Lung Nodule Identification and Classification of Lung CT Images using Deep Learning

Abstract

Lungs are very important organs of the human body and the respiratory system. And therefore, lung cancer claims a maximum number of deaths among all the cancers. Lung cancer starts from lung nodule which is converted and may result in malignant tumors. Identifying lung nodules at an early stage is a challenging task. By applying machine learning and image processing algorithms, reliability in the identification of lung nodules has been obtained to some extent.

Deep learning is evolving under machine learning that has claimed the efficiency in pattern recognition in several domains. Convolution Neural Network (CNN) which is a technique of deep learning that provides the advantage for automatically extracting the features from the images using convolution and max pooling methods. CNN models human brain activity from sensory inputs and has been applied to various activities such as driving cars, recognizing speech, and playing complex games. These machine learning algorithms gained popularity in recent years as an effective method for pattern recognition. A novel CNN architecture is proposed here to detect lung cancer in radiographic images.

In this research, the lung nodules have been classified using the CNN and a framework has been designed for earlier identification of lung nodules and their classification from the lung CT images. The computer aided diagnosis system proposed in here can assist the radiologists in cancer tumor identification based on various facts and studies done previously. The system can simplify the process of identification even in earlier stages by adding up the facility of a second opinion which makes the process simpler and faster.

The framework proposed here is made up of CNN, which is a technique under Deep Learning. The research work implements the framework on AlexNet and ZFNet architectures and one more proposed architecture and has trained the system for tumor detection in lung CT images for all these architectures. The tumor positions have also been pre dicted and outlined on the images along with the actual tumor positions. The accuracy for classification in tumorous and non-tumorous classes are 96.73%, 96.59%, and 99.31% respectively for AlexNet, ZFNet and the proposed architecture for taken dataset of lung CT images while the accuracy for classification in benign and malignant classes are 95.14%, 95.49%, and 99.19% respectively for AlexNet, ZFNet and the proposed architecture.