ENHANCING LUNG CANCER DETECTION: A DEEP LEARNING APPROACH WITH HYBRID-NET ARCHITECTURE ON THE LUNA16 DATASET

Abstract

This study focuses on lung cancer detection using deep learning techniques applied to the LUNA16 dataset, a comprehensive collection of CT scans specifically curated for pulmonary nodule analysis. The proposed methodology encompasses meticulous data preprocessing techniques, including denoising, normalization, image resampling, and data augmentation, aimed at optimizing the input data for robust model training and evaluation. The core of this research introduces the Hybrid-Net architecture, a novel framework tailored for lung cancer identification. This architecture integrates spatial-channel-temporal (SCT) attention networks and a unique hybrid pooling technique, designed to capture intricate patterns and global dependencies within the data. The model incorporates spatial and channel attention modules to enhance feature representation, followed by a temporal attention layer to detect temporal relationships within the sequential data. The study outlines the mathematical formulations and mechanisms of each component in the Hybrid-Net architecture, detailing the convolutional modules, attention mechanisms, and pooling operations. The results of the experiments demonstrated enhanced performance for the proposed framework in automating the segmentation of lungs and identifying infected areas in CT scan images. The framework achieved overall dice accuracies of 0.99 for the prediction of lung cancer disease.