Журнал "Программная инженерия"
Теоретический и прикладной научно-технический журнал
ISSN 2220-3397

Номер 8 2018 год

DOI: 10.17587/prin.9.369-374
Preprocessing for Enhancing the Classification of Pulmonary Data Sets using Convolutional Neural Networks
N. Esmaeilishahmirzadi, Postgraduate Student, nasibe.smaeili@gmail.com, Lobachevsky State University, Department of Computer Science, Nizhny Novgorod, Russian Federation, H. Mortezapour, hr.mortezapour@gmail.com, Postgraduate Student Ferdowsi University of Mashhad, Department of Computer Engineering, Mashhad, Iran

With growing lung cancer in the world the requirement of computer-aided diagnosis systems (CADs) has increased. In this work we present a method for enhancement of the data collection of pulmonary images using the method of data augmentation and filters in image processing. Our work consists of two components. We firstly present a novel method of image enhancement and secondly classify images into either nodule or non-nodule. For nodule classification the Residual convolutional neural network is proposed, which achieved better performance than state-of-the-art networks in our dataset. Images are extracted from the LUNA16 database derived from the LIDC-IDRI database. For image enhancement we use compound images obtained from image processing filters. These filters are like Gabor and Average and rotate at 90 and 180 degrees to produce images similar to LUNA16 images. We produce 8 images similar to the original image and combine them into one image. We thus produce a new dataset. After testing various convolutional networks on a new dataset we found that the Residual convolutional neural network (ResNet) provides an accuracy of 96.214 % in the classification task. The accuracy demonstrates that the proposed method performs desirably and acceptably compared to other methods.   

Ключевые слова: pulmonary nodule, lung images, residual neural network, deep learning, convolutional neural network CNN
Стр. 369–374