H-MGDM

Self-supervised pretraining on unlabelled histopathology patches has become the default route to a general-purpose tissue encoder, and masked reconstruction is its dominant recipe. Masked autoencoders hide a fraction of a regular grid of square tiles and train the network to paint them back. That grid is convenient but arbitrary: it cuts across glands, stroma and tumour nests without regard for where one tissue structure ends and the next begins, so the reconstruction task never asks the model about the spatial relationships between the entities a pathologist actually reads.

H-MGDM, the Dynamic Entity-Masked Graph Diffusion Model, replaces the grid with a graph. Developed at Xiamen University with collaborators at the University of Science and Technology of China, the University of Hong Kong, IMT Mines Alès and the University of Montpellier with CHU de Nîmes, and published at AAAI 2025, it segments each patch into tissue superpixels, treats those as graph vertices with edges between touching regions, and pretrains a latent diffusion model that reconstructs one half of the graph while conditioning on the other.

#Key Features

• Entity graphs instead of tile grids: SLIC superpixels define pathological entities, so masking removes whole tissue structures and the model must reason about their topological context to restore them.
• Complementary dual-subgraph objective: Each graph is randomly split into an encoder subgraph, which supplies the diffusion condition, and a decoder subgraph, which is the denoising target — the two roles are resampled every step.
• Latent rather than pixel diffusion: A first-stage entity autoencoder compresses each superpixel window into a latent vector, and the diffusion model operates entirely in that space, on vertex and edge features alike.
• Dynamic noise intensity: Sampling the diffusion timestep during pretraining varies the corruption level across examples rather than fixing a single mask strength, which the ablation shows matters as much as the graph structure itself.
• Inspectable readouts: Because the global representation is a readout over named tissue vertices, high-attention regions map back onto identifiable structures without a post-hoc saliency method.

#Technical Details

Pretraining uses three histopathology corpora: the Komura et al. pancancer set (1.6M images, 32 cancer types), PANDA (12.5M images from 11,000 Gleason-annotated prostate slides), and an in-house inflammatory bowel disease collection of 23M images with 360K patches labelled across nine tissue types. Graphs are built with a window size of 64 pixels over roughly 500 initial superpixels; the latent encoder uses a downsampling factor of 2. Diffusion runs over 1000 timesteps with a sigmoid schedule from 1e-7 to 2e-3, trained for 250 epochs with Adam at a 3e-4 learning rate. A masking ratio of 50–70% works best.

On tissue classification H-MGDM reaches 82.06% accuracy on the Komura set, 74.51% on PANDA and 86.23% on the IBD data, against 78.14 / 71.92 / 84.58 for DiffMAE, the closest diffusion-based baseline. For survival, features frozen after pancancer pretraining feed AB-MIL, PatchGCN and DeepSurv heads on TCGA-KIRC (512 cases), TCGA-ESCA (155 cases) and a primary-metastatic colorectal cohort (388 cases), with the best concordance indices of 71.17, 64.82 and 66.05. Averaged across downstream tasks the gain over prior pretraining methods is 5.18%.

#Applications

The intended use is as a frozen patch encoder for downstream computational pathology: cancer subtyping, Gleason grading, tissue-type classification, and multiple-instance survival models over whole slides. Groups working on prognostic modelling benefit most, since the survival results come from cohorts entirely separate from the pretraining corpora, and the entity-level attention gives a reviewable account of which tissue regions drove a risk score.

#Impact

H-MGDM shows that the masking unit in self-supervised pathology pretraining is a design choice worth revisiting: aligning it to tissue entities, rather than to a pixel grid, improves both accuracy and interpretability without a larger model or more data. Adoption is limited by what has been released. The repository contains networks, training and downstream code but no pretrained checkpoint, and it carries no license file, so reuse terms are undefined. Graph construction depends on the external Histocartography toolkit, and one of the three pretraining corpora is in-house and not distributable, which makes the pretraining run only partially reproducible.