An intelligent generative adversarial network multistage lung cancer detection and subtypes classification

Cancer remains a leading cause of global mortality, with lung cancer having the highest mortality rate. Early detection through Computed Tomography (CT) imaging is essential for accurately assessing lung cancer stages. However, existing deep-learning classification models for lung CT images often suffer from low efficiency and accuracy in early cancer diagnosis and are typically time-consuming. To address these challenges, this research introduces a novel two-stage network approach, the Deep Convolutional cross max-out kernel graph-based Generative Adversarial Network with Enhanced Prism refraction Search (DCGAN-EPS). This model enhances tumor detection accuracy and efficiency through advanced techniques. In Stage 1, adaptive wavelet denoising reduces noise and eliminates irrelevant pixels in lung CT images, followed by precise image segmentation using an improved attention U-Net. Stage 2 utilizes a weighted local binary pattern to improve feature extraction for tumor detection. The Enhanced Prism refraction Search (EPS) algorithm optimizes the neural network's loss parameters, significantly boosting detection accuracy. Additionally, integrating local interpretable model-agnostic explanations enhances the model's transparency and interpretability. Experimental results on the LIDC-IDRI and Chest CT-Scan datasets demonstrate the proposed method's robust performance, achieving accuracy, recall, precision, F1-Score, and specificity scores of 99.92%, 99.94%, 99.94%, 99.92%, and 99.90% on the LIDC-IDRI dataset, and scores of 99.4%, 99.2%, 99.5%, 99.4%, and 99.2% on the Chest CT-Scan dataset, underscoring its efficiency and suitability for clinical applications.