An efficient multiplier-based FPGA CNN accelerator for Parkinson's disease detection using hand-drawn circle images

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computer Engineering

Background:

• Parkinson's disease (PD) diagnosis can be aided by analyzing handwriting patterns.
• Field-Programmable Gate Arrays (FPGAs) offer efficient hardware acceleration for machine learning models.
• Optimized multiplier architectures are crucial for enhancing the performance of neural network accelerators.

Purpose of the Study:

• To develop and evaluate an FPGA-based convolutional neural network (CNN) accelerator for Parkinson's disease (PD) handwriting classification.
• To investigate the impact of different multiplier architectures (standard, approximate logarithmic, Karatsuba) on the accuracy and hardware efficiency of the CNN.
• To assess the feasibility of deploying lightweight CNNs with optimized multipliers for resource-constrained edge healthcare applications.

Main Methods:

• A lightweight binary CNN classifier was trained and deployed on FPGA hardware.
• Three multiplier architectures were evaluated: standard, approximate logarithmic, and Karatsuba.
• The model was quantized to Q4.12 fixed-point precision, and hardware inference was validated on the NewHandPD dataset, comparing FPGA outputs with software results using Mean Absolute Deviation (MAD).

Main Results:

• FPGA implementation achieved classification behavior closely aligned with software inference, demonstrating improved hardware efficiency.
• Under non-augmentation validation, FPGA accuracy was 89.73% (vs. 92.86% software); with standard augmentation, FPGA accuracy was 95.40% (vs. 97.83% software) using the Approximate Logarithmic Multiplier.
• The Approximate Logarithmic Multiplier showed promising results in balancing accuracy and efficiency for FPGA deployment.

Conclusions:

• Lightweight CNNs with optimized multipliers are feasible for resource-efficient edge healthcare applications, particularly for preliminary PD handwriting classification.
• The study serves as a proof-of-concept, highlighting the potential of hardware acceleration for medical diagnostics.
• Further validation with larger datasets is recommended for definitive clinical application.