Siemens Healthineers uses Anomaly Detector to identify anomalies in X-ray tube production

In late 1895, Wilhelm Conrad Röntgen discovered the existence of “X-rays,” a type of radiation that allows us to take images of internal objects, like bones and organs in the body. Only a year after that, the progenitors of Siemens Healthineers built the company’s first medical X-ray device. Today, more than 120 years later, Siemens Healthineers is a global corporation headquartered in Erlangen, Germany, employing nearly 66,000 people. The company designs and produces medical devices, including X-ray and MRI machines, which it provides to medical institutions across 70 countries. Using these medical devices, Siemens Healthineers’ customers help hundreds of thousands of patients each year.

Medical device production requires absolute precision, so Siemens Healthineers runs its X-ray devices and components through several quality tests, during which the production equipment and external sensors capture performance and condition data. To assess multiple data points collected during these processes at once, Siemens Healthineers uses machine learning models with multivariate anomaly detection capabilities. With this technology, Siemens Healthineers is able to monitor the condition of devices and components during the production process.

To accommodate the increasing amount of production data, as well as streamline data science activities within a common development environment, Siemens Healthineers decided to shift the implementation of the multivariate anomaly detection use case to the cloud. As they searched for the right cloud technology, they learned that Microsoft not only has a feature-rich machine-learning platform, but also provides dedicated machine learning models as a service.

Those services are promising, since they ease the integration of pretrained, customizable machine learning models and serve additional features like explainability. Thus, the overall development time can be reduced. For the anomaly detection use case, Siemens Healthineers successfully employed the newly released Anomaly Detector of Microsoft Azure Cognitive Services in a real-world use case.

A tiny flaw can disrupt patient care

In an X-ray tube, an electron beam is accelerated from a cathode onto an anode. When the electron beam impacts the anode, X-rays and heat are generated through a process called bremsstrahlung—more than 99 percent of the electron beam’s kinetic energy is converted into heat through this process. For some of its X-ray machines, Siemens Healthineers uses a kind of X-ray tube called a “rotating anode tube,” which enables the tube’s anode to uniformly absorb all that heat through rapid rotation of the anode.

“The liquid metal bearings allow for high rotation rates, on the one hand, and high heat dissipation from the anode out of the vacuum tube into the cooling oil on the other,” explains Dr. Jens Fürst, Head of Automation and Digitalization at Siemens Healthineers. “Liquid metal bearings are produced at the limits of what's basically physically possible, so that's why high-quality requirements are installed in production.”

Faulty liquid metal bearings may cause abrupt operational breakdowns, which can lead to reduced quality of patient care, increased wait times, and prolonged diagnosis and treatment.

Accommodating massive amounts of data and easy, quick implementation

“An advantage of collaborating in a cloud environment is the ability to directly share datasets at different stages among developers,” Selmaier says. “This is very practical throughout the development process, as it improves transparency and reproducibility.”

During one of their quality tests, Siemens Healthineers runs the bearing component of the X-ray tubes under operating conditions and uses IoT devices equipped with sensors to collect multivariate time series data. This data is used to draw conclusions about the condition of the bearing, forming the basis for the anomaly detection use case.

“The raw data collected adds up to several terabytes over the course of months,” Selmaier says. “Being able to choose between hot and cold data storage means that even these large amounts of data can be stored and accessed under reasonable economic conditions.”

Siemens Healthineers uses Microsoft Azure IoT Hub to connect the IoT devices to the cloud. The message queuing telemetry transport, or commonly referred to as MQTT, an open protocol for IoT communication, is used for a unidirectional data transfer.

Siemens Healthineers trained Anomaly Detector with good behavior patterns. During inference, the model identifies intervals within the multivariate time series that go against familiar patterns, describes them as anomalies, and brings them to the user’s attention. Anomaly Detector also provides severity scores and a contributor ranking, indicating how much each variable contributed to the corresponding anomaly. Siemens Healthineers uses this to make additional interpretations of the prediction results.

A successful partnership and an exciting future

For Dr. Fürst, the future with Microsoft and Azure is bright. “For the first run through, choosing Anomaly Detector was definitely the right way to go. We didn’t face any substantial limits that will prevent implementation. As Cognitive Services continue to evolve, I'm excited to see which additional functionalities will be offered in the future.”

“Collaborating with experienced professionals in the field of cloud computing and machine learning on a real use case led to an efficient and effective knowledge transfer. Within the short amount of time given, our colleagues on the Siemens Healthineers side got familiar with the required set-up steps of the cloud environment and tools,” says Johannes Distler, Project Manager Business Digitalization at Siemens Healthineers.

Because of this close working partnership with Microsoft, ease of implementation, access to training and tutorials, and having an enterprise-ready multivariate anomaly detector right out of the box, Siemens Healthineers is confident about taking their cloud-based solution to the next level.