A high-performance, hardware-based deep learning system for disease diagnosis.

Area of Science:

• Medical Science
• Computer Engineering
• Artificial Intelligence

Background:

• Deep learning models show human-level performance in medical science but are resource-intensive and not hardware-friendly.
• Existing hardware implementations face challenges due to complex algorithms and activation functions.
• Novel hardware-friendly activation functions offer potential for high throughput and accuracy.

Purpose of the Study:

• To develop a hardware-based neural network for accurate and efficient disease diagnosis.
• To utilize cost-efficient and accurate activation functions for improved hardware implementation.
• To demonstrate the system's capability in cancer detection and its potential for other diseases.

Main Methods:

• Proposed a novel hardware-based neural network architecture.
• Integrated hardware-friendly activation functions, Sqish and LogSQNL.
• Designed a reconfigurable system for versatile classification tasks.

Main Results:

• Achieved 98.23% accuracy in predicting human cancer presence.
• Classifies samples in a single clock cycle (15.75 nanoseconds).
• Demonstrated high performance with low resource utilization (983 slice registers, 2,655 LUTs, 1.1 kilobits memory).
• Processes 63.5 million samples/second and performs 20 giga-operations/second.
• System is 5-16 times cheaper and 4 times faster than existing solutions.

Conclusions:

• The proposed hardware system offers a significant advancement in efficient and accurate medical diagnosis.
• The reconfigurable design allows for broad applicability in classifying various diseases.
• This approach overcomes hardware limitations of deep learning in medical applications.