Towards automated and reliable lung cancer detection in histopathological images using DY-FSPAN: A feature-summarized pyramidal attention network for explainable AI

Abstract

Medical image classification is critical for accurate disease diagnosis, necessitating models that balance performance and interpretability. This study presents Dilated Y-Block-based Feature Summarized Pyramidal Attention Network (DY-FSPAN), a deep learning framework designed for robust feature extraction and classification. The architecture incorporates Y-blocks and attention mechanisms to enhance spatial feature representation while maintaining receptive field coherence. The proposed model achieves a classification accuracy of 98.5 %, surpassing existing approaches such as convolutional block attention networks, adversarial learning models, and multi-output 3D CNNs. To validate the efficacy of DY-FSPAN, we conduct an extensive experiment, including comparative benchmarking against state-of-the-art methods, robustness assessments, and ablation studies. The model’s structural improvements are tested through various configurations to assess the impact of key components, confirming the contribution of attention mechanisms to performance enhancement. Grad-CAM analysis was employed to visualize learned feature maps, highlighting the model’s focus on diagnostically relevant regions, thereby improving trust in AI-driven medical decision-making. From an explainable AI perspective, the proposed framework achieves superior classification accuracy and enhances interpretability, addressing a crucial requirement in medical imaging applications. The qualitative and quantitative analyses demonstrate that DY-FSPAN effectively localizes disease-specific features, making it a suitable tool for clinical use. The findings suggest that integrating attention-based architectures with optimized feature selection can significantly advance automated medical diagnosis. The model’s ability to improve diagnostic reliability while maintaining transparency underscores its potential for real-world deployment in healthcare settings.