How AI-enabled embedded modules are advancing medtech

For decades, medical device manufacturers have been at the forefront of innovation, utilising artificial intelligence (AI) since the 1990s to significantly enhance medical imaging and data analysis. The evolution of high-performance hardware AI accelerators and the advent of machine learning (ML) for training AI algorithms in the 2000s have further refined image analysis, leading to more accurate and efficient interpretations and, consequently, improved patient outcomes.

Today, a diverse range of medical devices benefit from AI integration. This includes high-performance stationary medical imaging equipment such as MRI and CT scanners, as well as more compact and mobile ultrasound and X-ray devices. Patient monitoring systems and surgical robots are also harnessing AI. Furthermore, AI is enhancing applications in laboratory settings, powering various blood analysers and genome sequencers.

The development, application and significance of AI in medical technology have gained substantial momentum over the past decade. This progress is driven not only by continuous advances in computer technology but also by optimisations in AI algorithms. It is a collective effort of medical professionals, technology developers and stakeholders in the healthcare industry that is driving this significant change.

Recent advancements in AI

A recent example of the importance of AI advancements for medical imaging is the innovative AI algorithm for MRI. This algorithm has revolutionised the scanning process, enabling scans to be completed in less than a minute, thereby reducing the time patients spend in the scanner. The key is the AI-based computing of upscaling or super-scaling. This process requires fewer images than traditional methods, meaning it is faster. The pre-trained AI interpolates a small number of individual images into a high-resolution overall image. Moreover, the AI can independently and accurately sharpen blurry areas within images, demonstrating the power of AI in improving medical imaging.

AI is also being integrated into endoscopy devices, for instance, to alert doctors to lesions during examination in real time and directing their attention to specific areas of concern. High-performance inference is crucial to ensure that the trained models execute swiftly to happen instantly. AI-based endoscopy devices provide doctors with a powerful tool to achieve more accurate clinical results and deliver better patient care.

Beyond diagnostic devices, AI plays a crucial role in therapeutic equipment such as ventilators. In these systems, AI algorithms are trained to continuously analyse critical patient data and automatically adjust parameters like respiratory rate, tidal volume and oxygen supply.

AI accelerators integrated

Dealing with massive volumes of data requires computing architectures capable of handling the data load. To achieve the necessary performance, engineers used either a dedicated general-purpose graphics processing unit (GPGPU), often integrated via the classic PCIe slot and relatively large and power-hungry, or a smaller AI accelerator card designed for an M.2 slot.

Today, more and more processor manufacturers are adapting their chip portfolios to meet the demands of AI. By integrating AI functions directly into their processors, many medical applications can now be realised easily, quickly and at a lower total cost of ownership (TCO), eliminating the need for additional accelerator or dedicated graphics cards.

The first generation of Intel Core Ultra processors (Figure 1) exemplifies this trend. These processors combine a CPU, a particularly powerful GPU and — for the first time — a neural processing unit (NPU) on a single chip. AI applications with high computing power demands can leverage the combined power of all three processors. In contrast, the CPU can be used to handle quick, lightweight tasks. The GPU that can also be used as a GPGPU for computational tasks is best for large workloads requiring parallel throughput, and sustained, heavily-used AI workloads requiring high performance per watt can be optimised to run on the NPU.

Figure 1: The Intel Core Ultra processors integrate a CPU and GPU as well as a dedicated AI engine — the NPU — for energy-efficient AI calculations. Image credit: Intel.

The Intel NPU executes ML algorithms and AI inferencing with approximately 20x greater energy efficiency than standard x86 instruction set architectures. For image classification tasks, applications can utilise the graphics unit as a GPGPU, achieving performance levels comparable to discrete GPUs. This results in 1.9x faster graphics or GPGPU processing, enabling a more detailed, meaningful and immersive user experience. These AI features can be easily implemented using a standardised computer-on-module (COM), particularly COM Express, without requiring developers to modify existing designs.

Modular compute provides high flexibility

As AI and its applications evolve, the flexibility of COM and carrier board solutions allows developers to adapt their products to new computing requirements with minimal integration effort and software modifications. They just need to follow two simple steps: unplug the old module and plug in the new one.

One such COM suitable for demanding edge AI workloads is the conga-TC700 (Figure 2). This COM Express Type 6 Compact module, powered by Intel Core Ultra processors, integrates all the necessary AI functions for the applications previously discussed.

The conga-TC700 provides application-ready AI capabilities in a plug-and-play COM Express form factor. Its 10-year availability and the ease of upgrading applications enable powerful real-time computing and offer high-performance AI functions for various medical applications, including surgical robots, medical imaging systems and high-resolution diagnostic workstations.

Figure 2: The conga-TC700 is suitable for real-time compute and AI applications requiring high reliability and fanless operation.

Developing and optimising AI models

Beyond the new edge AI capabilities of the Intel Core Ultra platform, Intel also offers the Intel Geti software framework. This comprehensive computer vision AI platform enables medical device engineers to rapidly develop AI models with limited coding resources. Developers benefit from a unified ecosystem spanning ML in the cloud to AI-accelerated edge devices.

The congatec COM ecosystem is further enhanced by the Intel open-source software AI toolkit, OpenVINO. This tool allows for the optimisation and transfer of pre-developed, often hardware-specific, AI models to the medical device manufacturer’s platform, regardless of where they were created. OpenVINO can also manage workload distribution, intelligently deciding which tasks should be handled by the CPU, GPU or NPU for maximum efficiency.

Conclusion

AI has been a longstanding focus in medical technology, predating its adoption in other industrial markets. AI is even being touted as the new operating system for medical devices. Recent advancements in semiconductor technology have yielded microprocessors with exceptionally high compute power and graphics performance. Featuring integrated NPU units, they enable faster, more accurate diagnoses while consuming less energy than predecessors. When implemented through COMs, today’s AI-supported medical devices become highly futureproof, making it easy to integrate upcoming technologies by simply swapping the module.

*Maximilian Gerstl is a Product Line Manager and AI expert at congatec.

^Zeljko Loncaric is Market Segment Manager for Medical and Infrastructure at congatec.

Top image caption: Portable ultrasound devices use AI to enhance imaging quality. Image credit: iStock.com/didesign021