Remaining Useful Life Prediction of Imaging Devices via Multi-Iteration SemiSupervised Modeling and Real-World Evaluation

Obtaining a reliable and substantial amount of ground truth from medical imaging devices is challenging due to several factors: simultaneous degradation of multiple components, reversible and irreversible failure modes, operators' sensitivity to degradations, external stressors, varying operating environments, and identical symptoms arising from different root causes. This paper proposes an iterative semi-supervised approach for training a machine learning (ML) model to predict the remaining useful life (RUL) of imaging devices. The method employs an iterative approach to model degradation modes in complex systems. In the initial iteration, an ML model is trained on all available ground truth data for a specific system, subsystem, or component (SSC) failure, with the objective of maintaining the false positive rate (FPR) below a predefined threshold. This step learns signatures of the most dominant degradation mode (a subset of the full dataset). In the subsequent iteration, the predicted ground truth labels from the first model, corresponding to the dominant degradation mode, are used to train a new model. Additional iterations are performed if the FPR for this mode exceeds the threshold. The FPR threshold is an optimization parameter, varying across modes and determined by evaluating model performance on validation data. Additionally, we introduce a novel evaluation method that simulates real-world deployment. Specifically, we evaluate the model by its initial predictions for each failure, focusing on whether it can detect failure signatures within a set time window after those predictions. This approach captures realistic failure scenarios, such as, operators' sensitivity to degradations, reversible failures, and environmental variations. Results on a component failure in a computed tomography (CT) machine demonstrate that the proposed multi-iteration semi-supervised modeling and evaluation approach significantly improves performance metrics (achieving over 80 % recall and 90 % precision) compared to traditional ML modeling and evaluation methods.

Keywords: Predictive Maintenance, Semi-Supervised Learning, Medical Imaging Devices, Remaining Useful Life, Real-World Evaluation.