Automated Lung Cancer Classification from 3D CT Scans Using Deep Convolutional Networks

Abstract

The research gives an artificial intelligence-based system that was designed to enhance early prediction and classification of lung cancer using 3D chest CT-scanned images. 3The system employs the use of deep learning by integrating a lean pipeline, segmentation of the lungs, detection of nodule candidates, and the subsequent classification of malignancy. In the system, lung tissue is initially isolated by threshold-based segmentation to distinguish it from other anatomy. However, it was not efficient to simply input whole segmented volumes in 3D Convolutional Neural Networks (3D-CNNs) due to high levels of noise and search space. To reduce this, a U-Net was developed as a modified version, which was trained on the LUNA16 database, which contains labeled nodules to identify candidate regions of the nodules. Such regions of candidates may allow for including false positives at times, but in turn, they allow the model to focus on the most relevant features. The defined regions are subsequently classified with a light and effective 3D-CNN, which achieved 92.6 % test accuracy. Unlike traditional CAD systems, our AI-based approach requires minimal amounts of labeled data and multiple (and complex) stages, and continues delivering high accuracy and generality. It then creates a compelling reason why it could be a good applicant in the search for lung cancer in a medical image, with scalability, efficiency, and early detection of lung tumors.

 Keywords: Artificial intelligence, lung cancer, CT imaging, deep learning, 3D convolutional neural networks, medical diagnostics.