fidoptPanTrack
Two consecutive portal venous CTs of a patient under treatment for pancreatic adenocarcinoma, the earlier fixed, with the lesions a radiologist segmented and linked between them (PanTrack, one German institution).
Read before comparing (8)
- Pool. 45 patients with pancreatic adenocarcinoma followed by portal venous CT at one German institution on identical Siemens protocols; a case is the patient's first consecutive pair with a lesion linked one to one, the earlier CT fixed. The 20 patients of PanTrackCT are left out.
- Truth. The ranked truth is the provider's TotalSegmentator maps on each CT, a network's view of 117 structures (graded automatic), as in same-patient-automatic's other sets. The landmarks are people's: the centres of the lesions one radiologist segmented and linked between the two CTs (tracking.json, merged lesions left out), equal to the provider's own centres to 0.0 mm on the test pairs; but most pairs hold one (31 in the 10 test pairs, median 1, 9 and 14 on two), so they are shown, not ranked. The lesions are shown as other labels.
- Change. The test pairs are 2 days to 11 months apart (median 52 days) under treatment: a tumour that shrinks or grows moves its centre, so the landmark error has a floor above zero.
- Images. The provider's CTs as delivered, 0.4 to 1.0 mm slices (0.4 mm for all test CTs but one at 1.0 mm) of 512 x 512 pixels of 0.63 to 0.87 mm, -8192 outside the reconstructed field; up to 1,300 slices, so a pair weighs over 1 GB in memory.
- Alignment. The two CTs come from two sessions in scanner coordinates: the start is a translation found by elastix on the bodies (elastix-translation).
- Start. After that start, no registration scores an organ Dice of 0.53 at the median of the test pairs (0.29 to 0.66), an HD95 of 12.2 mm, a lesion-centre error of 10.9 mm (6.3 to 26.9) and a lesion Dice of 0.45.
- Overlap. No other dataset of the corpus holds these CTs; LongitudinalCT is autoPET IV's melanoma cohort.
- Outside numbers. The provider reports lesion tracking with its LongiTrack models, on all 116 pairs, not on this draw.
No method has a result on every pair of this dataset yet.
* shown, not ranked: on this dataset it is checked against the output of a program.
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
- 10 test patients drawn with seed 2026 among the 25 outside PanTrackCT, then 10 development patients, the last of the same seeded order; one pair a patient, the first consecutive pair with a lesion linked one to one. The provider gives only a suggested training split.
- Truth
- contours by a program landmarks by experts lesion by experts the provider's TotalSegmentator maps of each CT (task total, 117 structures), made by a network on each image alone the centres of the lesions linked one to one between the two CTs, each the mean of its voxels in the radiologist's mask of that CT: shown, not ranked
- Source
- huggingface.co/datasets/mrokuss/PanTrack
huggingface.co/datasets/mrokuss/PanTrack/tree/5ce8a560509564148ffe55ad31ebc3da5a89cbfd
arxiv.org/abs/2605.23118
github.com/MIC-DKFZ/LongiSeg - Licence
- CC BY-NC 4.0
10 pairs for tuning, never tested
PanTrack_006, PanTrack_032, PanTrack_011, PanTrack_022, PanTrack_028, PanTrack_008, PanTrack_025, PanTrack_039, PanTrack_033, PanTrack_021