Automated Image Quality Evaluation of Structural Brain Magnetic Resonance Images using Deep Convolutional Neural Networks

EasyChair Preprint 650

Abstract

Automated evaluation of image quality is essential to assure accurate diagnosis and effective patient management. This is particularly important for multi-center studies, typically employed in clinical trials, in which the data are acquired on different machines with different protocols. Visual quality assessment of magnetic resonance imaging (MRI) data is subjective and impractical for large datasets. Data-intensive deep learning methods such as convolutional neural networks (CNNs) are promising tools for processing large-scale imaging datasets for automated quality assessment. In this study, we evaluate a CNN-based method for quality assessment of the Autism Brain Imaging Data Exchange (ABIDE) structural brain MRI dataset acquired from 17 sites on more than a thousand subjects. The CNN architecture consisted of an input layer, four convolution layers, two fully connected layers, and an output layer. A balanced set of 348 image volumes was used in the study. 60% of the data was used for training, 15% for validation, and 25% for testing. The results of the automated approach were compared with the evaluation by the radiologist. Performance of the CNN was assessed using the confusion matrix. The concordance in image quality labels between the expert and CNN was 86% (sensitivity = 81%, specificity = 92%). The present study shows that the proposed model can evaluate the image quality of brain MRI with higher classification accuracy compared to previous state-of-the-art classical machine learning algorithms.

Keyphrases: Convolutional Neural Networks, Radiology, deep learning, quality control

@booklet{EasyChair:650,
  author    = {Sheeba Sujit and Refaat Gabr and Ivan Coronado and Melvin Robinson and Sushmita Datta and Ponnada Narayana},
  title     = {Automated Image Quality Evaluation of Structural Brain Magnetic Resonance Images using Deep Convolutional Neural Networks},
  doi       = {10.29007/j68t},
  howpublished = {EasyChair Preprint 650},
  year      = {EasyChair, 2018}}