PathoSyn: Imaging-Pathology MRI Synthesis via Disentangled Deviation Diffusion

TL;DR

PathoSyn addresses the challenge of MRI pathology synthesis by modeling disease as a structured deviation $\mathbf{r}$ on a fixed anatomical substrate $\mathbf{x}_{\mathrm{sub}}$, enabling localized, controllable lesion generation. It introduces a deviation-space diffusion model conditioned on anatomy and lesion mask, coupled with seam-aware fusion and a boundary-stability regularizer to preserve anatomical fidelity. The approach yields higher perceptual realism, stronger mathematical disentanglement between anatomy and pathology, and improved downstream performance for segmentation and classification on brain tumor benchmarks, while enabling robust counterfactual disease progression modeling. By focusing stochasticity within the pathology domain and enforcing spatial coherence, PathoSyn provides a principled, data-efficient framework for generating clinically credible digital twins and for stress-testing AI diagnostic tools in low-data regimes.

Abstract

We present PathoSyn, a unified generative framework for Magnetic Resonance Imaging (MRI) image synthesis that reformulates imaging-pathology as a disentangled additive deviation on a stable anatomical manifold. Current generative models typically operate in the global pixel domain or rely on binary masks, these paradigms often suffer from feature entanglement, leading to corrupted anatomical substrates or structural discontinuities. PathoSyn addresses these limitations by decomposing the synthesis task into deterministic anatomical reconstruction and stochastic deviation modeling. Central to our framework is a Deviation-Space Diffusion Model designed to learn the conditional distribution of pathological residuals, thereby capturing localized intensity variations while preserving global structural integrity by construction. To ensure spatial coherence, the diffusion process is coupled with a seam-aware fusion strategy and an inference-time stabilization module, which collectively suppress boundary artifacts and produce high-fidelity internal lesion heterogeneity. PathoSyn provides a mathematically principled pipeline for generating high-fidelity patient-specific synthetic datasets, facilitating the development of robust diagnostic algorithms in low-data regimes. By allowing interpretable counterfactual disease progression modeling, the framework supports precision intervention planning and provides a controlled environment for benchmarking clinical decision-support systems. Quantitative and qualitative evaluations on tumor imaging benchmarks demonstrate that PathoSyn significantly outperforms holistic diffusion and mask-conditioned baselines in both perceptual realism and anatomical fidelity. The source code of this work will be made publicly available.

Figures (5)

• Figure 1: Conventional models synthesize images either by processing the entire image holistically or by fully segregating pathological from non-lesion regions. In contrast, PathoSyn explicitly disentangles anatomical structure from pathological alterations and performs generative modeling within a deviation space, thereby enabling controlled modulation of lesion characteristics while preserving the underlying anatomical substrate.
• Figure 2: The PathoSyn architecture consists of two encoders and two decoders that disentangle an image into an anatomical substrate and a pathology deviation. The anatomical encoder–decoder pair learns a stable anatomical substrate that preserves underlying tissue geometry, while the pathology encoder–decoder pair models localized pathology deviation driven by lesion appearance. A generative diffusion module refines the pathology deviation to capture diverse and spatially coherent variations. A fusion layer recombines the anatomical substrate with the pathology deviation, enabling reconstruction and controllable synthesis of pathological images by adjusting lesion severity, extent, and spatial distribution while maintaining anatomical consistency. Please refer to the training and inference in Alg. \ref{['alg:train']} and Alg. \ref{['alg:infer']}.
• Figure 3: Top: ROC curves with 95% bootstrap confidence intervals ($\downarrow$), where reduced detectability indicates greater alignment with real data. Bottom: Distribution of bootstrap AUC discriminability scores between real and generated images ($\downarrow$ is better).
• Figure 4: Qualitative comparison of pathological image synthesis. From left to right: the given image, Brain-LDM reconstruction, MaskDiff-Inpaint completion, PathoSyn (VAE-GAN), and PathoSyn (diffusion). The diffusion-based PathoSyn results further display the anatomical substrate and the corresponding pathology deviation fields, illustrating clearer separation between structural content and lesion-driven appearance, with improved spatial coherence and controllable pathology representation.
• Figure 5: Top left: segmentation performance on generated images from multiple models, including no augmentation and conventional augmentation baselines, reported by Dice coefficients to evaluate downstream utility. Top right: deep perceptual texture statistics quantifying alignment between generated and real images in feature space. Bottom left: distributions of intensity- and texture-based features comparing generated versus real samples, reflecting low-level photometric and mesoscopic consistency. Bottom right: empirical cumulative distribution functions of feature distances, with the x-axis denoting deviation from real data; curves concentrated toward the origin indicate improved fidelity and reduced distributional shift.