Abstract
Survival prediction in digital pathology is often hindered by the varying number of Whole Slide Images (WSIs) available per patient, a phenomenon termed "WSI distribution bias." We introduce WSI-Diffusion, a novel hierarchical diffusion model that generates representations of missing WSIs to fill these data gaps. By operating in embedding space rather than pixel space, the model efficiently synthesizes global slide blueprints and local patch features. Experimental results across five major cancer datasets show that WSI-Diffusion significantly improves prognostic accuracy and mitigates distribution bias, outperforming current state-of-the-art methods and large-scale foundation models.
Introduction
In digital pathology, survival prediction typically utilizes Whole Slide Images (WSIs)—gigapixel scans of tissue samples. A standard two-stage workflow involves segmenting WSIs into patches and training models to predict outcomes. However, clinical variability leads to a "WSI distribution bias," where some patients have many slides and others have very few. Models trained on these datasets often perform poorly on patients with fewer slides due to an incomplete histological view. Traditional data augmentation (e.g., flipping or rotating patches) fails to address this, as it only manipulates existing data rather than forecasting the missing biological insights an additional slide might provide.
Methods
We propose WSI-Diffusion, which generates entirely new WSI representations in representation space (embeddings) to bypass the computational costs of pixel-level gigapixel generation. The core of our approach is a two-stage hierarchical diffusion process:
- WSI-level Diffuser: This module generates a global "blueprint" for a missing slide, producing a patch consistency constraint (a high-level feature vector) and a tissue type distribution (proportions of neoplastic, inflammatory, or necrotic tissue).
- Patch-level Diffuser: Conditioned on the WSI-level blueprint, this module generates individual patch embeddings. A distribution-aware sampling strategy ensures that the generated patches align with the predicted tissue type distribution.
The generated representations are integrated with the patient’s actual slides and processed by a patient-level prediction module (using Convolutional Multiple Instance Learning) to calculate hazard risks via a Cox loss function.
Experimental Setup
The model was evaluated on five large-scale cancer datasets: NLST (Lung Adenocarcinoma and Squamous Cell Carcinoma), TCGA-LUSC, TCGA-LUAD, TCGA-BRCA (Breast Invasive Carcinoma), and TCGA-BLCA (Bladder Urothelial Carcinoma). These datasets include H&E-stained slides and corresponding clinical survival data.
Results and Discussion
WSI-Diffusion achieved superior performance across all datasets, with an average increase of 3.2% in C-index compared to existing baselines. Notably, it effectively mitigated distribution bias; in the NLST dataset, the performance gap between patients with few slides and those with many was reduced from 0.042 to 0.014.
Key insights from the study include:
- Hierarchy is Mandatory: Generating patches in isolation is insufficient. Without the WSI-level consistency constraint, generated patches lack the biological coherence necessary for accurate prediction.
- Outperforming Foundation Models: The model outperformed massive pathology foundation models like Prov-GigaPath and LongViT by approximately 3.5% in C-index, suggesting that targeted data generation can be more effective than simply scaling model parameters.
Conclusion
WSI-Diffusion addresses a critical bottleneck in digital pathology by generating prognostic insights for patients with limited tissue data. While the model currently relies on pre-defined tissue types for conditioning, future work will explore more granular histological features and improve the explainability of the generated representation space through techniques like clustering-based pseudo-labeling.
Summary updated