Optimal detection of suspected lung nodules using a novel convolution neural network

Subject Areas : journal of Artificial Intelligence in Electrical Engineering

Reza Majidpourkhoei ¹ , Mehdi Alilou ² , Kambiz Majidzadeh ³ , Amin BabazadehSangar ⁴

1 - Department of Computer engineering, Urmia Branch, Islamic Azad University, Urmia, Iran
2 - Department of Computer engineering, Urmia Branch, Islamic Azad University, Urmia, Iran
3 - Department of Computer engineering, Urmia Branch, Islamic Azad University, Urmia, Iran
4 - Department of Computer engineering, Urmia Branch, Islamic Azad University, Urmia, Iran

Received: 2019-05-26 Accepted : 2019-11-03 Published : 2019-12-01

Keywords: deep learning, Medical Image Processing, Computed Tomography, Computer Aided Detection, Lung nodules,

Abstract :

Lung cancer is among the deadliest cancers worldwide. One of the indications of lung cancers is lung nodules which can appear individually or attach to the lung wall. Therefore, the detection of the so-called nodules is complicated. In such cases, the image processing algorithms are performed by the computer, which can aid the radiologists in locating and assessing the nodule's feature. The significant problems with the current systems are the increment of the accuracy, improvement of other criteria in the results, and optimization of the computation costs. The present paper's objective is to efficiently cope with the aforementioned problems by a shallow and light network. Convolutional Neural Networks were utilized to distinguish between benign or malignant lung nodules. In CNN's networks, the complexity increases as the number of layers increases. Accordingly, in the current paper, two scenarios are presented based on State the art and shallow CNN method in order to accurately detect lung nodules in lung CT scans. A subset of the LIDC public dataset including N=7072 CT slices of varying nodule sizes was also used for training and validation of the current approach. Training and validation steps of the network were performed approximately in five hours, and the proposed method achieved a high detection accuracy of 83.6% in Scenario1 and 91.7% in Scenario2. Due to the usage of various validated database images and comparison with previous similar studies in terms of accuracy, the proposed solution achieved a decent trade-off between criteria and saved computation costs. The present work demonstrated that the proposed network was simple and suitable for the so-called problems. Although the paper attempted to meet the existing challenges and fill up the prevailing niches in the literature, there are still further issues that requires complementary studies to shape the tapestry of the knowledge in the field.