A Hybrid Optimization Framework Based on a Lightweight Dual-Branch Swin Transformer for Early Lung Cancer Detection and Classification

Abstract

Lung cancer arises from the uncontrolled growth of cells within lung tissues, posing a significant threat to respiratory functions crucial for oxygenation and carbon dioxide elimination in the human body. Early detection of malignant cells in the lungs is paramount, yet challenging due to the asymptomatic nature of early-stage lung cancer. Existing deep learning-based models for lung cancer detection require computational resources, presenting implementation challenges for healthcare centers. Thus, there is an urgent need for a sensitive and accurate approach to early lung cancer diagnosis. In this study, we propose an innovative approach to early lung cancer detection and classification, leveraging a hybrid process that integrates optimal segmentation, feature extraction, feature selection, and classification techniques. Our methodology combines the modified pity beetle optimization (MPBO) algorithm for optimal segmentation of CT scan and X-ray images, followed by feature extraction and optimization using DefenseNet with normative fish swarm optimization (NFSO) algorithm. Furthermore, we utilize lightweight dual-branch Swin transformer (LDB-ST) for early detection and classification of lung cancer subtypes, including adenocarcinoma, large cell carcinoma, and squamous cell carcinoma, in addition to identifying healthy lung tissue. We evaluate the performance of our model using benchmark datasets such as Chest CT Scans, LIDC-IDRI, and LUNA16. Comparative analysis with existing models demonstrates the superior performance of the proposed approach, with the LDB-ST model achieving a significant improvement of 5.801% in accuracy and 6.061% in F-measure. These findings highlight the potential of the proposed methodology to revolutionize early lung cancer diagnosis, offering enhanced accuracy and efficiency in clinical settings.