MSTF-Net: A Multi-Scale Transformer-Guided Feature Fusion Framework for Accurate Lung Cancer Detection from CT Images

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Lung cancer is one of the most common causes of cancer death worldwide, and early and accurate diagnosis is crucial to upsurge patient survival. Deep learning-based computer-aided diagnosis systems have demonstrated promising results, however they are difficult to recognise fine-grained local lesion traits and long-range contextual connections from CT scans. To overcome these limitations, to propose a Multi-Scale Transformer-Guided Feature Fusion Network for lung cancer diagnosis using the LIDC-IDRI dataset, named MSTF-Net. The proposed framework consists of CT image preprocessing, attention-guided lung nodule segmentation, multi-scale convolutional feature extraction, dual attention augmentation, Transformer-based contextual learning, adaptive cross-scale feature fusion and optimised classification. Multi-scale branches capture pulmonary nodules of various sizes and Transformer encoder models global contextual interactions of pulmonary structures. Experiments indicate the effectiveness of the framework. MSTF-Net outperforms Shuffle-DNN, CCHCO-DenResFT-Net, CHC-VO DCNN, FVCM-Net, and C-Swin with 99.14% accuracy, 98.96% precision, 98.72% recall, 98.84% F1-score, and 99.31% specificity. The lung nodule segmentation model reached Dice coefficient 97.42% and IoU 94.98%. The results suggest that MSTF-Net is a robust framework for automated lung cancer identification and can assist radiologists to make clinical decisions.