Dual-stage classification for lung cancer detection and staging using hybrid deep learning techniques

Lung cancer, a malignant disease originating in the lungs, presents significant challenges in early detection and staging. It occurs when abnormal lung cells grow uncontrollably, forming tumors that disrupt lung function. Timely detection is vital for better treatment outcomes. Staging, which assesses the cancer’s extent and severity, guides treatment decisions and prognosis predictions. Lung cancer diagnosis and staging face obstacles like symptom absence, similar imaging findings to other lung conditions, limited screening methods, invasive biopsies, complex staging procedures, variable tumor behavior, metastasis detection challenges, clinical overlap, comorbidities, and observer variability. Overcoming these hurdles is crucial for improving lung cancer care. In this study, we present a dual-stage classification model aimed at detecting and staging lung cancer using a combination of advanced deep learning techniques. We introduce a distinctive approach that commences with the development of a modified U-Net incorporating dual attention and pyramid atrous pooling. This modification enhances target segmentation accuracy, ultimately leading to improved detection precision. We further enhance our methodology by extracting texture, color, and shape features from the segmented target area. In the initial classification stage, we employ a hybrid Xception and custom CNN model, effectively distinguishing between normal and abnormal cases for tumor detection. In the subsequent stage, we extract additional locational features from the abnormal characteristics, utilizing them as input for our innovative hybrid adaptive learning neural network to achieve accurate lung cancer staging. This multistage approach represents a significant novelty in our study, aiming to enhance both detection and staging of lung cancer. To validate the model’s performance, we conducted experiments on several datasets, including LIDC-IDRI, NSCLC-radiomics–genomics, NSCLC-radiomics, and NSCLC radiogenomics. Our results demonstrate the effectiveness of our model in comparison with existing methods, as assessed through various quality evaluation metrics.