Patient-Specific Pseudo-Healthy Baselines

Updated 4 February 2026

The paper introduces an individualized synthetic template that simulates a patient's healthy anatomy from pathological scans to quantify anomalies.
It details deep generative and statistical approaches—including disentanglement GANs, diffusion inpainting, and biomechanical modeling—to achieve robust baseline reconstruction.
These frameworks support patient-specific longitudinal tracking, radiomics-based augmentation, and interpretable decision support through precise evaluation metrics.

A pathology-free patient-specific baseline, also termed a pseudo-healthy baseline, is an individualized synthetic anatomical reference that reconstructs what a patient’s scan would look like in the absence of disease or acquired pathology. This concept underlies a family of generative and statistical frameworks that seek to disentangle subject-specific anatomy from pathology-related deviations, thereby enabling robust anomaly quantification, personalized diagnosis, and interpretable decision support across medical imaging domains (Xia et al., 2020, Dalca et al., 2020, Marx et al., 2021, Wang et al., 29 Dec 2025, Chen et al., 17 Mar 2025, Zhang et al., 2022, Chen et al., 13 Jan 2026).

1. Conceptual Foundations and Motivation

The core objective of pathology-free patient-specific baselines is to generate, for a given individual, a reference image or structural state that represents their plausible healthy anatomy. This baseline can be computed retrospectively (from a pathological scan) or prospectively (as part of a disease progression model). Applications include:

Anomaly quantification: Measuring and localizing pathological deviations via residual analysis between observed and baseline anatomy (Xia et al., 2020, Dalca et al., 2020, Wang et al., 29 Dec 2025).
Longitudinal tracking: Disentangling disease-driven change from normal anatomical variation by computing subject-matched change trajectories (Dalca et al., 2020).
Augmentation and simulation: Generating counterfactuals for algorithm training, robustness assessment, and decision-support system benchmarking (Wang et al., 29 Dec 2025, Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026).

Traditionally, population-level templates are inadequate for capturing patient-specific anatomical idiosyncrasies. Pathology-free baselines resolve this by conditioning on the input subject, thus yielding image-level or biomechanical “healthy twins” for precise comparative analysis (Chen et al., 13 Jan 2026).

2. Deep Generative Approaches for Pseudo-Healthy Synthesis

Multiple deep learning paradigms address the pseudo-healthy synthesis task:

Disentanglement architectures: Models such as adversarial disentanglement-GANs explicitly factor image information into a healthy component and a pathology component (often a mask), allowing clean separation and targeted manipulation (Xia et al., 2020). The generator synthesizes a baseline by removing the pathological signal, while auxiliary networks enforce anatomical plausibility and cycle-consistency.
Segmentor-based discriminators: Replacing the binary discriminator with a pixel-wise segmentor sharpens lesion localization and enables direct adversarial harmonization of pathological and normal pixels (Zhang et al., 2022). The generator is trained to make the synthetic image indistinguishable from healthy anatomy from the segmentor’s perspective.
Diffusion models with inpainting: Diffusion-based masked inpainting (both 2D and 3D) is employed to fill in pathological regions by conditioning on the patient's own anatomy and the center-masked region, as in healthy persona synthesis for musculoskeletal MRI (Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026). This approach leverages powerful score-based models and supports downstream interpretability by precluding adversarial feature entanglement.
Decomposition with residual diffusion: The PathoSyn approach formalizes the image as an additive model $x = x_{\mathrm{sub}} + r$ , separating a deterministic anatomical baseline from a strictly lesion-localized stochastic deviation learned via a deviation-space diffusion process. The backbone U-Net reconstructs $x_{\mathrm{sub}}$ from masked images, constrained only by anatomic fidelity losses outside the lesion and inpainting priors inside (Wang et al., 29 Dec 2025).

Method	Pathology Separation	Input Requirements	Key Reference
Disentangle-GAN	Mask + image	Pathological + mask	(Xia et al., 2020)
Segmentor-GAN (GVS)	Masked cross-entropy	Pathological + mask	(Zhang et al., 2022)
Diffusion-inpainting	Masked region	Pathological + mask	(Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026)
PathoSyn diffusion	Additive deviation	Pathological + mask	(Wang et al., 29 Dec 2025)

3. Statistical and Biomechanical Patient-Specific Baselines

Beyond deep generative models, baseline computation spans:

Longitudinal biomechanical reference states: For patient-specific cardiac modeling, a pathology-free baseline comprises a subject’s stress-free reference geometry and passive parameterization, optimized such that the synthetic pressure-volume response fits empirical normal relations (e.g., the Klotz EDPVR), before any disease-induced remodeling (Marx et al., 2021). This approach delivers FE mesh initial states unbiased by acquired load or tissue change.
Population-informed, covariate-adaptive anatomical trajectories: Predictive modeling with mixed-effects regression augments a subject’s baseline anatomy by leveraging both fixed population dynamics and Gaussian-process nonparametric adjustments based on genetic, clinical, and image-derived features (Dalca et al., 2020). The individual trajectory yields a “healthy-predicted” scan for any future timepoint, enabling quantitative deviation mapping once observations are available.

4. Mathematical Formulations and Algorithmic Structure

The system comprises:

Generator $G(x_p)$ , Segmentor $S(x_p)$ , Reconstructor $R(\cdot)$ , and discriminators $D_x, D_m$ .
Adversarial, cycle, and mask-supervision losses, e.g.,

$\mathcal{L}_{GAN1} = \E_{x_h}\big[D_x(x_h)\big] - \E_{x_p}[D_x(G(x_p))] + \lambda_{gp} \text{(grad-penalty)}$

$\mathcal{L}_{CC1} = \E_{x_p} \|R(G(x_p), S(x_p)) - x_p \|_1$

The pseudo-healthy baseline is $\tilde x_h = G(x_p)$ .

Training: Forward noise process $q(x_t|x_{t-1})$ and learned denoising network $\epsilon_\theta(x_t, t, M)$ trained only on healthy data, with mask $M$ .
Inference: Starting from a masked pathological image, run T→0 reverse steps, imputing the masked region to generate the pathology-free persona $\hat x_0$ .

$x = x_{\mathrm{sub}} + r$

$x_{\mathrm{sub}}$ is estimated via masked U-Net, optimized with losses constrained inside/outside lesion.
No adversarial or stochastic regularization is applied to the baseline itself.

Iterative optimization of unloaded FE geometry $\mathbf X^*$ and material parameters $\theta$ to fit diastolic pressure-volume measurements.
Surrogate EDPVR models and Aitken-augmented fixed-point updates guarantee convergence.

5. Evaluation Metrics and Comparative Assessment

Evaluation strategies encompass synthetic image healthiness and fidelity, downstream task improvement, and clinical usability:

Healthiness ( $h$ ): Ratio of lesion-pixel volume as predicted by an external segmentor on the pseudo-healthy image to that on the pathological image (Xia et al., 2020).
Identity preservation ( $iD$ ): Masked multi-scale SSIM between original and reconstructed images outside lesions.
A-Dice: Area-under-curve Dice score measuring difficulty for a segmentor to refit lesion masks on generated images. Lower values reflect more effective lesion removal (Zhang et al., 2022).

Table: Typical performance comparison (ISLES FLAIR) (Xia et al., 2020)

Method	Identity ( $iD$ )	Healthiness ( $h$ )
CycleGAN	0.83	0.81
Ours (unpaired)	0.87	0.88
Ours (paired)	0.94	0.89

Clinical evaluations include expert ratings of anatomical plausibility, downstream segmentation/classification metrics (Dice, AUROC), and interpretability via human-explainable feature selection (Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026).

6. Downstream Applications and Interpretability

Patient-specific, pathology-free baselines support a range of analytical and translational tasks:

Deviational anomaly maps: Direct subtraction or ratio-based residuals localize abnormality, enabling patchwise or feature-level visualization (e.g., heatmaps or bar plots for radiomics) (Wang et al., 29 Dec 2025, Chen et al., 13 Jan 2026).
Radiomic fingerprinting: Coupling baseline features with classical radiomic extraction and patient-specific feature selection yields transparent, interpretable classification and biomarker discovery (Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026).
Counterfactual and data augmentation: By recombining baselines with stochastic deviation fields, models can simulate disease progression, regression, or generate synthetic cohorts for improved algorithm robustness (Wang et al., 29 Dec 2025).
Contrast enhancement: Synthetic healthy references enable contrast boosting through linear blending, improving lesion segmentation, especially in low-data regimes (Zhang et al., 2022).

7. Limitations, Extensions, and Open Challenges

Documented limitations include:

Dimensionality: Many frameworks are 2D or 2.5D due to memory constraints; extending to volumetric or time-resolved data is non-trivial (Xia et al., 2020, Zhang et al., 2022).
Ground truth absence: Paired healthy-for-pathological scans rarely exist for gold-standard evaluation. Metrics rely on segmentation/refitting proxies or indirect downstream performance improvements (Wang et al., 29 Dec 2025).
Mask dependence: Frameworks typically require high-quality lesion masks; weakly/unsupervised mask discovery remains underdeveloped (Zhang et al., 2022).
Multi-pathology scenarios: Addressing cases with overlapping or co-morbid pathologies necessitates more complex disentanglement (e.g., multiple mask channels) (Xia et al., 2020).
Interpretability vs. complexity: While radiomic baselines enable human-explainable insights, modeling high-dimensional variation and texture remains challenging (Chen et al., 17 Mar 2025, Chen et al., 13 Jan 2026).

A plausible implication is that future advances will combine learned uncertainty-weighted losses, scalable volumetric architectures, and hybrid statistical–generative frameworks to further improve both fidelity and clinical relevance of pathology-free baselines.

References (arXiv IDs: (Xia et al., 2020, Dalca et al., 2020, Marx et al., 2021, Wang et al., 29 Dec 2025, Chen et al., 17 Mar 2025, Zhang et al., 2022, Chen et al., 13 Jan 2026))