Prediction of Lung Cancer from Low-Resolution Nodules in CT-Scan Images by using Deep Features

Abstract

Prediction of lung cancer from CT-scan images is viewed as a challenging task in medical image analysis. It is because pulmonary nodules occupy less than 5% of a CT-scan image and vary highly in terms of their shape, texture, opacity and location, making it difficult to analyze by naked eye. To overcome this challenge, multiple approaches using digital processing of CT-scans are proposed. These existing approaches lack in addressing three main challenges in lung cancer prediction: the problem of low resolution data, lack of generalized nodule features, and cancer prediction in terms of a malignancy score. These challenges are addressed in the paper. In the proposed approach, CT-scan images are preprocessed and super-resolved to serve as input to a convolutional neural network for nodule region detection. Deep features engineered from different layers of this convolutional neural network are classified into two classes: benign and malignant. Additionally, these features are passed through a regressor to predict malignancy score. CT-scan images from a subset of the LIDC-IDRI dataset, called the LUNA16 dataset, are used for experimentation. An accuracy of 85.7% is achieved for the nodule classification task using the approach proposed in this paper, which is a substantial improvement over previous works in literature.