MIHIC-HybridNet: A hybrid vision transformer with multiscale channel–spatial attention for multi-class lung histopathological tissue region classification

Lung cancer is one of the most prevalent and deadly malignancies worldwide, and early diagnosis plays a critical role in improving patient survival rates. Although histopathological examination is the gold standard for the diagnosis of lung cancer, it is insufficient for the comprehensive analysis of tissue biomarkers. Multiplex immunohistochemistry (IHC) histopathological examination offers a more advanced alternative by allowing the visualization of multiple biomarkers in the same tissue section, enabling more detailed analysis of the complex tumor microstructure. This examination provides significant advantages, particularly in the diagnosis of heterogeneous tumors such as lung cancer. This study proposes an attention-based hybrid deep learning model, referred as MIHIC-HybridNet, for seven-class lung tissue region classification from multiplex IHC images. The proposed model integrates Convolutional Neural Networks (CNNs) and the Data-Efficient Image Transformer (DeiT) within a multi-branch framework, enabling the joint modeling of local spatial features and global contextual dependencies. In this model, a novel boosted fusion strategy is introduced to effectively combine CNN and the DeiT representations improving feature integration and overall classification performance. In contrast to conventional feature fusion approaches, the proposed fusion mechanism is structured around dedicated spatial and channel fusion gates, through which complementary local and global representations are adaptively weighted and refined to enhance discriminative feature learning. Furthermore, a Multi-Scale Attention (MSA) mechanism is incorporated into the architecture to enable the model to selectively focus on the most diagnostically relevant regions within the tissue images. The proposed model was trained and validated using the Multiplex IHC Histopathological Image Classification (MIHIC) dataset and ablation analyses are conducted to evaluate the contributions of different attention modules and the proposed fusion mechanisms. Experimental results demonstrate that an accuracy of 85.2% is achieved, and superior performance is obtained compared to state-of-the-art CNN- and Vision Transformer-based methods.