Convolutional Neural Network based Malignancy Detection of Pulmonary Nodule on Computer Tomography

Abstract

Without performing biopsy that could lead physical damages to nerves and vessels, Computerized Tomography (CT) is widely used to diagnose the lung cancer due to the high sensitivity of pulmonary nodule detection. However, distinguishing pulmonary nodule in-between malignant and benign is still not an easy task. As the CT scans are mostly in relatively low resolution, it is not easy for radiologists to read the details of the scan image. In the past few years, the continuing rapid growth of CT scan analysis system has generated a pressing need for advanced computational tools to extract useful features to assist the radiologist in reading progress. Computer-aided detection (CAD) systems have been developed to reduce observational oversights by identifying the suspicious features that a radiologist looks for during case review.

Most previous CAD systems rely on low-level non-texture imaging features such as intensity, shape, size or volume of the pulmonary nodules. However, the pulmonary nodules have a wide variety in shapes and sizes, and also the high visual similarities between benign and malignant patterns, so relying on non-texture imaging features is difficult for diagnosis of the nodule types. To overcome the problem of non-texture imaging features, more recent CAD systems adopted the supervised or unsupervised learning scheme to translate the content of the nodules into discriminative features. Such features enable high-level imaging features highly correlated with shape and texture.

Convolutional neural networks (ConvNets), supervised methods related to deep learning, have been improved rapidly in recent years. Due to their great success in computer vision tasks, they are also expected to be helpful in medical imaging. In this thesis, a CAD based on a deep convolutional neural network (ConvNet) is designed and evaluated for malignant pulmonary nodules on computerized tomography. The proposed ConvNet, which is the core component of the proposed CAD system, is trained on the LUNGx challenge database to classify benign and malignant pulmonary nodules on CT. The architecture of the proposed ConvNet consists of 3 convolutional layers with maximum pooling operations and rectified linear units (ReLU) activations, followed by 2 denser layers with full-connectivities, and the architecture is carefully tailored for pulmonary nodule classification by considering the problems of over-fitting, receptive field, and imbalanced data.

The proposed CAD system achieved the sensitivity of 0.896 and specificity of 8.78 at the optimal cut-off point of the receiver operating characteristic curve (ROC) with the area under the curve (AUC) of 0.920. The testing results showed that the proposed ConvNet achieves 10% higher AUC compared to the state-of-the-art work related to the unsupervised method. By integrating the proposed highly accurate ConvNet, the proposed CAD system also outperformed the other state-of-the-art ConvNets explicitly designed for diagnosis of pulmonary nodules detection or classification.