A deep stacked random vector functional link network autoencoder for diagnosis of brain abnormalities and breast cancer. (April 2020)

• Record Type: Journal Article
• Authors: Nayak, Deepak Ranjan; Dash, Ratnakar; Majhi, Banshidhar; Pachori, Ram Bilas; Zhang, Yudong
• Is Part Of: Biomedical signal processing and control. Volume 58(2020)
• Publication Date: 2020-04
• Subjects: Magnetic resonance imaging -- Autoencoder -- Random vector functional link network -- Deep neural network -- ReLU; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.101860
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A deep stacked RVFL autoencoder is proposed for diagnosis of brain abnormalities. The proposed SRVFL-AE network is free from additional fine-tuning step. It eliminates the requirement of hand-engineered features. A comparative analysis of SRVFL-AE with its counterparts is performed. Results show the superiority of the proposed method over existing methods. Abstract: Automated diagnosis of two-class brain abnormalities through magnetic resonance imaging (MRI) has progressed significantly in past few years. In contrast, there exists a limited amount of methods proposed to date for multiclass brain abnormalities detection. Such detection has shown its importance in biomedical research and has remained a challenging task. Almost all existing methods are designed using conventional machine learning approaches, however, deep learning methods, due to their advantages over machine learning, have recently achieved great success in various computer vision and medical imaging applications. In this paper, a deep neural network termed as stacked random vector functional link (RVFL) based autoencoder (SRVFL-AE) is proposed to detect the multiclass brain abnormalities. The RVFL autoencoders are the basic building blocks of the proposed SRVFL-AE. The main purpose of choosing RVFL as the core component of the proposed SRVFL-AE is to improve the generalization capability and learning speed compared to traditional autoencoder based deep learning methods. Further, the rectified linear unit (ReLU) activation function is incorporated in the proposed deep network to provide fast and better hidden representation of input features. To evaluate the effectiveness of suggested method, two benchmark multiclass MR brain datasets such as MD-1 and MD-2 are considered. The scheme achieved a greater accuracy of 96.67% and 95.00% on MD-1 and MD-2 datasets respectively. The efficacy of the model is also tested over a standard breast cancer dataset. The results demonstrated that our deep network obtains better performance with least training time and compact network architecture compared to its counterparts.