Accelerating Euclidean Distance Transforms: A Fast and Flexible Approach With Multi-Vendor GPU, Multi-Threading, and Multi-Language Support

Abstract

Euclidean distance transforms are fundamental in image processing and computer vision, with critical applications in medical image analysis and computer graphics. However, existing implementations often lack performance, flexibility, or cross-platform compatibility. This paper introduces a novel approach to accelerating Euclidean distance transforms using hardware-agnostic GPU acceleration, multi-threading, and cross-language support. Our method, implemented in Julia with Python bindings, supports multiple GPU platforms including NVIDIA CUDA, AMD ROCm, Apple Metal, and Intel oneAPI. Benchmarks demonstrate substantial performance improvements, achieving speedups by a factor of 250 for 2D and a factor of 400 for 3D transforms compared to optimized CPU implementations. We showcase the impact of our approach through two real-world applications: accelerating the Hausdorff distance loss function for medical image segmentation, achieving a 7.4-fold improvement in processing speed with enhanced accuracy, and enhancing a GPU-optimized distance transform-based skeletonization algorithm with performance gains up to a factor of 88. Our open-source implementation provides a flexible, high-performance solution for exact Euclidean distance transforms, advancing the state-of-the-art in medical image analysis and computer vision.