Inferring data from medical imaging with graphic processing units may soon be less frequent with the introduction of central processing units, due to fewer constraints on their memories.

Royal Philips and Intel put this idea to the test recently, pairing Intel Xeon Scalable processors with Philips’ OpenVINO toolkit to evaluate the efficiency of CPUs when applied to use cases in deep learning inference models, concluding that such a partnership meets objectives faster and provides consumers more affordable access to AI solutions.

“CPUs are already prevalent in embedded devices, edge servers, data centers and cloud,” Prashant Shah, director of engineering at Intel Health and Life Sciences, told HCB News. “By enabling deep learning workloads to run efficiently on CPUs, our customers don't need to add expensive accelerator cards to their solutions.”
The two tested their approach in a use case for predicting bone age based on X-ray bone scans and on another for lung segmentation based on lung CT scans.

With both solutions combined, detection speed for bone age prediction was 188 times faster, rising from an initial baseline test result of 1.42 images per second to a final rate of 267.1 images per second after optimizations.

Speed for the lung segmentation model was 38 times greater, surpassing the expected target of 15 images per second and improving from its baseline of 1.9 images per second to 71.7 images per second after optimizations.

Though primarily used to increase the efficiency of deep learning applications in medical imaging, GPU technology is hindered by memory constraints, which data scientists must work around when constructing models. In contrast, CPU technology exhibits fewer constraints and is capable of accelerating complex, hybrid workloads, including larger, memory-intensive models typically found in medical imaging.

In addition, CPUs serve as an efficient tool for deep learning inference applications which typically process workloads in small amounts or in a streaming manner.

The tests demonstrate that such workloads can be implemented through the use of AI applications and without the need for expensive hardware investments.

The Intel Xeon Scalable processor acts as an affordable, flexible platform for AI models while the OpenVINO toolkit helps to deploy pre-trained models for efficiency without sacrificing accuracy.

"Intel’s solution has the potential to enable our customers to use their existing hardware to its full potential, while achieving quality output resolution at exceptional speeds. If clinicians and data scientists want to build their own AI models, our HealthSuite Insights platform gives them access to advanced analytic capabilities to curate and analyze the healthcare data gathered in their own institution," John Huffman, chief scientific officer of data science and artificial intelligence at Philips, told HCB News. "These developments will support clinicians to continue to improve patient care."

Shah declined to disclose any other specific use cases subject to testing, only specifying that “Intel and Philips will continue to work together to further the use of Intel Xeon Scalable processors for deep learning uses.”

CT Homepage