How HPE and WEKA enhance healthcare through medical imaging AI

[TechExperts]

Learn how HPE and WEKA enable radiologists to use increasingly huge amounts of medical imaging data to train AI models that better assist clinicians in diagnostic testing.

By Robert Murphy, Director of Product Marketing , WEKA

A tsunami of medical imaging procedures – all with higher image counts and resolutions to analyze – are drowning the limited number of radiologists available to interpret them. This makes their chosen healthcare profession increasingly difficult.

Too much data, not enough radiologists. It's AI to the rescue!

This abundance of medical imaging data can be put to work to train today’s efficient Convolutional Deep Neural Network models[1] running on the latest NVIDIA GPU processors to assist clinicians in their diagnostic tasks. And none too soon, because all that medical image data must be read by increasingly overwhelmed diagnostic clinicians.

The data pipeline has now become the bottleneck in AI

For the last decade, most of the focus in artificial intelligence (AI) has been on GPU processing, and rightfully so, with all the advancements going on there. But GPUs have gotten so fast that data input into them has become the primary bottleneck to overall AI training performance.

Recent research by Google[2], Microsoft[3] and organizations around the world[4] are uncovering that GPUs spend up to 70% of their AI training time waiting for data. Looking at their data pipelines, that is no surprise. In Figure 1 below, you’ll see the typical deep learning pipeline described by NVIDIA[5] and widely used by organizations doing A

Figure 1. Typical multi-copy deep learning data pipeline

As shown here, at the beginning of each training epoch the millions of training medical images kept on high-capacity object storage are typically copied to a faster clustered storage system tier, and then copied again to GPU local storage that is used as scratch space for GPU calculations. Each copy introduces copy time latency and management intervention, slowing each training epoch considerably. Valuable GPU processing resources are kept idle waiting for data, and vital training time is needlessly extended.

HPE and WEKA have a better way

HPE Solutions for WEKA collapse the typical GPU-starving multi-hop AI data pipeline into a single “no-copy” high-performance, high-capacity data platform for AI. High-capacity object storage is “fused” with high-speed WEKA storage, sharing the same namespace, and accessed directly by the GPU with the NVIDIA GPUDirect Storage protocol, removing all bottlenecks as shown in Figure 2. below, with the physical implementation shown in Figure 3.

Figure 2. The HPE Weka GPUDirect Storage “no copy” data pipeline

 Figure 3. HPE Solutions for WEKA - AI innovation in medical imaging reference architecture

The data platform for AI

Not only has HPE WEKA removed all the performance-sapping hops from the typical AI data pipeline but Weka has also been purpose-built to handle the most challenging AI data IO patterns. This results in streamlined data movement between storage and GPU that vexes other storage systems: big files, small files, low latency read and writes. All these IO patterns are all deftly and uniquely handled by WEKA running on HPE ProLiant servers, removing any AI performance bottlenecks.

To learn more, check out this brief: HPE Solutions for WEKA – Data Storage for Artificial Intelligence in Medical Imaging. And contact your HPE representative for more information on this exciting advancement in medical imaging.

---

Meet our Insights Experts guest blogger, Robert Murphy, Director of Product Marketing at WEKA

Previously, Robert was the Big Data Program Manager for General Atomics Energy and Advanced Concepts and was responsible for High Performance Data and Analytics in IBM's Software Defined Environments organization. Before IBM, Bob held positions with increasing levels of responsibility at Hewlett Packard, Silicon Graphics, Oracle, and Sun Microsystems. He has a Biomedical Engineering Degree from Purdue University.

Insights Experts
Hewlett Packard Enterprise

twitter.com/HPE_AI
linkedin.com/showcase/hpe-ai/
hpe.com/us/en/solutions/artificial-intelligence.html

[1] An overview of deep learning in medical imaging

[2] “tf.data: A Machine Learning Data Processing Framework

[3] Analyzing and Mitigating Data Stalls in DNN Training

[4] Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters

[5] Beyond the Hype: Is There a Typical AI/ML Storage Workload?; CJ Newburn