HyCANet: a hybrid convolutional-attention network with cross-attention fusion for lung cancer histopathological...

Histopathology interpretation of lung cancer involves careful examination of tissue morphology, especially differentiating benign lung tissue, adenocarcinoma and squamous cell carcinoma. While deep-learning techniques have demonstrated promising success in digital pathology, a significant challenge remains in capturing both cellular and tissue-level context in a single model. To tackle this problem, in this paper, we propose a hybrid convolutional-attention network (HyCANet) for lung cancer histopathological image classification. The model is designed in a parallel architecture with two branches, one of which is Dynamic-CNN branch for learning local morphological patterns and the other is lightweight Vision Transformer branch for learning long-range contextual information within the image regions. In both of these two feature streams, the local and global representations are coupled together via a Cross-Attention Feature Fusion (CAFF) module, where both representations are allowed to interact prior to final classification. The performance of the proposed model was validated using the lung subset of the LC25000 dataset, which consists of images of lung tissue, lung adenocarcinoma, and lung squamous cell carcinoma. On the corrected stratified test set of 2,250 images, HyCANet achieved an accuracy of 96.44%, precision of 96.21%, recall of 95.97%, and F1-score of 96.08%. The model outperforms the CNN-only, transformer-only, serial hybrid, and pathology-specific encoder baselines. The ablation results further demonstrated that the Dynamic-CNN branch, ViT-lite branch and CAFF module all played a role in the final performance. The results were further validated on BreakHis, CRC-Tissue, TCGA-Lung, and PathMNIST, and HyCANet showed competitive performance on all the datasets. In general, the outcomes suggest that HyCANet is an appealing and understandable model for lung cancer histopathological image classification and would require further validation using patient-level, slide-level and multicenter clinical data sets.