Pathology-Informed Domain Randomization

• Pathology-informed domain randomization is a strategy that injects clinical priors into data augmentation to randomize imaging parameters while preserving key pathological structures.
• It integrates morphology, anatomical cues, imaging physics, and acquisition metadata to produce synthetic data with clinically relevant variability.
• PIDR supports both pixel-level and feature-level augmentation, enhancing model robustness and diagnostic accuracy across diverse imaging modalities.

Pathology-informed domain randomization (PIDR) is a family of augmentation and synthetic-data strategies that widens the training distribution by randomizing clinically meaningful variation while explicitly constraining pathology-relevant structure. In contrast to blind perturbations, PIDR ties the randomization process to morphology, anatomy, imaging physics, acquisition metadata, or expert diagnostic criteria, so that variability is expanded without corrupting the label-defining substrate. Across recent work, this idea appears in computational pathology as morphology-preserving target-aware diffusion, in radiology as lesion simulators grounded in physics and anatomical placement, in fetal MRI as phenotype-specific label-space deformation, and in 3D MRI as pathology-preserving outpainting (Zhang et al., 23 Jan 2026, Suzuki et al., 2022, Plana et al., 28 Aug 2025, Tan et al., 14 Jan 2026).

1. Conceptual scope and boundaries

In the recent pathology literature, PIDR is best understood as a specialization of domain randomization rather than a synonym for generic augmentation. Classical domain randomization exposes a model to wide variation in nuisance factors such as color, stain concentration, scanner characteristics, geometry, blur, or resolution, usually without explicit target-domain knowledge. Semi-supervised domain adaptation, by contrast, leverages unlabeled target data to reduce source–target mismatch. In computational pathology, this distinction is explicit: stain perturbations such as Macenko or Vahadane augmentation remain target-agnostic, whereas a target-aware latent diffusion model conditioned on cohort and tissue preparation falls on the PIDR side because it uses pathology priors and cohort metadata to control the randomization process (Zhang et al., 23 Jan 2026). The broader synthesis-driven neuroimaging literature similarly defines domain randomization as training on synthetic images generated from anatomical label maps with deliberately exaggerated variability in spatial structure, intensity, resolution, and artifacts; PIDR extends that scaffold by inserting pathology priors into the generative model (Hoffmann, 17 Jul 2025).

The motivation for PIDR is not only improved average accuracy but robustness under medically relevant shift. This is especially clear in pathology foundation models. A multiscanner study of 14 feature extractors on 384 identical H&E slides scanned on five devices found that most models encoded pronounced scanner-specific signatures; unsupervised embedding analyses showed scanner-separated structure, and calibration analyses showed that predicted probabilities for the same slide could shift systematically across scanners even when AUC remained comparatively stable. Fleiss’ $\kappa$ varied substantially across models and tasks, and LOWESS fits deviated from the diagonal, revealing scanner-dependent calibration bias (Thiringer et al., 7 Jan 2026). This suggests that acquisition variability is itself a pathology-relevant randomization axis, not merely a cosmetic nuisance.

2. Pathology-informed randomization axes

The defining feature of PIDR is that the axes of variation are selected from domain knowledge rather than sampled blindly. In computational pathology, one such axis is tissue morphology. The latent diffusion framework for lung adenocarcinoma prognostication conditions generation on UNI foundation-model features extracted from each source tile, alongside cohort identity and tissue preparation status. Morphology is thereby encoded as a structural prior, while cohort and preparation embeddings steer appearance toward target-domain staining, scanner, and preparation statistics (Zhang et al., 23 Jan 2026). A closely related but distinct mechanism appears in human-feedback-aligned cell synthesis, where a hematopathologist evaluates images against seven clinical criteria—cell size, nucleus shape and size, nucleus-to-cytoplasm ratio, cytoplasm color and consistency, chromatin pattern, inclusions, and granules—and those criteria are distilled into a reward model that operationalizes clinical plausibility (Sun et al., 2023).

In radiology and neuroimaging, the pathology-informed axes are often anatomical or physical rather than histomorphological. Goldilocks-curriculum domain randomization for pneumonia detection encodes lesion texture and morphology through fractal Perlin noise, opacity through Beer–Lambert-compliant intensity blending, and anatomical placement through a lung-field heuristic that keeps lesions in darker parenchymal regions while allowing rib and cardiac shadow occlusion (Suzuki et al., 2022). Fetal MRI work on corpus callosum dysgenesis defines the randomization axes at label level: complete agenesis, partial agenesis, hypoplasia, dysplasia, ventriculomegaly, cortical changes, and posterior fossa alterations are simulated by morphology operations and smooth deformations constrained by connectivity and plausibility checks (Plana et al., 28 Aug 2025). In 3D MRI outpainting, the pathology itself is held fixed and the randomized axis is the surrounding anatomical context; the model conditions on the real masked pathology and synthesizes only the non-pathological surroundings (Tan et al., 14 Jan 2026).

The synthesis-driven tutorial literature makes these axes explicit as parameterized controls over affine deformation, elastic warping, intensity assignment, bias field, blur, noise, gamma, downsampling, and field-of-view. The reported starter ranges include translation $[-30,30]$ mm, rotation $[-30^\circ,30^\circ]$, scaling $[90\%,110\%]$, elastic warp strength $[0,20]$ mm, bias field $[0\%,50\%]$, blur $[0,2]$ mm, gamma $[0.5,1.5]$, and downsampling factor $[1,4]$; pathology-informed extensions insert lesion classes, lesion probability maps, and modality-specific lesion appearance rules into the same framework (Hoffmann, 17 Jul 2025).

3. Generative and optimization mechanisms

PIDR does not correspond to a single generator class. Diffusion, simulators, label-space deformation, residual translation, and embedding-space perturbation have all been used, provided that the randomization is constrained by pathology priors.

A morphology-aware latent diffusion formulation appears in pathology SSDA. In that setting, a VQ-VAE maps tiles to latent space, a latent-space UNet predicts denoising noise, and the denoising objective is conditioned on morphology and metadata:

$$\mathcal{L}_{\mathrm{LDM}}=\mathbb{E}_{t,z_0,\epsilon}\left[\left\|\epsilon-\epsilon_\theta(z_t,t,c)\right\|_2^2\right],$$

where $[-30,30]$0 encodes UNI features, cohort identity, and tissue preparation. Classifier-free guidance is then used at inference, with condition dropout probability $[-30,30]$1 during training and guidance scale $[-30,30]$2 at sampling (Zhang et al., 23 Jan 2026).

Goldilocks-curriculum domain randomization uses a different principle. Instead of directly optimizing fidelity, it adapts simulator parameters so that synthetic data remain at a target difficulty for the current detector. The next simulator setting is chosen by

$[-30,30]$3

where $[-30,30]$4 is the current model’s performance on synthetic data and $[-30,30]$5 is the desired difficulty. Because $[-30,30]$6 is treated as a black box, Bayesian optimization selects simulator parameters over the pathology-informed space of lesion texture, size, edge sharpness, opacity, and placement (Suzuki et al., 2022).

Residual translation methods provide yet another mechanism. In PathoGAN, generators output an explicit pathology labelmap $[-30,30]$7 and inpainting $[-30,30]$8, and the translated image is composed by masked blending:

$[-30,30]$9

This exposes a pixelwise pathology probability while enabling sampling of unseen pathology from a latent distribution in the healthy-to-pathological direction (Andermatt et al., 2018).

Finally, PIDR can operate outside image space. PathoSCOPE synthesizes pathological embeddings rather than synthetic images. Its PiEG module initializes a perturbed embedding with Gaussian noise and refines it for $[-30^\circ,30^\circ]$0 steps by normalized gradient ascent on a global loss; the reported settings are $[-30^\circ,30^\circ]$1, $[-30^\circ,30^\circ]$2, and modality-dependent step size $[-30^\circ,30^\circ]$3 with best performance around $[-30^\circ,30^\circ]$4 for BraTS2020 and $[-30^\circ,30^\circ]$5 for ChestXray8 (Chin et al., 23 May 2025). This shows that pathology-informed randomization need not be pixel-level if the geometry of the feature space is itself clinically structured.

4. Histopathology-specific implementations

In computational pathology, the central technical tension is between appearance transfer and morphology preservation. The lung adenocarcinoma SSDA framework addresses this by training an LDM on unlabeled NLST and TCGA tiles while withholding target labels. For each labeled NLST tile, the model extracts UNI features, conditions on UNI plus target cohort and tissue-preparation embeddings, and generates a synthetic target-aware counterpart with the same subtype label. The downstream ViT-B/16 classifier is then trained on the union of real labeled NLST tiles and synthetic labeled tiles. On the held-out TCGA test cohort, weighted F1 improved from 0.611 to 0.706 and macro F1 from 0.641 to 0.716, with the best non-LDM baseline being Vahadane at weighted F1 $[-30^\circ,30^\circ]$6. The synthetic images achieved FID $[-30^\circ,30^\circ]$7, and an ablation showed that adding unlabeled TCGA data to LDM training improved target weighted F1 from 0.614 to 0.680 and macro F1 from 0.640 to 0.701, while source-cohort differences were not significant (Zhang et al., 23 Jan 2026).

A distinct pathology implementation focuses on expert-aligned plausibility rather than cohort adaptation. In bone marrow aspirate single-cell patches, a diffusion model was first pretrained and then audited by one hematopathologist over 3,936 synthetic images. The reward model trained on that feedback was used to fine-tune the generator away from clinically implausible regions. The average fraction of clinically plausible synthetic images improved from 0.21 to 0.75. On held-out test data, precision increased from 68.06 without feedback to 81.01 with feedback, recall from 52.00 to 56.74, and coverage from 56.98 to 84.57; a classifier trained on synthetic data improved from F1 60.33 without feedback to 75.80 with feedback and real-augmented reward training, compared with 79.03 when trained on real data (Sun et al., 2023). The paper’s explicit argument is that FID, precision, recall, and coverage alone do not encode clinical knowledge.

These works jointly define a pathology-specific interpretation of “preservation.” In the LUAD setting, preservation means retaining glandular and architectural patterns that determine labels such as lepidic versus acinar, papillary, solid, or micropapillary growth. In the single-cell setting, preservation means respecting nucleus morphology, chromatin, cytoplasm, inclusions, and granules. The common principle is that clinically relevant structure is treated as a hard or soft constraint, while appearance variation is permitted elsewhere.

5. Radiology and neuroimaging implementations

In radiography, PIDR has been instantiated through explicit simulators. Goldilocks-curriculum domain randomization was evaluated on a benchmark of 101 chest X-rays with difficult-to-identify pneumonia lesions and a simulator that inserts pseudo-lesions by combining fractal Perlin noise with Beer–Lambert intensity modeling. Across five folds, the best overall configuration, GDR with NVRM-SGD, achieved FAUC $[-30^\circ,30^\circ]$8, CPM $[-30^\circ,30^\circ]$9, and [email protected] $[90\%,110\%]$0, outperforming UDR with Adam and BayRn with Adam. Varying the target difficulty revealed a clear “just-right” zone: target FAUC $[90\%,110\%]$1 gave FAUC $[90\%,110\%]$2, CPM $[90\%,110\%]$3, and [email protected] $[90\%,110\%]$4, whereas lower or higher targets degraded performance (Suzuki et al., 2022).

In fetal MRI, PIDR is implemented at label level rather than texture level. Healthy super-resolution T2 fetal MRI labels are morphed to simulate complete agenesis, partial agenesis, hypoplasia, dysplasia, ventriculomegaly, cortical changes, and posterior fossa alterations, and FetalSynthSeg then synthesizes paired images from the modified labels. Training used only healthy subjects from KISPI and STA, with no real pathological annotations. The resulting SimPath models improved clinically relevant biomarker estimation, reducing corpus callosum length estimation error from 1.89 mm to 0.80 mm in healthy cases and from 10.9 mm to 0.7 mm in CCD cases, while maintaining performance on healthy fetuses and those with other pathologies. The paper further reports that SimPath50, where 50% of synthetic samples carry CCD-informed augmentations, provides a better trade-off than SimPath100, indicating that balanced pathology injection matters (Plana et al., 28 Aug 2025).

POWDR extends PIDR to full-volume 3D MRI outpainting. The method keeps a real lesion fixed, applies a 3D Haar DWT, conditions a wavelet-domain diffusion model on the masked pathology, and synthesizes anatomically plausible surroundings. A random connected mask training strategy addresses conditioning-induced collapse. Repeated sampling with the same pathology condition reduced cosine similarity from 0.9947 to 0.9580 and increased KL divergence from 0.00026 to 0.01494, indicating greater diversity outside the lesion. Adding 50 synthetic cases improved nnU-Net tumor segmentation Dice from 0.6992 to 0.7137. Tissue-volume analysis found no significant differences for CSF and GM between synthetic and real images (Tan et al., 14 Jan 2026).

Older weakly supervised translation work provides a precursor to PIDR. PathoGAN, trained on BRATS 2017 slices labeled only as healthy or pathological, learned to segment pathology and produce healthy inpaintings through residual generators. On the test set it achieved Dice $[90\%,110\%]$5 against $[90\%,110\%]$6 for a fully supervised MDGRU baseline, while also enabling sampling of unseen pathology via latent variation (Andermatt et al., 2018). More recent few-shot work shifts the same logic into feature space: PathoSCOPE combines a Global-Local Contrastive Loss with pathology-informed embedding generation and reports BraTS2020 image AUROC 89.19 and pixel AUROC 97.87 at $[90\%,110\%]$7 shots, alongside computational efficiency of 2.48 GFLOPs and 166 FPS (Chin et al., 23 May 2025).

6. Evaluation regimes and empirical behavior

Representative PIDR systems are evaluated with markedly different metrics because they target different clinical failure modes.

Setting	Evaluation focus	Reported outcome
Histopathology SSDA (Zhang et al., 23 Jan 2026)	Target-cohort classification	Weighted F1 $[90\%,110\%]$8; Macro F1 $[90\%,110\%]$9
Pneumonia GDR (Suzuki et al., 2022)	Sim2real lesion detection	FAUC $[0,20]$0; CPM $[0,20]$1; [email protected] $[0,20]$2
Bone marrow feedback alignment (Sun et al., 2023)	Expert plausibility	Clinically plausible fraction $[0,20]$3
CCD fetal MRI (Plana et al., 28 Aug 2025)	Biomarker accuracy	LCC error $[0,20]$4 mm in healthy; $[0,20]$5 mm in CCD
POWDR (Tan et al., 14 Jan 2026)	Segmentation utility	Dice $[0,20]$6 with +50 synthetic cases

A recurrent finding is that no single metric is sufficient. Fidelity measures can be encouraging yet incomplete: the LUAD diffusion model reports FID $[0,20]$7, and the paper interprets values below 10 as high fidelity, but the cell-synthesis work explicitly argues that domain-agnostic metrics cannot capture clinical sensibility (Zhang et al., 23 Jan 2026, Sun et al., 2023). Conversely, pathology foundation model evaluation shows that discriminative AUC can remain comparatively stable while embedding geometry and calibration shift across scanners, so embedding metrics such as Average Pairwise Cosine Distance, 1-Nearest-Neighbour match rate, Mantel correlation, Mean Intra-Scanner Distance, and Intersection-over-K consistency, as well as calibration analyses based on LOWESS, are necessary to expose scanner dependence (Thiringer et al., 7 Jan 2026).

Topology and anatomy-sensitive metrics occupy a special place in PIDR because many pathology priors are structural. Fetal MRI evaluation includes generalized Dice, Hausdorff distance at the 95th percentile, and Euler Difference, with the latter quantifying departures from the expected topology of the corpus callosum. In that study, synthetic-trained models achieved mean ED $[0,20]$8 against 12.04 for original dHCP CC annotations when compared to original dHCP CC topology, underscoring that topology-aware evaluation can reveal clinically relevant improvements not reducible to overlap alone (Plana et al., 28 Aug 2025). This suggests that PIDR should be evaluated as a multi-objective intervention spanning task performance, structural fidelity, diversity, and calibration.

7. Limitations, failure modes, and future directions

The principal limitations of PIDR are computational, epistemic, and supervisory. Latent diffusion in pathology required 2× NVIDIA L40S GPUs and approximately 76 hours of training; synthetic generation also incurs multi-step denoising cost (Zhang et al., 23 Jan 2026). Goldilocks curriculum learning adds Bayesian optimization with 35 iterations per timestep (Suzuki et al., 2022). Full-volume 3D wavelet diffusion is memory-intensive and was trained on an NVIDIA H100 with 200,000 iterations (Tan et al., 14 Jan 2026). Beyond compute, several methods depend on metadata quality or unlabeled target-domain representativeness; if cohort labels, tissue-preparation information, or target-domain data are noisy, the learned prior can imprint bias rather than robustness (Zhang et al., 23 Jan 2026).

A second limitation is that stronger pathology simulation is not necessarily better. In fetal MRI, SimPath100 could degrade healthy performance and some CCD metrics relative to SimPath50, indicating that over-weighting simulated pathology can shift the training distribution away from the clinically dominant regime (Plana et al., 28 Aug 2025). In human-feedback alignment, expert time is the bottleneck, only binary plausibility flags were collected, and the reward model reflects one expert’s preferences; the paper also notes reward-hacking risk and the absence of inter-rater reliability (Sun et al., 2023). In lesion simulation, fractal Perlin noise is well suited to infiltrative opacities such as pneumonia but is less suitable for sharply demarcated masses (Suzuki et al., 2022). In pathology robustness analysis, stable AUC can still conceal scanner-dependent calibration bias, so accuracy-centric validation remains a failure mode in its own right (Thiringer et al., 7 Jan 2026).

The future directions proposed across the literature are convergent. Pathology SSDA points toward refined conditioning for fine-grained morphology preservation, multi-cohort conditioning, self-supervised feature consistency, uncertainty estimation for synthetic labels, and faster diffusion samplers (Zhang et al., 23 Jan 2026). Fetal MRI work proposes topology-preserving or diffeomorphic constraints, multimodal priors, semi-supervised fine-tuning on small real pathological cohorts, and automated phenotype labeling from segmentation-derived biomarkers (Plana et al., 28 Aug 2025). The synthesis tutorial emphasizes learned generative priors, learned residual degradations, and tighter integration between pathology-aware simulators and general domain-randomized training pipelines (Hoffmann, 17 Jul 2025). Taken together, these directions indicate that PIDR is moving from heuristic augmentation toward a broader program of clinically constrained distribution design, in which pathology knowledge, acquisition knowledge, and uncertainty-aware evaluation are optimized jointly rather than treated as separate stages.