Outlier detection based on the neural network for tensor estimation

•Wavelet transform were used to extract features from the DWIs.•PCA was used to further reduce the features.•FNN was employed to construct the classifier.•A Rossler-based chaotic particle swarm optimization method was used to train the FNN.•The proposed method removed the influence of unexpected motion artifacts in DWI.

Rationale and objectivesDiffusion weighted imaging (DWI) is always influenced by both thermal noise and spatially/temporally varying artifacts such as subject motion and cardiac pulsation. Motion artifacts are particularly prevalent, especially when scanning an uncooperative population with several disorders. Some motion between acquisitions can be corrected by co-registration approaches. However, automated and accurate motion outlier detection of brain DWIs is an integral component of the analysis and interpretation of tensor estimation. Many different and innovative methods have been proposed to improve upon this technology. In this study, we proposed a classifier work frame, which can classify DWIs as normal images or motion artifacts.Materials and methodsThe procedure contains the following stages: first, we used the wavelet transform to extract features from the original DWIs; second, the principle component analysis was used to reduce the features; third, the forward neural network (FNN) was employed to construct the classifier; fourth, a Rossler-based chaotic particle swarm optimization method was proposed to train the FNN; fifth, the cost matrix was determined as the false negative (FN) which was 10 times larger than the false position (FP); and finally, the K-fold cross validation was chosen to avoid overfitting. We applied this method on 60 DWI datasets, including 50 training datasets and 10 test datasets.ResultsThe experimental results based on our DWI database showed that the proposed method can effectively extract the global feature from images and achieve better performance in tensor estimation by automatic unvoxelwise outlier rejection compared with manual and visual inspection, and previous voxelwise outlier rejection methods. We found that the motion artifact detection accuracy on both the training and test datasets was over 95.8%, while the computation time per DWI slice was only 0.0149 s.ConclusionThe proposed method could potentially remove the influence of unexpected motion artifacts in DWI acquisitions and should be applicable to other magnetic resonance imaging.