Real-time data analysis for medical diagnosis using FPGA-accelerated neural networks.

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Machine Learning

Background:

• Real-time analysis of patient data during medical procedures is crucial for improving success rates.
• Fast sensors necessitate efficient computational frameworks like Deep Neural Networks (DNNs) for real-time operation.
• Traditional computing platforms (CPUs, GPUs) face overheads from communication, memory access, and generic architectures, hindering real-time performance.

Purpose of the Study:

• To implement a low-latency Multi-Layer Perceptron (MLP) processor using Field Programmable Gate Arrays (FPGAs).
• To reduce data movement delays by enabling direct interfacing of sensors and processing units.
• To demonstrate the effectiveness of application-specific optimizations for real-time diagnostic feedback.

Main Methods:

• Developed a custom MLP processor on an FPGA.
• Integrated direct interfacing capabilities for sensors, storage, display, and actuators.
• Optimized compute pipelines for specific applications to improve resource utilization and reduce latency.
• Utilized mass-spectrometry data sets for real-time cancer detection.

Main Results:

• Application-specific parameter sizing reduced latency by an average of 20%.
• FPGA design demonstrated an average speedup of 144x over CPUs and 21x over GPUs.
• Showcased that excessive computing resources can decrease performance in latency-constrained applications.

Conclusions:

• FPGA-based MLP processors are vital for minimizing latency and maximizing resource utilization in real-time inference.
• Direct interfacing and processing of sensor data with FPGAs enable critical diagnostic feedback during medical procedures.
• This approach enhances the potential for successful patient outcomes through timely and accurate analysis.