Introduction: Artefacts appearing on magnetic resonance images can affect the quality of examination in a way to be confused with pathology or to cover the important information. Gibbs or ringing artefact is caused by the way in which the data are sampled and processed. On magnetic resonance images appears as multiple bright or dark lines parallel to edges of intensity change. Deep learning is a computer algorithm that takes metadata as an input and processes the data to compute the output. It can be used for detection of structures in images with use of convolutional neural networks. There are different methods of deep learning, but when no large dataset is available, the most useful method is transfer learning. Purpose: The aim of this study was to develop process of automatic detection of ringing artefact in magnetic resonance imaging for magnetic resonance scanner Philips Acieva 3.0 T TX with dStream system. Deep learning method that was used for detection is transfer learning. Methods: Dataset containing magnetic resonance images of phantom for quality assurance was produced for our research. Turbo spin echo pulse sequence in transversal plane was used for scanning, while we changed some of the scanning parameters. Slices, chosen for the research, were annotated by two independent observers in two categories: images with and without ringing artefact. With use of labelled dataset, we designed algorithm of detection of ringing artefact with use of transfer learning. As pretrained network to do transfer learning, we used convolutional neural network VGG16 and added two new layers, which we trained with use of our training dataset. Results: Automatic detection model for detecting ringing artefact on magnetic resonance imaging was tested on testing dataset and it showed great results. Accuracy of detecting ringing artefact on first type of magnetic resonance slice was 98 %, on second type 93 % and on third 98 %. All AUC values showing the quality of our build detection model, are above value 0,98. Discussion and conclusion: The accuracy of our build model can be compared with detection models in reviewed literature. There are no differences between models detecting different types of artefact or between models of detection algorithm using different type of pretrained convolutional neural network. Changing different scan parameters result in appearance of ringing artefact. Those parameters are matrix size, voxel size and number of averages. Phase encode direction does not affect on Gibbs artefact appearance. Automatic detection of artefacts on magnetic resonance images helps us to avoid the need for time-consuming manual review of images and it enables us to correct artefact with use of computer algorithms. With using automatic detection of artefact, we can ensure higher image quality of magnetic resonance imaging.
|
