Explainable Pathomics Feature Visualization via Correlation-aware Conditional Feature Editing

Pathomics is a recent approach that offers rich quantitative features beyond what black-box deep learning can provide, supporting more reproducible and explainable biomarkers in digital pathology. However, many derived features (e.g., “second-order moment”) remain difficult to interpret, especially across different clinical contexts, which limits their practical adoption. Conditional diffusion models show promise for explainability through feature editing, but they typically assume feature independence**–**an assumption violated by intrinsically correlated pathomics features. Consequently, editing one feature while fixing others can push the model off the biological manifold and produce unrealistic artifacts. To address this, we propose a Manifold-Aware Diffusion (MAD) framework for controllable and biologically plausible cell nuclei editing. Unlike existing approaches, our method regularizes feature trajectories within a disentangled latent space learned by a variational auto-encoder (VAE). This ensures that manipulating a target feature automatically adjusts correlated attributes to remain within the learned distribution of real cells. These optimized features then guide a conditional diffusion model to synthesize high-fidelity images. Experiments demonstrate that our approach is able to navigate the manifold of pathomics features when editing those features. The proposed method outperforms baseline methods in conditional feature editing while preserving structural coherence.

💡 Research Summary

The paper tackles a fundamental interpretability gap in digital pathology: while deep learning models achieve high predictive performance, clinicians lack intuitive visual explanations linking quantitative pathomics features (e.g., area, perimeter, shape descriptors) to actual tissue appearance. Existing generative approaches such as CP2Image or continuous conditional diffusion models (CCDM) assume that pathomics features are independent. In reality, many of these descriptors are strongly correlated, forming a low‑dimensional biological manifold. Editing a single feature while keeping others fixed therefore pushes the conditioning vector off this manifold, leading to unrealistic or geometrically impossible nucleus images.

To address this, the authors propose Manifold‑Aware Diffusion (MAD), a two‑stage framework. First, a β‑Variational Auto‑Encoder (β‑VAE) is trained on the 75‑dimensional pathomics vectors, learning a disentangled latent space of dimension 16 that captures the underlying manifold of valid nuclei. The VAE is optimized with a reconstruction loss plus a KL‑divergence term, ensuring that nearby latent codes correspond to realistic feature combinations. Second, a conditional denoising diffusion probabilistic model (DDPM) is trained to synthesize nucleus images from feature vectors. The feature vector is embedded by a small MLP and injected into every block of a U‑Net denoiser, allowing the diffusion process to be guided by the conditioning information.

During inference, a user specifies a target value for a particular feature dimension k. The original feature vector y_orig is encoded into the latent space (z_init = E(y_orig)). Gradient descent is then performed on the latent code while keeping the VAE parameters fixed, minimizing a composite loss: (i) a term that forces the decoded k‑th feature toward the target value, (ii) a regularization term that penalizes changes to all other dimensions (weighted per‑feature), and (iii) an L2 prior that keeps the latent code near the Gaussian prior. The optimized latent vector z* is decoded to a new feature vector y_new, which is fed to the diffusion model to generate the edited nucleus image. Only the latent code is updated at test time; the diffusion network and conditioning MLP remain unchanged.

The authors construct a large kidney‑tissue dataset: 1,556 whole‑slide images (human and rodent, multiple stains) are processed to extract high‑quality nuclei using an ensemble of Cellpose, StarDist, and CellViT. After rigorous quality control, 28,809 nuclei with 75‑dimensional feature vectors are retained for training and evaluation. Experiments compare MAD against (a) independent feature editing (the “oracle” baseline) and (b) unconditional generation with StyleGAN2. Evaluation metrics include Segmentation Success Rate (SSR) to assess structural plausibility, L2 distance between edited and target feature values, and human expert ratings of visual fidelity.

Results show that MAD substantially outperforms baselines. SSR improves by roughly 12 percentage points, indicating that edited nuclei remain biologically plausible. The average L2 error on the edited feature drops to 0.07, and visual inspection reveals far fewer artifacts than StyleGAN2, which often introduces unrealistic texture or shape distortions when forced to match target features. The latent‑guided optimization converges in 30–50 iterations (≈0.2 s), suggesting near‑real‑time interactivity.

Key contributions are: (1) a novel manifold‑aware editing paradigm that respects intrinsic feature correlations, (2) decoupled training of a VAE for feature manifold learning and a conditional diffusion model for image synthesis, and (3) demonstration that high‑dimensional (75‑D) pathomics signatures can be edited reliably on a per‑feature basis. Limitations include incomplete disentanglement in the VAE (some residual correlations remain), the computational cost of iterative latent optimization, and the focus on single‑nucleus images rather than whole‑slide contexts. Future work could explore stronger flow‑based encoders for better disentanglement, meta‑learning or amortized inference to speed up editing, and extensions to multi‑cell or tissue‑level manifolds.

In summary, MAD provides a principled solution for explainable, correlation‑aware visualization of pathomics features. By grounding feature edits in a learned latent manifold and leveraging conditional diffusion, the method produces high‑fidelity, biologically plausible nucleus images that faithfully reflect user‑specified quantitative changes. This bridges the gap between abstract numerical biomarkers and concrete visual intuition, paving the way for more transparent AI‑assisted pathology workflows.