Application of synthetic data in the training of artificial intelligence for automated quality assurance in magnetic resonance imaging

Abstract

AbstractBackgroundMagnetic resonance imaging scanner faults can be missed during routine quality assurance (QA) if they are subtle, intermittent, or the test being performed is insensitive to the type of fault. Coil element malfunction is a common fault within MRI scanners, which may go undetected for quite some time. Consequently, this may lead to poor image quality and the potential for misdiagnoses.PurposeDaily QA typically consists of an automated signal to noise ratio test and in some instances this test is insensitive to coil element malfunction. Instead of relying on daily QA testing, it was proposed to utilize patient images in conjunction with a trained neural network to detect coil element malfunction, even when it presents as a very subtle defect. The advantage to using patient images over phantom testing is real‐time monitoring can be achieved. This allows clinical staff to focus more on patient throughput without being burdened by daily testing.MethodsA neural network was trained using simulated coil failure in 3958 abdominal or pelvic images from 497 patients. The accuracy of the trained network was then tested on an unseen dataset of 109 images from which 44 patients which had coil element malfunction present. Five MRI radiographers were shown 249 images with and without real coil malfunction to assess their accuracy compared to the neural network in identifying the scanner fault.ResultsThe neural network achieved an accuracy of 91.74% in identifying coil element malfunction in the unseen data. Radiographers tasked with identifying coil element malfunction had an average accuracy of 59.99%. In the same test case, the neural network outperformed all radiographers with an accuracy of 91.56%.ConclusionThis work demonstrates that neural networks trained with artificial data can successfully identify MRI scanner coil element malfunction in clinical images. The method provided better accuracy than MRI radiographers (technologists) at identifying coil element malfunction and highlights the potential utility of AI methods as an alternative to support traditional QA. Further, our methodology of training neural networks with simulated data could potentially identify other faults, allowing centers to produce robust fault detection systems with minimal data.