Dual-Stream Attention Fusion with Metaheuristic Feature Optimization for Interpretable Lung Cancer Classification in Medical Imaging

Abstract

Detection and classification of lung cancer in chest computed tomography (CT) images is a pressing issue in medical imaging because it is challenging to differentiate a subtle morphologic distinction between benign and malignant lung nodules and further complicated by the fact that it is difficult to identify pathologic versus normal lung tissue. Conventional single stream deep learning methods do not inherently encode the complementary local and global characteristics that are required in the diagnosis of lung cancer. In this research work, a new dual-stream attention- based feature fusion system of semi-automatic classification of chest CT images on benign, malignant, and normal are proposed to detect lung cancer, overcoming the limitations of the traditional methods due to complementary feature extraction and intelligent fusion mechanisms. The methodology consists of a multiphase pipeline involved with dual-stream feature extractions by InceptionV3 (2 048-dimensional local features) and SE-ResNet18 (512-dimensional global features) frameworks, attention-based fusion mechanism with autonomy gating of adaptive feature weighting, metaheuristic feature selection through Recursive Feature Elimination that yields 98.05% dimensionality reduction (1,024 20 features), and classification optimization by LightGBM with Bayesian hyperparameter optimization.