A Novel FPGA-Accelerated Real-Time Approach for Early Detection of Alzheimer’s Disease Using Retinal OCT Images.

 

Background: Early AD screening is essential for timely interventions. Current gold-standard methods — MRI, PET — involve high costs, long scan times, and complex 3D processing.

Retinal OCT: Provides high-resolution 2D cross-sectional imaging during routine eye exams. RNFL thinning, GC-IPL loss, and microvascular alterations are emerging biomarkers for neurodegeneration — a patientfriendly point-of-care alternative.

Critical Gap: No embedded, real-time hardware-accelerated OCT system for AD risk stratification exists. FPGA offers deterministic, low-power, millisecond inference via fixed-point arithmetic.

Objectives
1. Design compact CNN hardware accelerator on FPGA for real-time OCT inference
2. INT8-quantized PL convolution + Transformer classifier on ARM PS (PYNQ-Z2)
3. Millisecond-latency, low-power (<2 W) inference for embedded OCT devices
4. Demonstrate PS+PL hybrid inference pipeline feasibility

Platform: PYNQ-Z2 (Zynq XC7Z020)
▸ Dual ARM Cortex-A9 @ 667 MHz (PS)
▸ 28 nm Artix-7 FPGA fabric — 85K LUTs, 220 DSP48E1s
▸ AXI4-Stream DMA for PS PL data transfer
▸ Vivado 2022.x synthesis & implementation