Abstract <p>Deep learning has achieved great success in handling the classification of high-resolution images, such as whole slide images (WSIs). However, traditional deep learning methods cannot directly process high-resolution images. They often resort to random sampling, wasting substantial resources on uninformative and redundant image patches, leading to inefficient utilization of WSIs. To overcome these shortcomings, we propose a novel high-resolution whole slide images classification network based on adaptive core regions(ACRNet). First, we retrieve high-probability informative patches from the sampled images using a pathology feature database and cosine similarity. Next, we construct a dynamic graph convolutional network with adaptive regions to learn contextual information from these high-probability patches. This includes the construction of the dynamic graph and the adaptive region GCN (Graph Convolutional Network) to generate vector representations of the contextual regions for final classification prediction. We conducted experiments on the TCGA-LUSC dataset and SHFPH-COAD dataset from our partner hospital. The results show that the proposed model achieved accuracy and AUC comparable to the advanced TransMIL method on both datasets, while reducing inference time by 16.3% and 21.8%, respectively.</p> Graphical abstract <p></p>