Hybrid Transformer-Based Framework for Multi-Type Lung Cancer Detection Using Attention-Guided Feature Fusion

Lung cancer is among the deadliest cancers worldwide, and early, accurate diagnosis can vastly improve survival. To address this work, we present a bool solution based on a transformer-based integrated framework for lung cancer detection on multi-type chest X-rays. The model combines two state-of-the-art Transformer-based architectures, the Swin-Transformer and the Vision Transformer (ViT), to mine local and global visual features in the images. The Swin Transformer uses a shifted window mechanism to capture hierarchical spatial features, while the ViT helps bring more contextual understanding to the whole image. These complementary features are then merged intelligently at the Attention Fusion Layer, which adaptively learns to weigh those extracted features, enriching the most representative parts while suppressing redundancy. The MLP classifier receives the cleaned feature vector for multi-class classification of lung conditions such as adenocarcinoma, large cell carcinoma, squamous cell carcinoma, and normal. We have performed a series of wide-ranging experiments on the publicly accessible Figshare Lung Cancer dataset, consisting of expert-labelled, variegated chest X-ray images, to demonstrate the potential of our method. Experimental results indicate that the proposed hybrid framework outperforms the UperNet and Swin-Unet models, yielding superior accuracy, precision, and F1-score across all classes. Overall, our findings suggest that hierarchical and global feature extractors fused in an attention-based way and classified using an MLP offer a strong and accurate approach to lung cancer detection. An ablation study evaluates the contributions of the proposed transformer architecture, UperNet and Swin-Unet. Simulated results show ViT-Swin-AGFF achieves superior classification performance over UperNet and Swin-Unet, highlighting the efficacy of AGFF in combining global and local features. For segmentation, the ViT-Swin-AGFF model outperforms the state-of-the-art UperNet and Swin-Unet. These findings suggest that hybrid transformer architectures with attention-based fusion enhance diagnostic accuracy, offering a promising approach for lung cancer classification in clinical settings. Future work will validate results with real-world datasets and explore segmentation mask integration. This approach has great potential as a computer-aided diagnostic tool to assist radiologists in making clinical decisions.