Transformer-Driven Dual Attention Volatility Network Enhanced by Federated Learning for Lung Cancer Detection and Classification

Lung cancer is a deadliest cancer at the global level, which requires accurate and early-stage diagnosis. The traditional systems also face various complexities. To solve these issues, this paper proposes a novel Federated Transformer-Based Dual Attention Volatility Network for lung cancer classification. In the initial stage the input images are subjected to preprocessing and augmentation approaches to improve the data quality. This model implements a decentralized federated learning architecture to ensure privacy-preserving model training over multiple clients. Federated Transformer-Based Dual Attention Volatility Network incorporates a Regime-Specific Transformer Block as its core element for extracting contextual spatial features and a Dual Channel Attention Mechanism for modeling the short- and long-range dependencies. A specific volatility classification branch is also incorporated to detect inter-class variations. Four different lung Computed Tomography image datasets were used in this work to elaborate experiments on the lung histopathology dataset exhibits that the Federated Transformer-Based Dual Attention Volatility Network achieved a superior accuracy of 98.02%, and outperforms conventional centralized and federated approaches based on accuracy and interpretability. The output of experiments conducted also revealed that the proposed approach attained a precision of 97.89%, recall of 98.23%, F1-score of 98.05%, and specificity of 98.16% and exhibited effective performance.