fidoptColorectal cancer TMA cores, H&E to CD44
One tissue-microarray core of a colorectal cancer, two sections of it scanned whole at 40x: haematoxylin and eosin (fixed) and CD44 immunohistochemistry (moving), with the provider's automatic epithelium mask on each.
10 test pairscolorectal cancerserial sections of the same corefixed H&E, moving CD44 IHC2D, distances in pxdelivered aligned
Read before comparing (9)
- Sections. Two series of six consecutive 4 µm sections were cut from each TMA block, and the deposit does not say which stain went where. A core's H&E and CD44 sections therefore lie an unknown 4 to 20 µm apart within one series, more across the two.
- Screening. All 247 cancer cores held in both stains were compared side by side: 46 show the same tissue by a distinctive feature at the same place (a cleft, a lumen, a mucin pool, a stroma band, a lymphoid aggregate, a cut edge), 56 visibly do not, 145 are unclear. One person judged on 300-pixel reductions, which cannot tell how far apart the sections are.
- Images. The provider's JPEGs, RGB, 7,224 to 10,000 pixels a side, with no pixel size: distances are in pixels, about 0.25 µm at 40x, which the deposit does not state. A method that reads no OME-Zarr gets level 2, about 2,400 pixels a side.
- Alignment. None: the pairs are used as delivered. On the ten test pairs that gives an epithelium Dice of 0.524 (0.261 to 0.647), within 0.03 of an elastix translation (0.526) and of the image centres put together (0.523) on every pair. A turn of up to 2.6 degrees is left to the methods.
- Baseline. With no registration the 10 test pairs score a median HD95 of 686 px (234 to 1403), large because each side has epithelium fragments the other lacks. On the 10 development pairs, Dice at level 4 is 0.17 to 0.82.
- Room. ANTs' SyN, the corpus's baseline, ends at a Dice of 0.532 against 0.530 at its former start, and no shift of the CD44 mask within 64 px raises Dice by more than 0.012 on any test pair. Mask areas differ by up to 1.5 times (H&E/CD44 0.66 to 1.30): glands change between sections, and two networks draw on two stains.
- Similarity. On A040-4 the two grey images start at a cross-correlation of 0.29, so SyN's field stays small and ends at a Dice of 0.669. With mutual information the Dice reaches 0.688, with cross-correlation on the haematoxylin channel 0.675.
- Truth. The provider's epithelium masks, from DeepLabV3 networks trained in NoCodeSeg on a pathologist's annotations, one for H&E and ISH, another for IHC. Ranked apart: automatic labels from two networks, unknown section distance, and a start close to the reachable Dice. The score measures how well large structures are kept, not cell-level correspondence.
- Overlap. No other dataset of the corpus reads this deposit (DIGQGQ and dataverse.no searched in every descriptor). Patients are anonymised A-numbers, with nothing else to link them.
No method has a result on every pair of this dataset yet.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 px | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| No registration | 0.524 ±0.082 | 730.70 ±274.47 | 0.000 | 0.00 | ||||||
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
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
The first test caseClick to look at every case.
- Test pairs
- 10 test cases drawn with seed 2026, then 10 development cases among the others, from a pool of 34 patients: one core a patient, the lowest-numbered of those whose two sections visibly show the same tissue (screen.py, screen.json). No patient is in both.
- Truth
- contours by a program The provider's binary epithelium masks, made by DeepLabV3 networks trained in the NoCodeSeg workflow on a pathologist's annotations, one network for H&E and ISH slides and another for IHC slides (the deposit's README): automatic, and from two different networks on the two sides.
- Licence
- CC0 1.0, DataverseNO record and release README
10 pairs for tuning, never tested
A067-6, A051-5, A047-4, A088-6, A037-4, A084-4, A097-4, A080-5, A079-5, A033-5