CTFN: A Hybrid CNN-Transformer Fusion Network with Explainability and Uncertainty Estimation for Enhanced Lung Cancer Classification

Abstract

Timely identification of cancer significantly improves a patient's likelihood of survival. Medical professionals assert that early diagnosis of lung cancer can decrease mortality rates. Broadly used deep learning systems that depend on convolutional neural networks (CNNs) face multiple challenges, including domain discrepancies and limited dataset sizes combined with contextual information deficits and dependence on pre-trained weights, while dealing with data imbalance issues. We propose a new hybrid framework Hybrid CNN-Transformer Fusion Network (CTFN) that combines local feature extraction through Convolutional Neural Networks into a system with Transformer modules specialized in capturing the global structural dependencies between features. In addition, the framework integrates explainability methods (Grad-CAM and occlusion sensitivity) and uncertainty estimation to improve interpretability and clinical reliability. The research utilizes Intel oneAPI to optimize model training on the LC25000 dataset that presents 15,000 histology images because of computation costs and dataset challenges. Our proposed model achieves better results than existing frameworks by delivering 99.6% accuracy alongside an F1-score of 0.9754. The integration between CNN and Transformer layers offers superior classification results that effectively address existing research deficiencies.