Dynamic Entity-Masked Graph Diffusion Model for histopathological image Representation Learning

TL;DR

The paper tackles the gap between natural-image pretraining and histopathology by introducing H-MGDM, a self-supervised framework that builds dynamic entity graphs from histology patches and learns representations through a latent graph diffusion process conditioned on complementary subgraphs. It combines a tissue-aware graph construction, latent space compression, and a diffusion-based denoising objective within a two-stage training paradigm to produce robust, interpretable patch representations. Across three large histopathology datasets and six downstream tasks, including cancer classification and survival analysis, H-MGDM consistently outperforms strong baselines and provides meaningful interpretability through graph attention mechanisms and diffusion visualizations. The proposed approach has practical implications for annotation-efficient analysis and prognosis in computational pathology, and sets the stage for broader applications in generation and integration of histopathological priors.

Abstract

Significant disparities between the features of natural images and those inherent to histopathological images make it challenging to directly apply and transfer pre-trained models from natural images to histopathology tasks. Moreover, the frequent lack of annotations in histopathology patch images has driven researchers to explore self-supervised learning methods like mask reconstruction for learning representations from large amounts of unlabeled data. Crucially, previous mask-based efforts in self-supervised learning have often overlooked the spatial interactions among entities, which are essential for constructing accurate representations of pathological entities. To address these challenges, constructing graphs of entities is a promising approach. In addition, the diffusion reconstruction strategy has recently shown superior performance through its random intensity noise addition technique to enhance the robust learned representation. Therefore, we introduce H-MGDM, a novel self-supervised Histopathology image representation learning method through the Dynamic Entity-Masked Graph Diffusion Model. Specifically, we propose to use complementary subgraphs as latent diffusion conditions and self-supervised targets respectively during pre-training. We note that the graph can embed entities' topological relationships and enhance representation. Dynamic conditions and targets can improve pathological fine reconstruction. Our model has conducted pretraining experiments on three large histopathological datasets. The advanced predictive performance and interpretability of H-MGDM are clearly evaluated on comprehensive downstream tasks such as classification and survival analysis on six datasets. Our code will be publicly available at https://github.com/centurion-crawler/H-MGDM.

Figures (8)

• Figure 1: Pathological slide inspection process from the overall view to details. Unlike comparison methods, H-MGDM focuses on masked pathological tissue regions rather than grid tiles in patches, constructing the masked subgraph with varying intensities and noise for reconstruction by complementary conditional subgraph.
• Figure 2: Overview of the H-MGDM pretraining stages. Conditional diffusion reverse process in the decoder. $\mathbf{G}_e$ and $\mathbf{G}_d$ are two complementary subgraphs of $\mathbf{G}$. $\mathbf{G}_d(t)$ are from the diffusion forward process $q_L$ of $\mathbf{G}_d$. The target is to denoise $\mathbf{G}_d(t)$ to $\hat{\mathbf{G}}_d^{(0)}(t)$ close to $\mathbf{G}_d$ at sampling time $t$.
• Figure 3: Original images and their attention heatmaps of five different categories of the PANDA dataset, showing the interpretability of our method under the pathological entity graph construction comparison with Dino.
• Figure 4: Kaplan-Meier Analysis of comparison methods and our framework. All patients from the five tests were pooled and analyzed. Each cohort is split into a high-risk (orange) and a low-risk group (blue) according to the median output of the cohort.
• Figure 5: T-SNE plots of pan-cancer samples' readout representations learning with H-MGDM and baseline methods.