Abstract
Computed tomography (CT) scans require precise and early lung cancer detection to produce better clinical results. High accuracy in deep learning approaches (DL) poses an existing challenge to interpret their functionality effectively. This research presents an innovative modular multi-backbone structure that combines channel-spatial attention together with explainable AI (XAI) methods for three-class lung cancer diagnosis (Normal, Benign, and Malignant). Research was carried out to evaluate six pre-trained CNN backbones (ResNet-50, VGG19, Inception-V3, EfficientNet-B0, MobileNet-V2, DenseNet-121) which received hybrid attention enhancement on the IQ-OTH/NCCD dataset. The experimental data showed four pre-trained models reaching perfect accuracy at 100 percent whereas the others attained precision, recall and F1-score rates of 99 percent. The training process of EfficientNet-B0 resulted in superior accuracy alongside minimal training duration. The XAI techniques (LIME, SHAP, Grad-CAM++) showed success in producing visual explanations which proved the model focused on important clinical areas and improved its interpretability. The framework provides optimal diagnostic performance alongside essential interpretability capabilities to present a strong and dependable system for lung cancer detection through computer assistance.
Recommended Citation
Obaid, Omar Ibrahim and ALazzawi, Abdulbasit
(2025)
"CSA-XAI: Channel–Spatial Attention and Explainable AI in a Modular Multi-Backbone Framework for Lung Cancer Classification,"
Iraqi Journal for Computer Science and Mathematics: Vol. 6:
Iss.
3, Article 26.
DOI: https://doi.org/10.52866/2788-7421.1286
Available at:
https://ijcsm.researchcommons.org/ijcsm/vol6/iss3/26
