2018/12/16 by Valérie Hayot‐Sasson, Hayot-Sasson, Valérie, Shawn T. Brown +3
Medicine · Neuroscience · #Advanced MRI Techniques and Applications #Brain Tumor Detection and Classification #Distributed #FOS: Computer and information sciences #Functional Brain Connectivity Studies #Parallel #and Cluster Computing (cs.DC)
paper · pdf · doi:10.48550/arxiv.1812.06492
openalex publication_date 2018/12/16 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Neuroimaging datasets are rapidly growing in size as a result of advancements\nin image acquisition methods, open-science and data sharing. However, the\nadoption of Big Data processing strategies by neuroimaging processing engines\nremains limited. Here, we evaluate three Big Data processing strategies\n(in-memory computing, data locality and lazy evaluation) on typical\nneuroimaging use cases, represented by the BigBrain dataset. We contrast these\nvarious strategies using Apache Spark and Nipype as our representative Big Data\nand neuroimaging processing engines, on Dell EMC's Top-500 cluster. Big Data\nthresholds were modelled by comparing the data-write rate of the application to\nthe filesystem bandwidth and number of concurrent processes. This model\nacknowledges the fact that page caching provided by the Linux kernel is\ncritical to the performance of Big Data applications. Results show that\nin-memory computing alone speeds-up executions by a factor of up to 1.6,\nwhereas when combined with data locality, this factor reaches 5.3. Lazy\nevaluation strategies were found to increase the likelihood of cache hits,\nfurther improving processing time. Such important speed-up values are likely to\nbe observed on typical image processing operations performed on images of size\nlarger than 75GB. A ballpark speculation from our model showed that in-memory\ncomputing alone will not speed-up current functional MRI analyses unless\ncoupled with data locality and processing around 280 subjects concurrently.\nFurthermore, we observe that emulating in-memory computing using in-memory file\nsystems (tmpfs) does not reach the performance of an in-memory engine,\npresumably due to swapping to disk and the lack of data cleanup. We conclude\nthat Big Data processing strategies are worth developing for neuroimaging\napplications.\n