Automatic MR image quality evaluation using a Deep CNN: A reference-free method to rate motion artifacts in neuroimaging.

Motion artifacts on magnetic resonance (MR) images degrade image quality and thus negatively affect clinical and research scanning. Considering the difficulty in preventing patient motion during MR examinations, the identification of motion artifact has attracted significant attention from researchers. We propose an automatic method for the evaluation of motion corrupted images using a deep convolutional neural network (CNN). Deep CNNs has been used widely in image classification tasks. While such methods require a significant amount of annotated training data, a scarce resource in medical imaging, the transfer learning and fine-tuning approaches allow us to use a smaller amount of data. Here we selected four renowned architectures, initially trained on Imagenet contest dataset, to fine-tune. The models were fine-tuned using patches from an annotated dataset composed of 68 T1-weighted volumetric acquisitions from healthy volunteers. For training and validation 48 images were used, while the remaining 20 images were used for testing. Each architecture was fine-tuned for each MR axis, detecting the motion artifact per patches from the three orthogonal MR acquisition axes. The overall average accuracy for the twelve models (three axes for each of four architecture) was 86.3%. As our goal was to detect fine-grained corruption in the image, we performed an extensive search on lower layers from each of the four architectures, since they filter small regions in the original input. Experiments showed that architectures with fewer layers than the original ones reported the better results for image patches with an overall average accuracy of 90.4%. The accuracies per architecture were similar so we decided to explore all four architectures performing a result consensus. Also, to determine the probability of motion artifacts presence on the whole acquisition a combination of the three axes were performed. The final architecture consists of an artificial neural network (ANN) classifier combining all models from the four shallower architectures, which overall acquisition-based accuracy was 100.0%. The proposed method generalization was tested using three different MR data: (1) MR image acquired in epilepsy patients (93 acquisitions); (2) MR image presenting susceptibility artifact (22 acquisitions); and (3) MR image acquired from different scanner vendor (20 acquisitions). The achieved acquisition-based accuracy on generalization tests (1) 90.3%, (2) 63.6%, and (3) 75.0%) suggests that domain adaptation is necessary. Our proposed method can be rapidly applied to large amounts of image data, providing a motion probability p∈[0,1] per acquisition. This method output can be used as a scale to identify the motion corrupted images from the dataset, thus minimizing the time spent on visual quality control.