Macenko Stain Normalization in Histopathology

• Macenko stain normalization is a template-based method that models H&E slides in optical density space using PCA/SVD and the Beer–Lambert law.
• It standardizes histopathology images by estimating stain vectors and non-negative concentrations, facilitating improved cross-dataset tissue classification.
• Its effectiveness is task-dependent: while it boosts classification accuracy, it may hinder segmentation performance due to boundary sensitivity and over-normalization.

Macenko stain normalization is a template-based color standardization method for histopathology images, principally used for hematoxylin and eosin (H&E) slides to reduce inter-laboratory, inter-batch, and inter-scanner stain variability. In its canonical form, it maps RGB intensities into optical density (OD) space, estimates two dominant stain vectors by PCA/SVD, solves for non-negative stain concentrations under a Beer–Lambert model, rescales those concentrations to match a target template, and reconstructs a normalized RGB image. Within computational pathology, it is widely treated as a physics-inspired normalization procedure for H&E data, but later work shows that its downstream utility is strongly task-dependent: it can substantially improve cross-dataset tissue classification, yet it can be neutral or harmful for boundary-sensitive segmentation workloads (Ciompi et al., 2017).

1. Definition and mathematical basis

Macenko normalization assumes that H&E appearance can be modeled in OD space as a low-dimensional linear mixture of two dominant stains, hematoxylin and eosin. The canonical OD transform is

$$\mathrm{OD} = -\log\!\left(\frac{I + \epsilon}{I_0}\right),$$

with typical choices $I_0 = 255$ and $\epsilon \in [0,1]$, often $\epsilon = 1$. A later segmentation study used the OD definition $I_{OD} = -\log_{10}(I)$, reflecting an implementation-dependent variant of the same absorbance-domain representation (Yeung et al., 2023).

Under the Beer–Lambert model, OD is approximated as a linear combination of stain contributions. In the two-stain H&E setting, this is written as

$\mathrm{OD} \approx S\,C,$

where $S \in \mathbb{R}^{3 \times 2}$ is the stain matrix and $C \in \mathbb{R}^{2}$ or $\mathbb{R}^{2 \times N}$ contains non-negative stain concentrations. The method is therefore both a color standardization procedure and a stain-separation model: it estimates a slide-specific stain basis, then replaces that basis and associated concentration scale with those of a target template (Ciompi et al., 2017).

Conceptually, Macenko differs from direct color-statistics matching. The 2017 colorectal study contrasts it with a stain-specific standardization method by Ehteshami Bejnordi et al. that uses a template-derived look-up table, whereas Macenko is described there as physics-inspired, OD-domain, unsupervised, and reliant on PCA/SVD plus robust percentile scaling. The same study also notes a key assumption: two dominant stains and approximately linear Beer–Lambert behavior (Ciompi et al., 2017).

2. Canonical algorithmic pipeline

The standard pipeline begins by discarding background or nearly white pixels, since stain estimation is unstable when stain concentration is negligible. A common criterion is to retain only pixels satisfying

$\lVert \mathrm{OD} \rVert_2 > t_{\mathrm{OD}},$

with typical $I_0 = 255$0, often approximately $I_0 = 255$1. This background handling is important because low-OD pixels can skew stain-vector estimation and introduce scanner-related artifacts (Ciompi et al., 2017).

The retained OD vectors are then stacked into $I_0 = 255$2, and PCA/SVD is used to estimate the two-dimensional stain subspace. With $I_0 = 255$3 denoting the first two left singular vectors, the projection is

$I_0 = 255$4

For each projected vector $I_0 = 255$5, an angle is computed as $I_0 = 255$6. The two stain directions are then obtained from percentile extremes of the angular distribution, with typical $I_0 = 255$7 and $I_0 = 255$8. These extreme directions are mapped back to 3D OD space, normalized to unit length, and arranged as the columns of the stain matrix $I_0 = 255$9 (Ciompi et al., 2017).

Per-pixel stain concentrations are then estimated by non-negative least squares,

$\epsilon \in [0,1]$0

or by a pseudoinverse followed by a non-negativity clamp, $\epsilon \in [0,1]$1. To normalize to a template, the target stain matrix $\epsilon \in [0,1]$2 is estimated from a reference slide. Concentrations may then be rescaled using robust maxima or upper percentiles, commonly $\epsilon \in [0,1]$3, before OD is recomposed as

$\epsilon \in [0,1]$4

The final RGB image is recovered with

$\epsilon \in [0,1]$5

followed by clamping to $\epsilon \in [0,1]$6 (Ciompi et al., 2017).

Several implementation details recur across studies. Stain vectors are constrained to be unit-norm and non-negative, NNLS enforces non-negative concentrations, and OD thresholding is used for background removal. Patch-based use is common, but whole-slide normalization is also used; in the colorectal classification study, normalization was applied to $\epsilon \in [0,1]$7 patches and also illustrated on $\epsilon \in [0,1]$8 tiles (Ciompi et al., 2017).

3. Template dependence and implementation practice

Macenko normalization is fundamentally template-dependent. In the colorectal classification study, a representative rectal cancer whole-slide image was selected as the template from which target stain parameters were derived; both the stain-specific standardization baseline and Macenko were applied relative to that reference. The practical guidance given there recommends choosing a high-quality representative slide with broad tissue coverage and good staining quality, while avoiding artifacts, heavy background, and atypical staining (Ciompi et al., 2017).

Typical practical settings reported for Macenko include $\epsilon \in [0,1]$9, $\epsilon = 1$0, $\epsilon = 1$1, angular percentiles near $\epsilon = 1$2 and $\epsilon = 1$3, and concentration scaling near the $\epsilon = 1$4 percentile. The same source recommends applying normalization to both training and test data, applying stain normalization before data augmentation, and ensuring that patches or tiles contain enough tissue pixels for robust stain estimation; alternatively, stain vectors may be estimated once per slide and reused per patch. It also warns that too low an OD threshold includes background, whereas too high a threshold can exclude lightly stained tissue (Ciompi et al., 2017).

Later benchmarking work reinforces the centrality of template choice. In a multicenter study spanning 66 laboratories across 11 countries, identical unstained sections from skin, kidney, and colon blocks were stained locally and then digitized on the same Hamamatsu NanoZoomer-XR at $\epsilon = 1$5, $\epsilon = 1$6pixel; the reference whole-slide image for each organ was chosen computationally as the sample whose red-to-blue mean intensity ratio was closest to 1. Macenko, Reinhard, and Vahadane were then applied at whole-slide level on WSIs resampled to $\epsilon = 1$7, with the authors noting that Reinhard and Macenko “seem to benefit” from whole-WSI normalization (Khan et al., 23 Jun 2025).

A recent abstract on multi-target stain normalization frames this single-reference dependence as a limitation of “traditional staining normalization approaches, e.g. Macenko,” and proposes using multiple reference images to improve robustness against diverse staining patterns in external data (Ivanov et al., 2024). This suggests that template choice is not merely an implementation detail but a structural property of the method.

4. Role in colorectal tissue classification

The 2017 colorectal study provides a direct empirical demonstration of Macenko normalization in a ConvNet pipeline for tissue classification. The rectal cancer cohort comprised 74 H&E whole-slide images from 74 patients, scanned at $\epsilon = 1$8 magnification with pixel resolution $\epsilon = 1$9, with nine annotated tissue classes: tumor, stroma, necrosis, muscle, healthy epithelium, fatty tissue, lymphocytes, mucus, and blood. The network was a fully convolutional 11-layer architecture with $I_{OD} = -\log_{10}(I)$0 RGB input, trained by 5-fold cross-validation with ADAM, categorical cross-entropy, balanced batch sampling, and rotation augmentation at $I_{OD} = -\log_{10}(I)$1, $I_{OD} = -\log_{10}(I)$2, and $I_{OD} = -\log_{10}(I)$3. On this rectal cancer task, the reported overall 9-class accuracy was $I_{OD} = -\log_{10}(I)$4 (Ciompi et al., 2017).

The central finding concerns cross-dataset generalization from the rectal cancer cohort to the public colorectal dataset of Kather et al. Without stain normalization, six-class mapped cross-dataset accuracy was $I_{OD} = -\log_{10}(I)$5. Applying stain normalization only at test time improved accuracy substantially: $I_{OD} = -\log_{10}(I)$6 with the stain-specific method and $I_{OD} = -\log_{10}(I)$7 with Macenko. The best reported configuration was stain normalization of both training and test data with the stain-specific method, reaching $I_{OD} = -\log_{10}(I)$8, whereas normalizing training data only and leaving test data unnormalized reduced accuracy to $I_{OD} = -\log_{10}(I)$9 (Ciompi et al., 2017).

Within that study, Macenko therefore functions as an effective domain-alignment baseline. The reported improvement from $\mathrm{OD} \approx S\,C,$0 to $\mathrm{OD} \approx S\,C,$1 under test-time Macenko normalization is large, though still slightly below the alternative stain-specific normalization. The paper further notes qualitative failure modes: Macenko sometimes “forced” colors, tended to suppress red coloration in blood cells, and failed on three adipose patches. The same work therefore presents a nuanced view: Macenko can materially improve cross-dataset classification, yet its OD-based remapping can over-correct diagnostically relevant chromatic cues (Ciompi et al., 2017).

5. Limits, failure modes, and task dependence

A common misconception is that stain normalization is uniformly beneficial across computational pathology tasks. The 2023 segmentation study directly challenges that view by evaluating Macenko normalization, via StainTools defaults, as one of several normalization baselines for instance segmentation on both Masson’s trichrome cardiomyocyte data and H&E nuclei data. Metrics were instance-level $\mathrm{OD} \approx S\,C,$2 and $\mathrm{OD} \approx S\,C,$3, reported with 95% bootstrap confidence intervals and Wilcoxon signed-rank significance testing (Yeung et al., 2023).

On Masson’s trichrome cell segmentation, the baseline achieved overall $\mathrm{OD} \approx S\,C,$4 and $\mathrm{OD} \approx S\,C,$5, whereas Macenko achieved $\mathrm{OD} \approx S\,C,$6 and $\mathrm{OD} \approx S\,C,$7. The study reports that Macenko was significantly worse than baseline for both $\mathrm{OD} \approx S\,C,$8 $\mathrm{OD} \approx S\,C,$9 and $S \in \mathbb{R}^{3 \times 2}$0 $S \in \mathbb{R}^{3 \times 2}$1. On H&E nuclei segmentation trained on DigestPath and tested across CoNSeP, CRAG, GlaS, and PanNuke, the baseline overall performance was $S \in \mathbb{R}^{3 \times 2}$2 and $S \in \mathbb{R}^{3 \times 2}$3, while Macenko yielded $S \in \mathbb{R}^{3 \times 2}$4 and $S \in \mathbb{R}^{3 \times 2}$5; all normalization approaches, including Macenko, were significantly worse than baseline with all $S \in \mathbb{R}^{3 \times 2}$6 (Yeung et al., 2023).

The explanation offered in that work is task-specific. Normalization can introduce small structural changes or artifacts, which are particularly harmful for instance segmentation because segmentation requires precise object boundaries and pixel-level accuracy. The study also argues that standardizing to a single template can itself impose a domain shift relative to naturally diverse staining, thereby reducing exposure to useful real-world variation. In that framework, Macenko remained valuable not as an inference-time normalization transform but as a decomposition tool: the proposed stain-specific augmentation extracted per-image stain color matrices via Macenko and used them to learn realistic stain distributions for training-time augmentation (Yeung et al., 2023).

The 2017 colorectal classification paper identifies related but more localized failure modes. It notes over-normalization, possible suppression of genuine chromatic cues such as red blood cells, and instability on certain adipose patches. It also recommends careful template selection, verification of stain vectors, scanner-specific adaptation when necessary, and pre-filtering of pen marks, folds, and out-of-focus regions. This suggests that Macenko’s behavior depends not only on task class but also on the stability of stain-vector estimation under scanner, tissue, and artifact conditions (Ciompi et al., 2017).

6. Benchmarking, extensions, and contemporary uses

The 2025 multicenter benchmark offers a broad comparative picture of Macenko under controlled staining variability. Because the tissues were cut from the same blocks and scanned on the same device, the dataset isolated staining variation while minimizing biological and acquisition confounders. In that setting, Macenko showed acceptable structural preservation but weaker color-transfer fidelity than the best-performing alternatives. For skin, Macenko yielded histogram intersection $S \in \mathbb{R}^{3 \times 2}$7, PCC $S \in \mathbb{R}^{3 \times 2}$8, Euclidean distance $S \in \mathbb{R}^{3 \times 2}$9, JS divergence $C \in \mathbb{R}^{2}$0, FID $C \in \mathbb{R}^{2}$1, and SSIM $C \in \mathbb{R}^{2}$2. For kidney, the corresponding values were $C \in \mathbb{R}^{2}$3, $C \in \mathbb{R}^{2}$4, $C \in \mathbb{R}^{2}$5, $C \in \mathbb{R}^{2}$6, $C \in \mathbb{R}^{2}$7, and $C \in \mathbb{R}^{2}$8. For colon, they were $C \in \mathbb{R}^{2}$9, $\mathbb{R}^{2 \times N}$0, $\mathbb{R}^{2 \times N}$1, $\mathbb{R}^{2 \times N}$2, $\mathbb{R}^{2 \times N}$3, and $\mathbb{R}^{2 \times N}$4. The study describes Macenko as mid-tier to lower on color transfer and FID, and reports qualitative blue artifacts in erythrocytes and lamina propria, especially in colon, while histogram matching and Reinhard were more robust in color fidelity (Khan et al., 23 Jun 2025).

At the same time, recent work uses Macenko less as a terminal normalization endpoint and more as a structured parameterization of stain variability. A 2026 paper on Calibrated Adversarial Stain Augmentation operates in the Macenko stain parameter space, treating stain directions and concentration scaling as perturbable variables constrained by the Beer–Lambert model. Budgets are calibrated from multi-center statistics via the DKW inequality, and the method reports $\mathbb{R}^{2 \times N}$5 slide-level accuracy and $\mathbb{R}^{2 \times N}$6 worst-group accuracy on Camelyon17-WILDS, compared with $\mathbb{R}^{2 \times N}$7 and $\mathbb{R}^{2 \times N}$8 for a Macenko-normalized baseline. In that formulation, Macenko serves as a physically grounded coordinate system for augmentation rather than as a fixed template-matching transform (Hong, 12 May 2026).

Taken together, these results define Macenko stain normalization as both a classical normalization algorithm and a foundational stain model. It remains valuable when H&E slides exhibit moderate inter-site variation, when OD-space decomposition is desirable, and when computational simplicity matters. However, the empirical record shows that its benefits are conditional: it can strongly improve cross-dataset colorectal tissue classification, can underperform for segmentation, and can be sensitive to reference choice, background handling, and tissue-specific chromatic structure. Contemporary research increasingly treats those limitations not as reasons to discard the method, but as motivations to generalize it through multi-target normalization, augmentation, and domain-robust learning (Ivanov et al., 2024).