JISE


  [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23]


Journal of Information Science and Engineering, Vol. 42 No. 5, pp. 1435-1459


GAN-Augmented Lightweight CNN-GRU Hardware IP Core for Neurological Disorder Classification

I. NAURIN BAHADUR+ AND LAKSHMI BOPPANA
Department of Electronics and Communication
National Institute of Technology Warangal
Warangal, Telangana, 506004 India
E-mail: in712107@student.nitw.ac.in+; lakshmi@nitw.ac.in

Medical practitioners require a portable/wearable device which consumes less power to detect neurological disorders with high accuracy. The work presented in this paper proposes a lightweight hybrid CNN-GRU (Gated Recurrent Unit) architecture for the diagnosis of neurological diseases (ADHD, Alzheimer’s and Epilepsy) using multiclass classification. Since EEG data is the time-series data, the proposed model is developed by adding an GRU layer in the convolution Neural Network (CNN). The hardware complexity of the proposed model in terms of number of parameters is reduced by preprocessing the data using the Discrete Wavelet Transform (DWT) which reduces the number of kernels required in the convolution layer. The proposed work also addresses the problem of less accuracy in diagnosing the neurological disorders due to the small dataset by employing Generative Adversarial Network (GAN) for data augmentation. The GAN created in this study attains a discriminator accuracy of 53%, a discriminator loss of 0.62, a generator loss of 0.53, and a mean square error (MSE) of 100, illustrating its capacity to produce realistic synthetic data. Fine-tuning is used to improve the model accuracy. The simulation results indicate that the suggested CNN-GRU model attains a high classification performance, with an average accuracy of 98.5% and an Area Under the Curve (AUC) of 0.98. The CNN-GRU model, trained using both real & GAN-generated EEG data and implemented on an Artix-7 FPGA to classify EEG signals associated with ADHD, Alzheimer's and epilepsy disorders. The verilog model developed for the architecture of the proposed CNN-GRU model is synthesized using 32nm technology cells after functional verification using Synopsys Design Compiler. The post layout simulations of the proposed architecture are performed and the parasitics are extracted using Synopsys IC compiler. The proposed architecture exhibits a latency of 476.7μs for the multi-class classification to detect neurological disorders. The proposed architecture of the CNN-GRU model consumes 1.119mW of power with 0.9V operating voltage source and occupies an area of 2.40mm2.

Keywords: ASIC, CNN-GRU, DW, EEG, FPGA, Generative Adversarial Network (GAN), neurological disorders

  Retrieve PDF document (JISE_202605_07.pdf)