Medical imaging or diagnostic equipment such as Computed Tomography (CT), Ultrasound, and Magnetic Resonance Imaging (MRI) devices play a critical role in modern healthcare. But while these devices enable healthcare providers to better diagnose their patients' and provide an optimal treatment plan, they are also very expensive to maintain. As these devices have become increasingly connected, opportunities to the role of machine learning (ML) for medical equipment maintenance has become important to keeping them up and running in order to provide the best patient care and maximize revenues.
Cost and Complexity
Medical devices are expensive due to their complexity. They are engineered by combining many different components that are prone to failure and will need to be replaced multiple times during the lifetime of the device. Some parts fail due to wear and tear from normal usage, however, parts also often fail due to abnormal usage or varying environmental conditions. For example, a part may overheat and breakdown due to high usage or a part may break down because the cooling system is not functioning correctly.
Identifying Sensor Readings
To ensure optimal operating conditions medical imaging devices rely on multiple sensors. For example, a CT scanner has sensors for monitoring tube temperature, water temperature, fan speed, air temperature, water flow, and other important variables. Each sensor periodically records its readings, which can be used to determine whether the tracked variable is in the normal range. However, it is difficult to accurately identify which sensor readings are in the normal range and which are not. Currently, these thresholds are manually defined often leading to misclassified normal values as abnormal and vice-versa. Missing an abnormal sensor reading means missing an opportunity to take corrective action in time. As a result, a problem may go unnoticed resulting in the failure and replacement of an expensive part.
ML-based Techniques for Anomaly Detection
Machine learning provides a better technique, known as anomaly detection (AD), for identifying abnormal sensor readings accurately. ML-based AD techniques use historical data to train a model that can be used for detecting anomalous sensor values. Multiple ML-based techniques are available for anomaly detection, but the right technique depends on the characteristics or distribution of data. If the data or sensor readings have a normal distribution, a probability density function can be used for identifying outliers or abnormal values.
A probability density function consists of two parameters, mean and standard deviation of data. These are calculated from historical values of a sensor reading. For example, if we wanted to detect abnormal water temperature readings from a CT Scanner, we calculate the mean and standard deviation of historical water temperature readings. These then become the model parameters for the ML model that is used for detecting anomalous water temperature logged by a CT scanner. The ML model, in this case, is nothing but a function that calculates the probability of a sensor reading given its historical mean and standard deviation.
The same technique can also be used for detecting an abnormal combination of readings from two or more different sensors, however, detecting abnormal combinations of readings from different sensors is much more difficult than from a single sensor. The reading from each sensor by itself may not be abnormal, but the combination of these readings may be abnormal. For example, assume we are monitoring four different environmental factors, water inlet temperature (wit), water outlet temperature (wot), air inlet temperature (ait) and air outlet temperature (aot). An ML-based anomaly detection model, in this case, will calculate the probability for the set (wit, wot, ait, aot) readings. Assume that the normal range for wit, wot, ait, and aot is 10 to 90. If at some point in time, wit is 15, wot is 85, ait is 18, and aot is 86, each of these individually seems normal, but the probability of the combination (15, 85, 18, 86) may be 0.01% and thus represent an abnormal combination.
The Power of ML Based Anomaly Detection
Machine learning makes it easy to detect anomalies in sensor readings recorded by a device. With the help of an ML-based anomaly detection solution, a medical imaging device operator can easily spot problems and take corrective actions. They can fix issues before those become bigger issues and cause expensive parts to fail, rendering a system unusable. By predicting part failures, healthcare organizations can prevent unplanned downtime and expensive repairs, thus maximizing the revenue and providing better patient care.