MagniSynth: paired low/high-magnification microscope crops

Pairs of (low-magnification crop, real high-magnification capture) of the same physical region, for training models that predict a higher-magnification image from a lower-magnification one. All images are 8-bit grayscale; captures are 5120x3840 at every magnification, so a 5x -> 20x pair is a x4 upscaling task.

Contents

• raw/ source captures (5x, 10x, 20x), original folder layout.
• crops/ the region of the lower-magnification image matching each high-magnification capture, at native low-magnification pixel density.
• pairs.jsonl one row per pair (143 rows, 114 accepted by a reviewer).

Loading

from datasets import load_dataset
from huggingface_hub import hf_hub_download

pairs = load_dataset("duyle240820/magnisynth", split="train")   # one row per pair
good = pairs.filter(lambda r: r["decision"] == "accept" and r["quality_flag"] == "ok")
row = good[0]
crop = hf_hub_download("duyle240820/magnisynth", row["crop_path"], repo_type="dataset")   # low-mag crop
high = hf_hub_download("duyle240820/magnisynth", row["high_image"], repo_type="dataset")  # real high-mag capture

Image paths in pairs.jsonl are relative to the repository root.

How to use pairs.jsonl

decision is a human review outcome: accept, regenerate or manual_needed (null = not reviewed). Train on accept rows. label_source is manual_anchor (a reviewer located the crop by hand) or auto_reviewed (the matcher proposed it and a reviewer accepted); manual anchors are stronger evidence. Reviewer identities and free-text notes are not included.

quality_flag is an automatic sanity score, not a guarantee of alignment: correlation between the high-magnification image (blurred, reduced to the crop's size) and the crop. low_correlation is below 0.3, weak_correlation below 0.5. Some accepted pairs score near zero and should be audited or excluded. padded crops extend past the low-magnification frame (background-filled) and should be masked in any loss.

Quality flags in this release:

• low_correlation: 9
• ok: 104
• padded: 15
• weak_correlation: 15

License

No license has been granted: all rights reserved by the dataset's owners. You may not copy, redistribute or build on these images without their permission.

Limitations

• Very small: two specimens. Split by specimen (leave-one-specimen-out), never by random patch -- neighbouring captures overlap physically and one high-magnification image appears in several pairs.
• Crops are integer-pixel slices of the low-magnification image, so even a perfect match is up to 0.5 low-mag pixel (2 high-mag pixels at x4) off.
• The low-magnification crop is a separate optical capture, not a downsample of the high-magnification image; it is softer and its noise differs.
• Models trained on this can invent plausible detail. Do not use outputs for measurement without checking against real captures.