Predicting Head Motions During Brain MRI Scans Using Deep Learning Models
MRI is a medical machine with a large magnet and radio waves to scan organs and body structures, especially in the head scans. During the scanning process, head motion is a major source of error for brain MRI. There is no direct method to assess head motion during MRI scans. To solve this problem, we generate deep learning models to predict the head motion based on an in-scanner video obtained from an in-scanner eye tracker. There are two applications in the medical field. One application is to save time for both patients and doctors since the doctor could pause MRI scans when detecting large head motion. The other application is to deblur the low-quality head MRI pictures based on the head motion. We generated a classification model that classifies each video frame as “moving” or “no-moving”. Besides, we built a regression model that predicts exact head motion values in rotation and translation directions. The Convolutional Neural Network is based on “VGG16” with seven dense layers. For the classification model, it achieves an overall accuracy of 83.84% with 89.26% and 72.91% for the “moving” and “no-moving” frames. For the regression models, we use the R-squared score to evaluate the model of rotational motion. The performance for validation data is excellent. For testing data, the model does not work as well as that on the validation data but the performance still be of a high standard for the practical value.
Computer science|Electrical engineering
Yuan, Hui, "Predicting Head Motions During Brain MRI Scans Using Deep Learning Models" (2020). ETD Collection for Fordham University. AAI27959273.