An adaptive feature fusion framework of CNN and GNN for histopathology images classification

Abstract

This study introduces an Adaptive Feature Fusion Classification Network (AFFCNet) designed for cancer detection in histopathology images. AFFCNet leverages Convolutional Neural Network (CNN) and Graph Neural Network (GNN) as parallel feature extractors, significantly improving the ability to capture complex histopathological features. The network includes an adaptive feature fusion module that weights and fuses features from the two branches using adaptive scaling factor and attention mechanisms. The fused features are subsequently utilized to construct a graph structure. The global feature aggregation unit then performs sampling and aggregation on this graph to extract high-level semantic features. Experimental results demonstrate the effectiveness of AFFCNet. On the BRACS dataset, a breast cancer subtype pathology image dataset containing 4391 images, the model achieved an F1-score of 67.23 %, representing a 2.83 % improvement over previous methods. On the LC25000 dataset, a pathology image dataset of lung and colon tissues containing 25,000 images, it achieved Precision, Recall, Specificity, Accuracy, and F1-score of 99.84 %, 99.84 %, 99.96 %, 99.84 %, and 99.84 %, respectively, showing improvements of 1.36 %, 0.37 %, 0.34 %, 0.45 %, and 0.52 % compared to existing approaches. These results highlight AFFCNet's capability to leverage advanced semantic features and achieve competitive performance compared to state-of-the-art methods.