fidoptTotalSegmentator pleural_pericard_effusion, pericardial effusion
Chest CT of a patient with a pericardial effusion or haemopericardium at the University Hospital Basel, the effusion drawn by hand by a radiology resident under a cardiothoracic radiologist (Wilder-Smith et al. 2022).
15 test casesheartimage CT
Read before comparing (9)
- Deposit. TotalSegmentator's training set for its task pleural_pericard_effusion: 600 Basel chest CTs at 1.5 mm, each with a map of lung, pleural effusion and pericardial effusion. Its README cites Wilder-Smith et al. 2022 for the pericardial effusion and Sexauer et al. 2022 for the pleural one; no source says which CT comes from which study.
- Pool. 301 deposit CTs are CTs of the pericardial study's release (zenodo.org/records/6384747), their corners within 1.25 mm. The 157 with an effusion there carry the same effusion in the deposit. The 144 without one, the study's negative cohort, are not cases: a truth that holds nothing scores nothing.
- Truth. The pericardial effusion only, the one label people are known to have drawn on these CTs: one of two residents traced it on the 1 mm CT under a cardiothoracic radiologist. No source says who drew the deposit's lungs and pleural effusions on these CTs: they are left out, and a method's are not scored.
- Agreement. 40 CTs drawn by both residents agree at a Dice of 0.69 +/- 0.02, and the study's nnU-Net reached 0.75 +/- 0.01. A method should not be expected to agree with the truth much better than the two readers agree.
- Cases. Patients of the study's positive cohort, the effusion more than 4 mm thick, simple or a haemopericardium. On the test cases it holds 38.8 to 410.6 mL in 2 to 101 pieces (thin layers break up at 1.5 mm), at a median value of 2 to 38 HU.
- Revised maps. On test case ABDJqMkDEtKf5n0x the deposit's map holds 215.9 mL where the release holds 137.8 (Dice 0.67), and on development case pr1c8CMWVo9dPty6 18.7 against 22.5 mL. Someone revised them after the study, the deposit does not say who; the map as deposited is used.
- Images. The deposit's files byte for byte: 1.5 mm voxels, Hounsfield units from -1024 once read through the header's intercept. Four Siemens scanners, a soft tissue kernel (30f), CTs cropped to the chest. No mask is given. At 1.5 mm the default surface Dice at 1 mm is strict; the provider sets no tolerance.
- Outside models. TotalSegmentator's pleural_pericard_effusion model was trained on this deposit and has seen every case; so has the study's own nnU-Net. The study's test CTs are not public: no outside number on these cases.
- Overlap. No dataset of the corpus holds these CTs: none of the 1,228 CTs of TotalSegmentator's CT dataset lies within 5 mm of any of the 600 by its corners, and no corpus image with 1.5 mm voxels is one of the 20 cases built first. PleThoraCT's CTs are Maastricht's. A later dataset built from the deposit or the study's release leaves these 29 out.
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
The first test caseClick to look at every case.
- Test cases
- 15 test CTs drawn with seed 2026 among the 143 outside development, and 14 development CTs, the last of the same seeded order. The pool is the 157 deposit CTs that are CTs of the pericardial study's own public release and hold an effusion there (prepare.py --pool); one CT a patient. The provider publishes no test set.
- Truth
- contours by experts the pericardial effusion, simple or bloody, drawn by hand in 3D on the 1 mm CT by one of two radiology residents (postgraduate years 2 and 4) under a board-certified radiologist with 15 years of cardiothoracic imaging, all with one software and one window (Wilder-Smith et al. 2022), delivered resampled to 1.5 mm with its CT; the deposit's lung and pleural effusion labels are left out
- Source
- zenodo.org/records/20272295
github.com/wasserth/TotalSegmentator
doi.org/10.3390/diagnostics12051045
zenodo.org/records/6384747
doi.org/10.1097/RLI.0000000000000869
doi.org/10.1016/j.ejrad.2020.109233 - Licence
- CC BY 4.0
14 cases for tuning, never tested
brrIipWxMcQnbZS0, 8szUoIVdqc9vwsap, Hv8ybw34FKYCFaJN, pr1c8CMWVo9dPty6, dlJG144PXNuiNXxT, RiaKk5DdcR4IVGD1, 5zMoKFNbowWW2F0p, V7FftwWA4CRr7FDE, vVTKvd2wO2MMm2Ct, 5McigRxGQyUNfP0w, sXmZYMxLfNjpPThk, eEB77WeMupLMvQsg, bEB5YdQ9Awbnxh3V, F9IREmgPAC5EUAoe