LungFormer-XAI: A Dual-Attention CNN–Transformer Framework for Explainable Lung Cancer Detection and Classification learning

Abstract

Lung cancer is one of the leading causes of cancer-related mortality worldwide, highlighting the need for accurate and timely diagnosis. The capabilities of the Computer-Aided Diagnosis (CAD) system are significantly enhanced with the development of deep learning, but the existing approaches are not sufficient to capture the local and global contextual information and features related to lung lesions, are limited in terms of interpretability, and only utilize a single imaging modality. This research suggests a new multimodal deep learning framework called LungFormer-XAI to overcome these limitation. It uses DenseNet121, Vision Transformers, multimodal feature fusion, Explainable AI (XAI), and dual attention mechanisms to automatically detect and classify Lung Cancer. This method uses Grad-CAM and SHAP to give visual explanations in an understandable manner. Experiment results show that LungFormer-XAI had an accuracy of 97.2%, which is higher compared to other models such as ResNet50 (92.8%), DenseNet121 (94.3%), and EfficientNet-B0 (95.1%). In addition, the proposed framework had a precision of 96.8%, a recall of 97.5%, specificity of 96.6%, an F1 score of 97.1%, and an AUC-ROC of 0.989. Therefore, the results prove that LungFormer-XAI is an accurate and clinically applicable model for automated lung cancer diagnosis.