiologist confirmed (about 20% of them)
- Source
- zenodo.org/records/3893508
zenodo.org/records/3893498
zenodo.org/records/3893496
ribfrac.grand-challenge.org
doi.org/10.1016/j.ebiom.2020.103106 - Licence
- CC BY-NC 4.0
30 cases for tuning, never tested
RibFrac186, RibFrac210, RibFrac319, RibFrac53, RibFrac357, RibFrac74, RibFrac260, RibFrac464, RibFrac306, RibFrac262, RibFrac271, RibFrac2, RibFrac428, RibFrac3, RibFrac247, RibFrac69, RibFrac178, RibFrac45, RibFrac116, RibFrac258, RibFrac398, RibFrac97, RibFrac266, RibFrac401, RibFrac113, RibFrac297, RibFrac362, RibFrac414, RibFrac152, RibFrac424