Item

Abstract

Sometimes magnetic resonance imaging (MRI) data is distorted by severe artifacts, due to non optimal sequence settings. Analyses of those images are going to produce false results and will confound affected studies. For that reason it is important to detect and remove datasets containing artifacts. Manual quality control of all images in large cohorts is not feasible; automatic methods to detect relevant artifacts are needed.
We trained a fully convolutional feed forward network to detect artifacts caused by imperfect fat suppression in MRI diffusion measurements (DWI) of the human head (128x128x72@1.7mm³/voxel, 60 diffusion directions, b=1000 s/mm²). The deep neural network was pre-trained for two epochs on an auto-encoding task using one million unlabeled slices from 2000 different DWI datasets. After pre-training the encoder was separated and extended with one additional layer. The prediction task was trained on 1200 labeled slices from 100 different DWI datasets with binary cross-entropy objective. The optimal training state was selected posterior based on the best accuracy.
Inference on 3303 before unseen slices achieved an accuracy of 0.82 (true positives: 179, true negatives: 2537, false positives: 26, false negatives: 561).
Exploration of misclassified images often showed positive response to very bright spots and negative response if the contrast between artifact and surrounding voxels was low.
Considering, that only few labels were available the achieved results are very promising. Predicting several slices for each dataset and applying majority voting might be used to further reduce false predictions.