Improving the Pulmonary Nodule Classification Based on KPCA-CNN Model

Abstract

Lung cancer mortality, the main cause of cancer-associated death all over the world, can be reduced by screening risky patients with low-dose computed tomography (CT) scans for lung cancer. In CT screening, radiologists will have to examine millions of CT pictures, putting a great load on them. Convolutional neural networks (CNNs) with deep convolutions have the potential to improve screening efficiency. In the examination of lung cancer screening CT images, estimating the chance of a malignant nodule in a specific location on a CT scan is a critical step. Low-dimensional convolutional neural networks and other methods are unable to provide sufficient estimation for this task, even though the most advanced 3-dimensional CNN (3D-CNN) has extremely high computing requirements. This article presents a novel strategy for reducing false positives in automatic pulmonary nodule diagnosis from 3-dimensional CT imaging by merging a kernel Principal Component Analysis (kPCA) approach with a 2-dimensional CNN (2D-CNN). To recreate 3-dimensional CT images, the kPCA method is utilized, with the goal of reducing the dimension of data, minimizing noise from raw sensory data while maintaining neoplastic information. The CNN can diagnose new CT scans with an accuracy of up to 90% when trained with the regenerated data, which is better than existing 2D-CNNs and on par with the best 3D-CNNs. The short duration of training, and certain accuracy shows the potential of the kPCA-CNN to adapt to CT scans with different parameters in practice. The study shows that the kPCA-CNN modeling technique can improve the efficiency of lung cancer diagnosis.