Faster CPU can accelerate ability to use AI on medical images

Tests suggest that computer processors can accelerate the speed with which artificial intelligence can be applied to image studies.

That’s good news for the future of using deep learning to derive insights and find anomalies on medical images. The use of AI in imaging was being constrained by limitations in the memory constraints of graphics processing units (GPUs).

Recent testing by Philips and Intel demonstrated the value of high-power central processing units in applying deep learning to imaging studies. The companies used Intel-brand scalable processors and a Philips AI toolkit on two deep learning inference models. One involved X-rays of bones for bone-age prediction modeling, while the other was on CT scans of lungs for lung segmentation.

The tests showed speed improvements of 38 to 188 times for the studies over baseline measurements. AI techniques such as object detection and segmentation can help radiologists identify issues faster and more accurately, which can translate to better prioritization of cases, better outcomes for more patients and reduced costs for hospitals, the companies’ researchers contend.

Also See: Google’s DeepMind to create AI product to spot eye disease

The companies say the results bode well for the future use of AI in interpretation of radiological studies. While GPUs are crucial for their ability to work well with images, they have inherent memory constraints that data scientists have had to work around in building some AI models.



CPUs, on the other hand, aren’t constrained by the memory limits of GPUs and thus can accelerate complex, hybrid workloads, including memory intenstive models typically found in medical imaging. The companies say the two studies suggest that CPUs processors can better meet the needs of data scientists than GPU-based systems.

Philips contends that it might be able to offer less expensive AI solutions that could run on CPU-based systems.

For healthcare organizations, faster processing times will become more critical as medical image file sizes grow to one gigabyte and beyond, due to improved medical image resolutions, the companies say. More healthcare organizations are using deep learning inference to more quickly and accurately review patient images, and they want to do so without buying expensive new infrastructure.

CPU-based technology can enable organizations “to use their existing hardware to its maximum potential, while still aiming to achieve quality output resolution at exceptional speeds,” says Vijayananda Jagannatha, chief architect and fellow of data science and AI at Philips HealthSuite Insights.