4.6 Article

Lung cancer detection and classification with DGMM-RBCNN technique

Journal

NEURAL COMPUTING & APPLICATIONS
Volume 33, Issue 22, Pages 15601-15617

Publisher

SPRINGER LONDON LTD
DOI: 10.1007/s00521-021-06182-5

Keywords

Lung cancer; Gaussian noise; Region growing; Deep Gaussian mixture model; Gaussian distributions; Dimensionality reduction; ROC curves; Nonlinear model; Martins' correlation coefficient

Ask authors/readers for more resources

This paper introduces a method for lung cancer diagnosis using a combination of deep Gaussian mixture model and region-based convolutional neural network. By preprocessing, segmentation, and feature extraction of images, the diagnostic accuracy can be significantly improved, outperforming traditional methods.
Lung cancer holds a vital spot amidst the causes of cancer deaths among humans. The best way to maximize the survival rate of the patient with cancer is the prior prediction of cancerous cells. But various traditional methods fail to diagnose cancerous cells in their earlier stage. The traditional methods also lead to reduced accuracy. This paper is a mere attempt to demonstrate the methods for diagnosing the disease earlier and also to enhance the accuracy rate. So, at first, the LIDC dataset is considered to be used in this investigation to deal with the higher volume of the scanned image, to attain maximum accuracy. Second, Gaussian noise is excised using Gaussian and Wiener filter. Here pre-processing is carried out to reduce noise in the images. Third, region growing segmentation is used to achieve accurate segmentation of ROI (Region of Interest). In a region growing segmentation, seed points are selected and the adjacent pixels are merged to attain a larger region. After this, features that are extremely significant for a nodule of interest such as area, perimeter, entropy, intensity, or statistical-based features are extracted. From these extracted features, dimensionality is reduced using deep Gaussian mixture model in region-based convolutional neural network [DGMM-RBCNN]. The proposed model is a network with multiple layer latent variables. Here, at every layer, the variables follow Gaussian distributions. Therefore, DGMM forms a cluster of Gaussian distributions to offer a nonlinear model and to describe the image data more flexibly. To eliminate over-parameterized solutions, Gaussian-based dimensionality reduction by designing an overfitting model is used. This is applied at every layer of architecture thereby giving the outcome in deep mixtures of factor analyses. Here, accuracy, sensitivity, specificity, F-measure, ROC curves, and Martins' correlation coefficient are considered as performance metrics. The simulation was carried out in a MATLAB environment, to achieve an accuracy of about 87.79% during the 18thepoch for training and testing the image samples. The false-positive rate could also be determined through this investigation. The anticipated DGMM-RBCNN shows a better and best trade-off than the prevailing systems.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.6
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available