Room: Exhibit Hall | Forum 8
Purpose: To detect and classify artifacts in structural brain MR images using a deep convolutional neural network (DCNN).
Methods: This work included 932 brain images rated by two expert raters to have acceptable image quality, out from 1112 autism patients and healthy control subjects in the Autism Brain Imaging Data Exchange (ABIDE) database. The 932 brain volumes were corrupted with simulated patient motion and added noise performed in the spatial frequency domain (k-space). The constructed dataset had 2796 volumes, equally representing 3 different categories: artifact-free, motion corrupted, and noisy images. The constructed dataset was subdivided into training (60%), validation (20%), and testing (20%). Each of these partitions had equal representation of all image classes. The architecture of the DCNN consisted of an input layer, six convolution layers, one fully connected layer, and an output layer to classify each slice in the volume. The convolutional layers used 3x3 kernels and 2x2 max pooling operation for down sampling. The pooled feature map from the last convolutional layer was flattened into a single vector and fed to a fully-connected layer. The output layer provided the final image class. Three DCNNs were trained separately for the three MR image planes (axial, coronal, and sagittal). The predictions from the DCNNs were compared with the ground truth artifact type, and the classification accuracy was calculated.
Results: The accuracy for image quality evaluation and classification of different artifacts of the test set were 0.80, 0.81 and 0.83 when using axial, coronal and sagittal planes, respectively.
Conclusion: This work shows the potential of deep learning for detecting corrupted images and classifying the different artifacts in structural brain MRI in multi-center studies. Future investigations will include more artifact types. Detection and possible correction of artifacts will greatly enhance the diagnostic utility of MRI.
Funding Support, Disclosures, and Conflict of Interest: The authors gratefully acknowledge funding from NINDS/NIH grant #1R56NS105857-01, Endowed Chair in Biomedical Engineering, and Dunn Foundation. We acknowledge the support from the Texas Advanced Computing Center, Austin, TX for providing access to Maverick2 cluster.