A Configurable and Fully Synthesizable RTL-Based Convolutional Neural Network for Biosensor Applications.

Area of Science:

• Hardware design for biosensors
• Digital signal processing for biomedical applications
• Machine learning in healthcare

Background:

• Biosensor technology is crucial for disease detection through DNA identification.
• Developing efficient hardware architectures for biosensor data analysis is a significant challenge.
• Current methods require optimized computational approaches for real-time disease detection.

Purpose of the Study:

• To propose a synthesizable register-transistor level (RTL) based convolutional neural network (CNN) architecture for biosensor applications.
• To optimize hardware overhead and reduce computational complexity for DNA identification.
• To achieve high accuracy and efficiency in disease detection using biosensors.

Main Methods:

• Implemented a parallel computation of multiplication and accumulation (MAC) approach for optimized hardware.
• Utilized multiplier bank sharing to reduce the implementation area during convolutional and fully connected operations.
• Trained the CNN model in MATLAB® on the MNIST® handwritten dataset and validated using ModelSim®.

Main Results:

• Achieved 92% accuracy in disease detection through DNA identification using the proposed RTL-based CNN.
• The design significantly reduces arithmetic calculations and hardware overhead.
• Demonstrated efficient real-time processing with a total execution time of 8.6538 ms.

Conclusions:

• The proposed RTL-based CNN offers a hardware-efficient solution for biosensor applications.
• The architecture enables accurate and rapid DNA identification for disease detection.
• The design is suitable for implementation in modern CMOS technology, occupying 9.986 mm² with a power requirement of 2.93 W.