Big Data For Medical Image Analysis: A Performance Study

Big data systems can be used to facilitate powerful medical image analysis at scale. Understanding their behaviors in this context can lead to many benefits, ranging from superior infrastructure configurations to optimized parallel algorithm implementations. This paper is, to our knowledge, a first step towards developing such an understanding for state-of-the-art big data platforms. We characterize a representative medical image segmentation pipeline, detailing the per-stage CPU, memory, I/O reads and writes, and execution time patterns. This characterization has already helped us overcome a bottleneck persistently causing analysis to crash unexpectedly, and avoid poor architecture choices on storage and parallel execution.