One PNG carries the MRI and its AI mask
Research demo, not for diagnosis. / 研究用デモであり、診断用ではありません。 ← Demos / デモ一覧
Reversible data hiding stores a compressed segmentation mask in the pixels of an 8-bit PNG. This walkthrough uses a pre-embedded public lumbar T2 slice. Nothing here needs a network. Embedding is not run live (median 27.9 s per image on CPU in the mid-slice record).
Open the pre-embedded PNG
This file looks like an ordinary 8-bit MRI PNG (512×640). The payload lives in the pixels, not in sidecar tags. Embedded payload: 14048 bits (lzma).
Extract the mask and overlay it
The mask was read from the PNG pixels only, then overlaid in the same colors as the existing figure: vertebra red, canal green, disc blue.
Restore the original image
After extraction, the scan is restored bit-exactly. Exact match: 327,680 / 327,680 pixels
Change one pixel
One pixel at row 256, column 320 was changed from 187 to 188. Then extraction was run again on that altered copy.
Intact PNG: extract and match succeeded. Altered PNG: Extraction still finished, and the mask still matched (0 labels differed). The restored image did not: 1 / 327,680 pixels differed.
This is one slice and one pixel. Whether other locations behave the same is not yet checked.
Keys: 1–4 or arrows. This file is offline and does not upload the image.