OpenCLIPER: An OpenCL-Based C++ Framework for Overhead-Reduced Medical Image Processing and Reconstruction on...

Area of Science:

• Medical Image Processing
• High-Performance Computing
• Software Engineering

Background:

• Medical image processing algorithms face computational cost limitations.
• Dedicated hardware like Graphics Processing Units (GPUs) offer speedups but require complex management.
• Developers often struggle with housekeeping tasks (device management, data transfer, synchronization) when using accelerators.

Purpose of the Study:

• To present an OpenCL-based framework for seamless integration of dedicated computing devices in medical image processing.
• To enable developers to concentrate on image processing tasks rather than device management.
• To abstract away complexities of device selection, initialization, data handling, and kernel compilation.

Main Methods:

• Developed an OpenCL-based framework automating device discovery, initialization, and data transfers.
• Implemented a system where data structures are defined once, independent of the computing device.
• Utilized pinned memory/buffer mapping for optimized data transfer performance.
• Designed algorithms as chained mathematical operators (processes) with split initialization and core workloads.

Main Results:

• The framework significantly reduces the code required for device management and kernel compilation.
• Developers can utilize computing devices almost immediately and effortlessly.
• Algorithm code becomes independent of the target device type (CPU or GPU).
• Processes can split initialization from core workload, avoiding performance penalties in chains and loops.

Conclusions:

• The proposed framework streamlines the use of dedicated computing devices for medical image processing.
• It enhances developer productivity by minimizing overhead associated with hardware acceleration.
• The approach facilitates efficient and portable medical image analysis on diverse hardware platforms.