An Intelligence-Based Hybrid CNN-GAT Framework Optimized by the Whale Optimization Algorithm for Clinical Lung Cancer Classification from Chest CT Images.
Authors
Affiliations (6)
Affiliations (6)
- Department of Computer Engineering, Ard.C., Islamic Azad University, Ardabil, Iran. [email protected].
- Electrical and Computer Engineering, University of Denver, Denver, USA.
- Department of Computer Engineering, Ard.C., Islamic Azad University, Ardabil, Iran. [email protected].
- Department of Computer Engineering, Ard.C., Islamic Azad University, Ardabil, Iran.
- Department of Computer Engineering, CT.C., Islamic Azad University, Tehran, Iran.
- Department of Computer Engineering, Ta.C., Islamic Azad University, Tabriz, Iran.
Abstract
Early and accurate detection of lung cancer remains a central challenge in intelligence-based medicine, where robust imaging informatics solutions are required to interpret complex chest CT data. This study proposes a novel hybrid convolutional neural network-graph attention network (CNN-GAT) framework, optimized by the Whale Optimization Algorithm (WOA), for clinical three-class classification of lung cancer (benign, malignant, normal) from chest CT images. The proposed architecture employs a ResNet-18 backbone to extract deep spatial representations, which are subsequently modeled as a graph structure and processed by graph attention layers to capture non-linear relational dependencies among localized CT features. To enhance generalization and mitigate overfitting, WOA is used for automated hyperparameter tuning, including learning rate, batch size, hidden channel dimensions, and dropout rates. The framework is evaluated on a curated chest CT dataset and benchmarked against standard CNN and compound EfficientNet architectures. Experimental results demonstrate that the proposed intelligence-based framework substantially outperforms the baseline models, achieving a test accuracy of 98.7%, precision of 98.5%, recall of 98.9%, F1-score of 98.7%, a Matthews correlation coefficient of 0.975, and an area under the curve of 0.994. In addition, the model is highly efficient, with approximately 4.1 million trainable parameters and an average inference time of 0.035 s per CT scan, making it suitable for real-time deployment. In conclusion, integrating graph-based topological intelligence with meta-heuristic optimization on top of a lightweight CNN backbone yields a highly accurate, generalizable, and computationally efficient diagnostic framework. The proposed CNN-GAT + WOA model shows strong potential for seamless integration into clinical workflows as an automated decision-support tool for high-stakes lung cancer screening from chest CT images.