MRI Condition Monitoring with Explainable AI and Feature Selection
Emmanouil Christoforou, Krelis Blom, Qi Gao, Mesrur Boru, Tanju Cataltepe · 2022 30th Signal Processing and Communications Applications Conference (SIU) · 2022
Monitoring multisensor MRI devices using anomaly detection for multivariate time series consists a challenging task. In this use case, we investigate and provide explanations for abnormalities detected in sensors during failed scans and propose a framework for MRI sensor condition monitoring using XAI and feature selection. Sensor properties are preprocessed (normalized, aggregated and resampled) and used to generate statistical features that are fed to machine learning models for Anomaly Detection (AD), namely Isolation Forest. A feature selection step using an XGBoost classifier is applied before the AD models to improve performance by removing unrelated sensors properties. Explanations for the total output of the models and selected instances are provided using SHAP (SHapley Additive exPlanations). Results are presented using plots with anomaly scores against scan statuses and SHAP explanations for selected anomalous instances along with the top highly contributing sensor properties. Our models were able to successfully detect all failed scans with unknown causes as anomalies, achieving a total accuracy up to 95%. Models that successfully detect anomalies during such failed scans, could provide a tool for a comprehensive condition monitoring of the device using Explainable Artificial Intelligence (XAI) for the domain experts.