Motion Corruption Detection in Breast DCE-MRI

Motion corruption can result in difficulty identifying lesions, and incorrect diagnoses by radiologists in cases of breast cancer screening using DCE-MRI. Although registration techniques can be used to correct for motion artifacts, their use has a computational cost and, in some cases can lead to a reduction in diagnostic quality rather than the desired improvement. In a clinical system it would be beneficial to identify automatically which studies have severe motion corruption and poor diagnostic quality and which studies have acceptable diagnostic quality. This information could then be used to restrict registration to only those cases where motion correction is needed, or it could be used to identify cases where motion correction fails. We have developed an automated method of estimating the degree of mis-registration present in a DCE-MRI study. We experiment using two predictive models; one based on a feature extraction method and a second one using a deep learning approach. These models are trained using estimates of deformation generated from unlabeled clinical data. We validate the predictions on a labeled dataset from radiologists denoting cases suffering from motion artifacts that affected their ability to interpret the image. By calculating a binary threshold on our predictions, we have managed to identify motion corrupted cases on our clinical dataset with an accuracy of 86% based on the area under the ROC curve. This approach is a novel attempt at defining a clinically relevant level of motion corruption.