Neural network for step anomaly detection in head motion during fMRI using meta-learning adaptation

Quality assessment and artifact detection in functional magnetic resonance imaging (fMRI) data is essential for clinical applications and brain research. Subject head motion remains the main source of artifacts - even the tiniest head movement can perturb the structural and functional data derived from the fMRI. In this paper, we propose an end-to-end neural network technology for detecting step anomalies with training on partially synthetic data with adaptation to a specific small set of real data. A procedure for generating a synthetic dataset for training and a module for automated labeling of real data is developed. A recurrent neural network model for detecting step anomalies is proposed. A method for the model adaptation to a small set of real data based on one-step meta-learning is developed. An experimental verification of the accuracy is carried out in the problem of detecting step anomalies using a sliding window of 10, 15, and 24 pixels. The experiments have shown the proposed technology to provide the detection of stepwise anomalies with an accuracy of 0.9546.