Automated detection of motion artifacts in brain MR images using deep learning and explainable artificial intelligence

Marina Manso Jimeno,Keerthi Sravan Ravi,Maggie Fung,John Thomas Vaughan Jr.,Sairam Geethanath
2024-02-14
Abstract:Quality assessment, including inspecting the images for artifacts, is a critical step during MRI data acquisition to ensure data quality and downstream analysis or interpretation success. This study demonstrates a deep learning model to detect rigid motion in T1-weighted brain images. We leveraged a 2D CNN for three-class classification and tested it on publicly available retrospective and prospective datasets. Grad-CAM heatmaps enabled the identification of failure modes and provided an interpretation of the model's results. The model achieved average precision and recall metrics of 85% and 80% on six motion-simulated retrospective datasets. Additionally, the model's classifications on the prospective dataset showed a strong inverse correlation (-0.84) compared to average edge strength, an image quality metric indicative of motion. This model is part of the ArtifactID tool, aimed at inline automatic detection of Gibbs ringing, wrap-around, and motion artifacts. This tool automates part of the time-consuming QA process and augments expertise on-site, particularly relevant in low-resource settings where local MR knowledge is scarce.
Computer Vision and Pattern Recognition,Machine Learning
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: in brain magnetic resonance imaging (MRI), automatically detect the artifact problem caused by patient movement. Specifically, the author has developed a method based on deep learning and explainable artificial intelligence (XAI) to detect rigid motion artifacts in T1 - weighted brain images and classify them into three categories through a classification model: no motion, slight motion, and severe motion. ### Background of the problem 1. **Importance of quality assessment** - Quality assessment (QA) of MRI data is crucial for ensuring the accuracy of image analysis, interpretation, or diagnosis. - Artifact detection is a key step in the QA process and usually needs to be carried out during the imaging session to decide whether a repeat scan is required. 2. **Impact of motion artifacts** - Voluntary or involuntary movement of patients is a common source of artifacts in MRI imaging, which may lead to image blurring and ghost artifacts, reduce image quality, and affect quantitative analysis and interpretation. - Especially in long - scan sequences (such as 3D T1 - weighted imaging), motion artifacts are more common. 3. **Limitations of existing methods** - Manual detection of artifacts is time - consuming and depends on the knowledge of on - site MR physics experts. Especially in resource - limited environments, technicians may lack such professional knowledge. - Existing automatic detection methods mostly use binary classification models and cannot distinguish motion artifacts of different severities. ### Solutions in the paper 1. **Model architecture** - Use a lightweight 2D convolutional neural network (CNN) for three - class classification (no motion, slight motion, severe motion) for fast inference and applicability in resource - limited environments. - Use Grad - CAM heat maps to explain model prediction results, helping to identify failure modes and improve pre - processing steps. 2. **Dataset and experimental design** - Use publicly available datasets (such as IXI, HCP, etc.) for training and testing, including data of synthetic motion artifacts. - Verify the model performance in a prospective dataset and compare it by using the average edge strength (AES) as an indirect motion indicator. 3. **Performance evaluation** - The model has been tested on multiple retrospective and prospective datasets, showing good generalization ability. - Evaluate the model performance through multiple evaluation metrics (such as precision, recall, AUROC, etc.) and use Grad - CAM heat maps to explain model decisions. ### Conclusion The method proposed in this study can effectively detect motion artifacts in T1 - weighted brain images and provide more flexible quality assessment through multi - class classification. This tool can be integrated into the clinical workflow, improve scanning efficiency, and assist inexperienced technicians in quality assessment in resource - limited environments.