Author: Giovanni Iacca, Gloria Miori, Laura Orsingher, Daniele Ravanelli, Annalisa Trianni 👨🔬
Affiliation: Department of Information Engineering and Computer Science, University of Trento, Medical Physics Department, S.Chiara Hospital, APSS 🌍
Purpose: This study aims to leverage artificial intelligence (AI) to predict and identify performance degradation in Digital Radiography (DR) systems, enabling proactive maintenance and minimizing clinical disruptions.
Methods: Traditionally, medical physicists conduct periodic Quality Control (QC) tests on DR systems to verify performance. However, these tests provide only intermittent snapshots, and system degradation may go undetected between evaluations. In this study, we focus on a non-uniformity QC test, which assesses four key parameters: Local Signal Non-Uniformity (SLNU), Global Signal Non-Uniformity (GSNU), Local Signal-to-Noise Ratio Non-Uniformity (LSNRNU), and Global Signal-to-Noise Ratio Non-Uniformity (GSNRNU), with predefined tolerances. We analyzed a dataset of 1825 images from 19 DR systems (spanning five manufacturers) collected over 10 years. Each image was processed using an ImageJ script to extract these parameters. To predict system failures, a hybrid AI model was developed, combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, which perform well at capturing spatial and temporal dependencies in image data. The dataset was split into 80% training and 20% validation sets, with standardization and data augmentation techniques applied to address class imbalance, particularly for negative (failures) samples.
Results: After 100 training epochs, the model achieved an accuracy of 88.2% on validation set, effectively predicting whether DR system performance met the established QC criteria. The use of data augmentation improved the model's ability to generalize, particularly for cases with fewer negative labels.
Conclusion: This study demonstrates the feasibility of using AI-based models to predict DR system performance degradation based on long-term QC data. The predictive tool developed here has the potential to reduce the time to detect suboptimal system performance, revolutionizing traditional QC practices and enhancing system reliability.