Explainable Pathomics Feature Visualization via Correlation-aware Conditional Feature Editing

TL;DR

This work addresses the interpretability gap in pathomics by introducing Manifold-Aware Diffusion (MAD), a framework that jointly models correlated pathomics features and nucleus images. MAD leverages a disentangled feature manifold learned with a $eta$-VAE and a conditional diffusion model trained on nucleus images, enabling latent-guided edits that move a target feature toward a desired value while adjusting correlated features to stay on the manifold. At test time, latent optimization in the VAE space yields a feasible feature vector, which then conditions diffusion-based image editing to produce high-fidelity edits that preserve structural coherence. Across real and geometric datasets, MAD outperforms baselines in feature control and image quality, demonstrating its potential for explainable, interactive visualizations of quantitative pathomics signatures in digital pathology.

Abstract

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.

Figures (9)

• Figure 1: Feature editing in pathomics. (a) In generic object editing (e.g., an apple), attributes such as stem length and size operate independently. Modifying one feature does not impose constraints on others. (b) Pathomics features often exhibit intrinsic correlations (e.g., Area and Perimeter). The "Current Method" modifies the target feature (Area) while fixing the correlated feature (Perimeter), creating a geometric conflict and resulting in an infeasible edit. The proposed "Correlated-Aware" framework updates the correlated attribute alongside the target, ensuring the output remains within a feasible biological structure.
• Figure 2: Independent editing versus manifold-aware editing of correlated pathomics features. The turquoise surface depicts the manifold of real nuclei in the feature space of area, perimeter, and orientation. The green point marks the input nucleus. To decrease area, independent editing (red dashed path) changes the area coordinate while keeping the other features fixed, reaching the oracle editing point (red cross), which lies off the manifold and leads to an out-of-distribution image. Manifold-aware editing (blue path) follows the manifold and jointly updates correlated features, reaching the MAD editing result (purple point) on the manifold. The images on the right show the input image, the image generated under the oracle editing condition, and the on-manifold result produced by our proposed model (MAD).
• Figure 3: Training and inference pipeline of MAD. (a) A conditional diffusion model learns to use a feature vector to reconstruct the nucleus image. A pre-trained MLP encodes the feature vector into a conditioning embedding for the diffusion U-Net. (b) A disentangled VAE is trained on pathomics features. The encoder maps a feature vector to a latent variable $z$ and the decoder reconstructs the feature vector. (c) Given an input nucleus and a target feature value, the encoder initializes a latent variable from the feature of the input image. Through latent optimization, the decoded feature moves toward the target feature along the learned feature manifold. The bottom plots illustrate this trajectory in the feature space. The optimized feature vector is then used as the condition for the diffusion U-Net to edit the input image.
• Figure 4: Qualitative results between unconditional generation (StyleGAN2) and conditional generation (Ours). Each row represents the traversal of a specific feature value from low to high. The left block shows StyleGAN2, which takes the target feature vector as input and generates a nucleus for each target value. The right block shows MAD, which edits a single input nucleus to match the same sequence of target values. StyleGAN2 samples follow the target feature but can introduce artifacts. MAD keeps the nuclei appearance close to the input image while changing the designated feature.
• Figure 5: Qualitative results for conditional editing. Each row specifies a target trajectory for one nucleus feature. All three models take the same input nucleus (blue box) and the same sequence of target feature values. For Stable Diffusion and MAD without VAE, the edited nuclei stay close to the input nucleus across target values, which indicates that the conflicting conditions are not effectively reflected in the images. In contrast, MAD produces edited nuclei that follow the target feature change while preserving the overall appearance of the input nucleus.