A Re-configurable Software-Hardware CNN Framework for Automatic Detection of Respiratory Symptoms
MetadataShow full item record
Type of Work20 pages
Citation of Original PublicationRashid, HA., Ren, H., Mazumder, A.N., Sajadi, M.M., Mohsenin, T. (2022). A Re-configurable Software-Hardware CNN Framework for Automatic Detection of Respiratory Symptoms. In: Adibi, S., Rajabifard, A., Shariful Islam, S.M., Ahmadvand, A. (eds) The Science behind the COVID Pandemic and Healthcare Technology Solutions. Springer Series on Bio- and Neurosystems, vol 15. Springer, Cham. https://doi.org/10.1007/978-3-031-10031-4_4
RightsThis item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.
Access to this file will begin on 10-30-2024
Detection of respiratory symptoms has long been an area of extensive research to expedite the process of machine aided diagnosis for various respiratory conditions. This chapter attempts to address the early diagnosis of respiratory conditions using low power scalable software and hardware involving end-to-end convolutional neural networks (CNNs). We propose RespiratorNet, a scalable multimodal CNN software hardware architecture that can take audio recordings, speech information, and other sensor modalities belonging to patient demographic or symptom information as input to classify different respiratory symptoms. We analyze four different publicly available datasets and use them as case studies as part of our experiment to classify respiratory symptoms. With regards to fitting the network architecture to the hardware framework, we perform windowing, low bit-width quantization, and hyperparameter optimization on the software side. As per our analysis, detection accuracy goes up by 5% when patient demographic information is included in the network architecture. The hardware prototype is designed using Verilog HDL on Xilinx Artix-7 100t FPGA with hardware scalability extending to accommodate different numbers of processing engines for parallel processing. The proposed hardware implementation has a low power consumption of only 245 mW and achieves an energy efficiency of 7.3 GOPS/W which is 4.3× better than the state-of the-art accelerator implementations. In addition, RespiratorNet TensorFlow model is implemented on NVIDIA Jetson TX2 SoC (CPU + GPU) and compared to TX2 single-core CPU and GPU implementations to provide scalability in terms of off-the-shelf platform implementations.