@Article{cmes.2026.078892,
AUTHOR = {Muhammad Waqar, Zeshan Aslam Khan, Arthur Chang, Zhishan Guo, Chun-Liang Lai, Chuan-Yu Chang},
TITLE = {Explainable Context-Aware Fusion Network for Non-Small Cell Lung Cancer Analysis with Application to Smart Healthcare Systems},
JOURNAL = {Computer Modeling in Engineering \& Sciences},
VOLUME = {147},
YEAR = {2026},
NUMBER = {1},
PAGES = {0--0},
URL = {http://www.techscience.com/CMES/v147n1/67134},
ISSN = {1526-1506},
ABSTRACT = {Lung cancer (LC) is among the dangerous cancers spreading progressively, and a timely LC diagnosis becomes a dire need of the time. Various imaging-based studies have been conducted for accurate LC examination through computed tomography (CT), X-ray, and histopathology. Worldwide, the proportion of LC-affected patients in hospitals is growing, thereby increasing imaging data for fast processing and early examination. To facilitate histopathological imaging-based automated and timely decision making for accurate LC prediction, a Context Aware Fusion Network (CAFNet) for holistic feature learning and spatially localized feature learning is proposed in this study for the efficient extraction and processing of global as well as local features, respectively. CAFNet exploits histopathological tissues to ensure local and global attributes uniformity for extracting contextual information. The conducted research achieves histopathological image enhancement using median filtering (MF) and contrast-limited-adaptive-Histogram-Equalization (CLAHE). Moreover, the classifying power of the proposed CAFNet is enhanced through superior attributes extraction strategies, such as Mobile Inverted Bottleneck Convolution (MIBConv) employed with Spatial Attention with Residual Learning (SARL) and Channel Attention with Residual Learning (CARL). An innovative, partially adaptive optimization approach is utilized to fine-tune the degree of adaptivity in the learning process of the network. The descriptive behavior of CAFNet is explored through explainable artificial intelligence (XAI) strategies like Gradient-Weighted Class Activation Mapping (GradCAM) and Local Interpretable Model-Agnostic Explanation (LIME). The proposed network achieved an improved average classification accuracy of 7.36% while reducing models’ complexity by 85% to 99% as compared to the existing benchmark models. The study also addresses users’ accessibility challenges by providing a web-based interface using Gradio for users’ real-time interaction.},
DOI = {10.32604/cmes.2026.078892}
}