Efficient FPGA Implementations for Gradient Boosted Decision Trees with Applications in Real-Time Medical Imaging Systems

Abstract

Implementing artificial intelligence on embedded platforms is a critical challenge, especially for applications requiring low power consumption, high performance, and real-time processing. The choice of algorithms plays a significant role in determining resource allocation, energy efficiency, and computational performance, making it essential to select models that balance these factors effectively. Medical applications are a key area where intelligent embedded systems have the potential to revolutionise diagnosis, surgical assistance, and patient care. This thesis addresses the challenges of deploying efficient artificial intelligence models on field-programmable gate arrays, leveraging their energy efficiency, computational capability, and flexibility advantages. The research contributes by developing, optimising, and applying machine learning models tailored for medical imaging systems. The first contribution involves designing an intelligent image-processing system for surgical operations. This system integrates advanced image enhancement methods with lightweight machine-learning models. It includes optimised techniques for contrast enhancement, convolution operations, and edge detection, establishing the foundation for real-time, efficient image processing. The second contribution focuses on optimising gradient boosted decision tree models for deployment on field-programmable gate arrays. By introducing training methods that consider the impact of reduced precision during model development, this work reduces the computational and memory demands of these models while maintaining high accuracy. These optimised models significantly minimise resource usage and increase processing speeds, making them well-suited for use in energy-constrained, real-time devices. The final contribution is to develop an adaptive hardware design for gradient boosted decision tree models. This system allows for dynamic updates to model parameters without the need for reconfiguring the hardware, enabling it to adapt seamlessly to changing data and requirements. This adaptability is critical for medical applications, where evolving data necessitates continuous model updates while maintaining high performance and reliability. Together, these contributions advance the field of embedded artificial intelligence by providing practical strategies for deploying intelligent systems in resource-limited environments. This work makes a major contribution to the future development of low-power, high-performance medical imaging systems and adaptive artificial intelligence solutions.