English | Français

CHROMA V2

Calibrated Histological Reconstruction Of Multiplexed Absorbances

A stain normalizer for HES slides (Hematoxylin–Eosin–Saffron) that works in optical density and separates brightness from chromaticity before decomposition. Single file, numpy + scikit-learn, no deep learning.

Placenta patches at x20. Left column: native image (inter-slide staining variability). Middle: CHROMA V2. Right: Macenko. CHROMA harmonizes color while staying close to the native image (DINOv2 cosine distance ≈ 0.02–0.03 vs native, against ≈ 0.05–0.09 for Macenko on the same patches) and preserves the pigments and red blood cells that Macenko erases.

Why

HES staining varies from slide to slide (reagents, bath age, scanner). This color drift disturbs both the pathologist's eye and encoders (DINOv2, CTransPath…). Existing normalizers each have a weakness:

Method	Limitation
Macenko (SVD, 2 components H&E)	Ignores saffron, unstable per slide, destroys pigments (meconium, hemosiderin) and red blood cells
Reinhard (Lab)	Separates brightness/chromaticity but in a perceptual space, not a physical one

CHROMA is the synthesis: OD space (physical, additive via Beer-Lambert) + brightness/ chromaticity separation (Reinhard's idea) + a biological 3-stain decomposition.

The physical principle

The 3 HES stains behave like the 3 subtractive CMYK primaries: H ≈ Cyan, E ≈ Magenta, S ≈ Yellow, plus a residual = Key (pigments). Moving to optical density makes the mixture additive:

OD = -log10(I / I0)          (Beer-Lambert law, I0 = bright background per slide)

The V2 novelty: remove the gray axis before decomposing

The gray axis ĝ = [1,1,1]/√3 in OD encodes thickness/brightness variation without color. In V1 it contaminated the saffron channel (saffron would "light up" on gray tissue). V2 removes it before decomposition:

T   (brightness)     = OD · ĝ                  tissue thickness, scalar per pixel
Sat (saturation)     = ‖OD − T·ĝ‖              chromatic intensity
c   (concentrations) = pinv(S⊥) · (OD − T·ĝ), clamped ≥ 0   on chromaticity only
K   (residual)       = OD − (T·ĝ + S⊥·c)       unexplained pigments

where S⊥ = S − ĝ(ĝᵀS) is the stain matrix projected out of the gray axis. Result: brightness is carried by T, never by the stains → saffron no longer captures gray.

7 interpretable channels

RGB → OD ─┬─ T   [tissue thickness]
          ├─ Sat [chromatic intensity]
          └─ chromaticity ─ decomposition ─┬─ H [hematoxylin / nuclei]
                                           ├─ E [eosin / cytoplasm]
                                           ├─ S [saffron / collagen]
                                           └─ K [residual: meconium, hemosiderin, formalin]

Total: H, E, S, K, T, Sat + normalized image

The K channel is left untouched during normalization: this is the key advantage over Macenko, which wipes out these diagnostic pigments. The gray-decontaminated H channel isolates pure basophilia and turns out to be a strong descriptor of necrosis and inflammation.

Automatic calibration

No annotation, no human intervention. We pool tissue pixels (Otsu mask on saturation) from ~10 HES slides, run a 3-component NMF anchored on a canonical HES basis (otherwise NMF is non-identifiable on the HES cone), and assign H/E/S by dominant absorption channel. A single site-specific artifact: a JSON calibration file.

Usage

import numpy as np
from PIL import Image
import chroma

# 1) calibrate a TARGET on ~10 RGB thumbnails of HES slides (once per site/scanner)
target = chroma.auto_calibrate([np.asarray(Image.open(p).convert("RGB")) for p in thumbs])
norm = chroma.CHROMA(target)

# 2) for each slide to normalize: calibrate its SOURCE and build the PIL→PIL function
src = chroma.auto_calibrate([np.asarray(Image.open("slide.png").convert("RGB"))])
stain_fn = chroma.make_stain_fn(src, norm)

normalized = stain_fn(Image.open("patch.png"))     # normalized RGB PIL.Image

# 3) decomposition into 7 channels (2D maps)
channels = norm.decompose(np.asarray(Image.open("patch.png").convert("RGB")))
H, E, S, K, T, Sat = (channels[k] for k in ("H", "E", "S", "K", "T", "Sat"))

stain_fn plugs directly in as a torchvision.transforms.Lambda in an embedding pipeline. Cost < 1 s per 224×224 patch on CPU.

Built-in self-check:

python chroma.py     # gray decontamination, reconstruction ≈ identity, gains OK

Calibration format (JSON)

{
  "method": "NMF_auto_chroma_separated",
  "chroma_version": "2.0",
  "I0": [253.7, 253.6, 253.7],
  "stain_vectors": { "H": [...], "E": [...], "S": [...] },
  "stats": {
    "mu": [...], "sigma": [...],
    "brightness_mu": 0.092, "brightness_sigma": 0.05,
    "sat_mu": 0.053, "sat_sigma": 0.03
  },
  "calib_id": "4d75cb66"
}

Dependencies

numpy, scikit-learn, opencv-python, Pillow (see requirements.txt). No framework, no ORM, no GPU required.

Authors

• Rémi Mathevet — fetal pathologist, Besançon University Hospital (CHU) — design, specification, clinical validation
• Claude (Anthropic) — implementation, gray-separated decomposition, QC tooling

License

Apache-2.0