M4Image forensicsMicroscopyLayer 1 (Deterministic)

Bleed-Through

Reads the correlation between the colour channels of a fluorescence micrograph. A genuine acquisition shows moderate spectral bleed-through, so channels are neither perfectly independent nor identical; near-zero correlation suggests channels assembled separately and near-one suggests one channel copied into another. Channels with no signal are ignored.

Technical description

Splits the image into red, green, and blue channels, tiles each into 16 by 16 blocks, and computes the Pearson correlation of every channel pair in each block, r = sum((a - a_bar)(b - b_bar)) / sqrt(sum((a - a_bar)^2) sum((b - b_bar)^2)). The mean absolute correlation per pair is mapped to a score: the natural band 0.1 <= c <= 0.95 scores 0, c below 0.1 scores 5(1 - c/0.1) toward independence, and c above 0.95 scores 5(c - 0.95)/0.05 toward a copy. A channel whose standard deviation is below 1.0 carries no signal and is excluded, since bleed-through is undefined against an empty channel. The final score averages the scorable pairs.

How it works

Layer 1 (deterministic): computes per-block Pearson correlations for each channel pair, averages the absolute correlation per pair, and scores only pairs whose two channels both carry signal. A near-zero correlation scores toward independence and a near-one correlation toward a copy, while the natural band scores 0; the final score averages the scorable pairs and is 0 when fewer than two channels are active. Blocks outside the natural band are reported with bounding boxes, labelled independent or copied.

Why this matters

Multi-channel fluorescence figures are a common vehicle for fabrication because a reader cannot judge by eye whether the channels belong together. Two fluorophores on one specimen share the sample and overlapping emission spectra, and the filters cannot perfectly separate them, so genuine channels show a moderate, spatially structured correlation called spectral bleed-through. Channels generated or sourced independently lack that shared origin and tend to zero correlation, and a channel duplicated into another approaches correlation one; both extremes are physically implausible for a single acquisition.

Score thresholds

0-1: Channel correlations sit in the natural bleed-through band, consistent with a genuine multi-channel acquisition
2-3: One channel pair is near-independent or near-identical, a possible assembly or duplication or a genuinely separated label
4-5: A channel pair is essentially independent or essentially copied, consistent with channels assembled separately or one duplicated into another

Limitations

Fluorophore correlation is biology-dependent, so the low-correlation cue is the weaker one: two genuinely separated labels, such as a nuclear stain and a sparse membrane marker, legitimately show near-zero correlation, so a low score is suggestive, not conclusive. The high-correlation cue is firmer because near-identical channels rarely arise from independent fluorophores. The 16-pixel block bounds resolution and the absolute-correlation average is upward-biased on small blocks. The screen reads rendered RGB rather than the original acquisition channels, so unusual fluorophore-to-RGB mappings or more than three channels are only partially represented. Inactive channels are excluded. Field-of-view reuse across panels is indicator M2.