fidoptORION colorectal cancer, H&E to Hoechst of one section
A 2048-pixel square of one colorectal cancer section at 0.325 µm a pixel: its H&E (source) and the Hoechst channel of the immunofluorescence image taken before staining (target), from the ORION study (Nature Cancer 2023).
10 test casescolorectal tissuesame sectionsource H&E, target Hoechst2D, distances in µmresearch only
Read before comparing (8)
- Registration. The two images are the provider's pixels on one grid: the authors registered the H&E onto the fluorescence image (palom), and nothing is resampled. By the phase correlation of the two nuclear stains over tiles of 256 pixels, that registration is off by 0.2 to 0.6 µm a case, 1.2 µm on CRC05.
- Square. Where the section changed between the two scans the registration fails, so the square is drawn in the H&E's tissue until the nuclei of the two stains agree: the first draw holds for 19 specimens, CRC05's second is kept.
- Target. Hoechst is one of the image's 19 channels. It marks the nuclei that the H&E's haematoxylin also shows: what is to be made is in the source.
- Mask. The H&E's tissue, found automatically by colour (max minus min over red, green and blue above 20, mean under 235), not a pathologist's outline.
- Range. truth.dynamic_range, 15686, is the median of the ten test targets' own 99.9th percentiles inside the mask (10,242 to 24,201 counts), the rule of BraTS2021_Syn. No target reaches the 16-bit ceiling (at most 49,953).
- Level. The target is in the scanner's counts, kept as stored, and its level is a matter of staining and acquisition the H&E does not show: its median inside the mask runs from 404 (CRC09) to 5,227 (CRC36) on the test cases. Part of any method's error is that level, and the MAE's mean weighs the brighter specimens most.
- Baseline. Inside the mask of the test cases, with KonfAI's criteria, a constant at the case's own median, an oracle that knows the level, scores an MAE of 1,885 counts (496 to 3,516), a PSNR of 15.0 dB (10.5 to 21.6) and an SSIM of 0.49 (0.30 to 0.80); a zero image an MAE of 875 to 5,875.
- Overlap. No other descriptor names the ORION study, its bucket or its paper, and OrionCRC's specimens are held out, so neither this target nor another square of the same section is shown there.
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the target image, and over it the method's synthetic image. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 10 test specimens drawn with seed 2026 and the other 10 for development, one case a specimen. The pool is the 20 of the public bucket's 40 specimens (CRC01 to CRC40) that OrionCRC does not hold: its 20 are held out, since its moving image is their Hoechst under a known warp. The study gives no patient key, so a specimen is taken as a patient.
- Truth
- target image as released the Hoechst channel of the specimen's immunofluorescence image as stored (16-bit counts), on the H&E's grid through the authors' own registration of the H&E onto it (palom), off by 0.2 to 0.6 µm a case (median over tiles of 256 pixels) on 19 cases and 1.2 µm on CRC05
- Source
- www.tissue-atlas.org/atlas-datasets/lin-chen-campton-2023
doi.org/10.1038/s43018-023-00576-1
zenodo.org/records/7637988 - Licence
- None stated for the images, on the bucket, the atlas page or the repository's README. The companion record (Zenodo 7637988, MIT) covers the index and the code, not the images.
10 cases for tuning, never tested
CRC11, CRC20, CRC15, CRC02, CRC32, CRC08, CRC16, CRC13, CRC25, CRC17