Embedded Systems for Healthcare and Industrial Applications

The development of AI for embedded/edge computing systems presents unique challenges and requirements that differ from those of traditional server or desktop systems. While servers and desktops can benefit from their powerful hardware, embedded systems must address strict resource constraints, energy efficiency requirements, and specific real-time requirements. 
The successful implementation of AI on embedded/edge devices therefore requires specialized approaches, optimization techniques, and AI solutions tailored to the specific needs of these systems.

The design of compact AI models that run with high performance on embedded hardware and, depending on the embedded system, also manage with low hardware resources, requires a great deal of experience. This is one of Chimaera's core competencies. We have been developing individualized AI networks consisting of layers since 2012, since the release of the first CNNs such as AlexNet. As a result, we have gained a lot of experience in dealing with various AI technologies and are therefore always up to date with the latest AI architectures.

To achieve optimal performance for an embedded system like the NVIDIA Jetson family, trained AI models must be compressed from 32-bit single precision to either integer or 16-bit floating point (float16). This significantly reduces memory consumption, allowing AI models to calculate more features on smaller hardware and can even run faster in some cases. 
Our development team is the leading expert in the field of numerics, with extensive experience converting already trained AI models to other bit representations, especially for embedded systems. This allows us to use advanced AI models on very small embedded systems, making them available at an attractive price.

That might interest you:

• AI for Healthcare with Jetson Nano 2GB
  https://developer.nvidia.com/embedded/community/jetson-projects/ai_for_healthcare
   
• “There are algorithms that even allow to apply post-training quantization and reduce bit depth and thus save valuable memory resources while maintaining accuracy. This raises new applications on embedded systems like NVIDIA Jetson Family or low-end workstation GPU cards.”
  https://github.com/ist-daslab/gptq

Embedded computing systems, also known as edge computing, are distinguished by their robust hardware at competitive prices. Additionally, they can be integrated directly into a device as an independent system with low energy consumption. 
Depending on the respective application, these can be very small embedded systems that can even fly with drones, such as the Jetson Nano. Alternatively, embedded systems with higher performance are also feasible, such as the NVIDIA Jetson Orin, which enables the processing of very large 3D image or video data. As it is a stand-alone system, it is well-suited for risk management of medical devices.

We approach every task with enthusiasm, whether it is in the field of medical technology or industrial image processing. Our team possesses extensive know-how in the field of medical technology, enabling us to develop AI projects in accordance with MDR regulations from the outset. We therefore offer the development of “AI software as a medical device”. 
With over 18 years of experience developing solutions for radiology, we have a strong background in 3D image processing. 
The development of AI models for 3D images, such as CBCT, CT, or MR, is a challenging task that requires significant experience and expertise. The results are of a higher quality overall, as 3D features capture the image content more consistently than a layer-by-layer operation. 3D data is often available as a stack of 2D section data. If the data is only processed as 2D images, the complete 3D context and therefore important information is lost. Our experience has shown that the recognition rates of real 3D AI models are therefore higher. This is our specialty.

Well-known models such as nnUnet offer an easy way to train a model for research applications. While such models streamline the optimization of hyperparameters to enhance usability during development, they demonstrate suboptimal runtime performance when deployed. These applications are more complex frameworks, making them an ideal starting point for developing a prototype. In the case of extensive 3D applications, however, the time required for inference can be several minutes. 

It is essential to monitor the requirements for the final system at all times, as this is where the actual development begins at Chimaera. 
We offer tailor-made 3D AI models optimized for the respective task and hardware, offering high accuracy and very short inference runtimes up to real-time applications. These are precisely the characteristics required for the development of embedded solutions.