Timely Detection of Lung Nodule Malignancy Using 3D Convolutional Neural Networks

Lung cancer remains one of the leading causes of cancer death worldwide, with most cases diagnosed at a late stage. This paper proposes a 3D convolutional neural network (3D CNN) model to classify accurately malignant lung nodules from computed tomography (CT) scans in an attempt to help radiologists and enhance patient care. With the publicly available LIDC-IDRI dataset of 1,018 CT scans on 1,010 patients, our method uses volumetric context to gain spatial information of nodules, a domain where conventional 2D methods fall short. The proposed novel 3D CNN achieves a classification accuracy of 93.0%, sensitivity of 91.0%, and area under the curve of 0.95, which are significantly better than a baseline 2D CNN (85.0% accuracy) and a support vector machine model developed from radiomics (80.0% accuracy). This study introduces a novel multi-scale 3D CNN with spatial and channel attention for malignancy detection, outperforming 2D CNN and support vector machine baselines. The model leverages volumetric CT context and task-specific preprocessing, including voxel normalization, intensity clipping, and data augmentation to extract discriminative features. Our architecture demonstrates robust performance across varied nodule types and has strong potential as a computer-aided diagnosis tool in early lung cancer detection.