Hybrid CNN-Based Transformer Pipeline With Radiomic Fusion for Multi-Class Lung Cancer Detection

Abstract

Early detection of lung cancer remains challenging due to high intra-class variation and inter-class similarity in Computed Tomography (CT) images. In this paper, we propose a hybrid deep learning model that combines convolutional, attention-based, and transformer-guided representations to address these challenges in multi-class lung cancer classification. For deep feature extraction, we use the EfficientNetV2-S architecture augmented with a Convolutional Block Attention Module to emphasize salient spatial and channel information. A transformer encoder captures global contextual dependencies, and texture-based radiomic features are incorporated to further enrich the representation. The resulting features are fused into a single embedding, which is then classified as normal, benign, or malignant. Experiments on the IQ-OTHNCCD dataset demonstrate that the proposed framework achieves superior performance across multiple metrics, accuracy, recall, precision, F1 score, and AUC, and outperforms state-of-the-art methods.