Abstract:Predicting labels of gigapixel whole-slide images (WSIs) and localizing regions of interest (ROIs) with high precision are of great interest in computational pathology. The existing methods are mainly based on multi-instance learning (MIL) approach and its variants. However, such algorithms primarily treat the instances of slides as independent samples, which cannot effectively describe the relationship between instances in WSI. This would cause only a subset of high-score instances can be located, which is not applicable in clinical scenarios. Context-aware attention fusion adaptive hypergraph convolution network (CAF-AHGCN) is proposed to adaptively establish the local and global topology of cropped image patches spatially arranged in the slides. The framework combines hypergraph embedding representation with attention-based MIL pooling aggregation in a hierarchical feature fusion manner, which fully preserves high-order spatial structural correlations of the patches to make a slide-level prediction. Here, an adaptive hypergraph convolutional network is designed to dynamically adjust the correlation strength between the hypergraph nodes as the model goes deeper. Poly-1 loss and residual connection are also applied to prevent over-smoothing and improve the generalization ability of deep CAF-AHGCN model. We verified the superiority of CAF-AHGCN on two datasets, CAMELYON16 and TCGA-NSCLC. The results showed that the ACC, AUC and F 1 score predicted by our model outperform other state-of-the-art algorithms on both datasets. The heat maps obtained by CAF-AHGCN are highly consistent with the pixel-wise annotated label of the WSI. The results show that CAF-AHGCN not only achieves high-accuracy label prediction for WSI, but also provides patch-wise human-interpretable features in ROI localization heatmaps. The outstanding performance of CAF-AHGCN framework provides a new perspective for future clinical applications of computer-aided diagnosis and intelligent systems.

Hierarchical Attention Guided Framework for Multi-resolution Collaborative Whole Slide Image Segmentation

Super-Resolution Based Patch-Free 3D Medical Image Segmentation with Self-Supervised Guidance

Generating Hypergraph-Based High-Order Representations of Whole-Slide Histopathological Images for Survival Prediction

HSG-MGAF Net: Heterogeneous Subgraph-Guided Multiscale Graph Attention Fusion Network for Interpretable Prediction of Whole-Slide Image

DHUnet: Dual-branch Hierarchical Global–local Fusion Network for Whole Slide Image Segmentation

Memory-Efficient Sparse Pyramid Attention Networks for Whole Slide Image Analysis

Semi-Supervised Instance Segmentation in Whole Slide Images via Dense Spatial Variability Enhancing

DSCA: A dual-stream network with cross-attention on whole-slide image pyramids for cancer prognosis

Multi-scale representation attention based deep multiple instance learning for gigapixel whole slide image analysis

A deep learning based multiscale approach to segment cancer area in liver whole slide image

HoloHisto: End-to-end Gigapixel WSI Segmentation with 4K Resolution Sequential Tokenization

CAF-AHGCN: context-aware attention fusion adaptive hypergraph convolutional network for human-interpretable prediction of gigapixel whole-slide image

Multi-scope Analysis Driven Hierarchical Graph Transformer for Whole Slide Image Based Cancer Survival Prediction

Histopathology WSI Encoding based on GCNs for Scalable and Efficient Retrieval of Diagnostically Relevant Regions

Gigapixel Whole-Slide Images Classification using Locally Supervised Learning

Cancer Sensitive Cascaded Networks (Csc-Net) for Efficient Histopathology Whole Slide Image Segmentation

A hybrid network for automatic hepatocellular carcinoma segmentation in H&E-stained whole slide images

ConSlide: Asynchronous Hierarchical Interaction Transformer with Breakup-Reorganize Rehearsal for Continual Whole Slide Image Analysis

W-Net: Dual Supervised Medical Image Segmentation Model with Multi-Dimensional Attention and Cascade Multi-Scale Convolution

Encoding Histopathological WSIs Using GNN for Scalable Diagnostically Relevant Regions Retrieval.

Deep Hierarchical Representation And Segmentation Of High Resolution Remote Sensing Images