Introduction

Patient motion is a frequent cause of degraded image quality in MR imaging. Mild motion artifacts do not necessarily render an image clinically useless, yet severe artifacts require the acquisition to be repeated. Taking the decision to repeat or not requires considerable skill from the system operator. This study explores the possibility of making this decision in real time by generic image feature extraction and machine learning immediately after scan completion. The aim is to base the decision only on a limited set of generic image features that can be calculated very fast.

Methods

Expert radiologists have classified the images of 200 neurological patient exams into acceptable and inacceptable image quality, with the acquisitions of the latter class needing to be repeated. Images with artifacts not caused by motion were excluded from the study. All acquisitions are 2D and performed at 1.5T, but include different orientations (axial, sagittal and coronal), various contrasts (including T1, T2, T2*, T1 FLAIR and T2 FLAIR) and both cartesian and propeller readouts. For the images of inacceptable quality, non-expert readers further identified the slices clearly affected by motion and discarded the other image slices. Finally, the images were randomly split into training and testing data for machine learning, resulting into the split of figure 1. For each labeled image slice, a total of 108 brain image agnostic features were calculated either on the entire slice, or simple decompositions thereof by e.g. fore/background thresholding, wavelet decomposition or discrete cosine transform. The features extracted included noise metrics, focus measures¹, grey level co-occurrences and classical texture features². These features were further supplemented with 12 meta data entries from the image header, including e.g. the repetition and echo time, the flip angle and the slice orientation. At first a support vector machine (2 class C-SVM with radial basis function kernel, LibSVM 3.22³) was trained to classify slices of acceptable and inacceptable image quality based on the full set of 120 features. The optimal SVM hyperparameter values were determined by grid search and cross-validation with a leave (all slices of) 15 series out strategy. To partially offset the large imbalance between the number of slices with acceptable and inacceptable quality, the inacceptable slices were weighted higher accordingly, yet without attempt to optimize this weight. Next, recursive feature elimination⁴, eliminating just a single feature at a time, was performed to the 120 features to rank them according to their importance for the classification. Afterwards, a new support vector training was performed using only the 20 top-ranking features. Finally, the performance of the 2 SVM’s is compared.

Results

The 120 feature SVM yielded an overall accuracy of classification of 90.1% on the validation data and 85.7% on the testing data. Figure 2 provides a complete overview of the performance. The 20 feature SVM yielded only slightly worse performance with an overall classification accuracy of 87.7% on the validation data and 83.3% on the testing data (see also figure 2). For both SVM’s, the classification performance on acceptable slices is further somewhat higher than for inacceptable slices. From the 120 features, the recursive feature elimination selected 2 meta data features: the repetition time and the flip angle of the sequence.

Discussion

The decision of whether an MRI acquisition is of sufficient quality for diagnosis can be difficult to make. This study has investigated the possibility to automate/facilitate this decision in real time by generic image feature extraction and machine learning. The study was limited to brain imaging, but included a wide range of different protocols of clinical interest. Furthermore, the features used do not require any brain specific image processing like e.g. gray/white matter segmentation. This ensures fast and reliable feature extraction and it also makes the method easily applicable to other body parts and protocols. The results of the study are encouraging, especially when taking into account that the results provided are for the classification of individual slices. In practice, a decision for an entire series volume is required based on the classifications of the composing slices. The misclassification of just a few slices doesn’t have to lead to the misclassification of the entire volume. The study has also demonstrated that, unlike in other studies⁵, a limited set of features can already be sufficient for good performance. The results finally also suggest that the inclusion of specific sequence parameters is beneficial for the classification accuracy.

Conclusion

Generic image feature extraction in combination with support vector machine learning is a promising method to judge the clinical quality of MR images.

Acknowledgements

No acknowledgement found.