[Paper Review] Deep Learning using K-space Based Data Augmentation for Automated Cardiac MR Motion Artefact Detection

İlkay Öksüz, Bram Ruijsink+7 more|arXiv (Cornell University)|Aug 15, 2018

Acute Ischemic Stroke Management12 references4 citations

TL;DR

This paper proposes a 3D spatio-temporal convolutional neural network (CNN) with k-space based data augmentation to automatically detect motion artefacts in cardiac MR images. By simulating realistic motion artefacts through k-space corruption, the method significantly improves recall (up to 0.652) and F1-score (0.722) on a highly imbalanced UK Biobank dataset of 3,465 CMR scans, outperforming state-of-the-art methods.

ABSTRACT

Quality assessment of medical images is essential for complete automation of image processing pipelines. For large population studies such as the UK Biobank, artefacts such as those caused by heart motion are problematic and manual identification is tedious and time-consuming. Therefore, there is an urgent need for automatic image quality assessment techniques. In this paper, we propose a method to automatically detect the presence of motion-related artefacts in cardiac magnetic resonance (CMR) images. As this is a highly imbalanced classification problem (due to the high number of good quality images compared to the low number of images with motion artefacts), we propose a novel k-space based training data augmentation approach in order to address this problem. Our method is based on 3D spatio-temporal Convolutional Neural Networks, and is able to detect 2D+time short axis images with motion artefacts in less than 1ms. We test our algorithm on a subset of the UK Biobank dataset consisting of 3465 CMR images and achieve not only high accuracy in detection of motion artefacts, but also high precision and recall. We compare our approach to a range of state-of-the-art quality assessment methods.

Motivation & Objective

• To address the challenge of automated motion artefact detection in large-scale cardiac MR imaging datasets like the UK Biobank.
• To overcome the severe class imbalance in CMR image quality assessment, where good-quality images vastly outnumber those with motion artefacts.
• To develop a realistic data augmentation strategy that generates synthetic motion artefacts using k-space corruption for improved model generalization.
• To evaluate the performance of a 3D CNN-based approach against state-of-the-art machine learning and image quality metrics on real clinical CMR data.
• To enable fully automated, high-throughput quality assessment for large population studies, supporting accurate cardiac function quantification.

Proposed method

• A 3D spatio-temporal CNN is trained to classify 2D+time short-axis CINE CMR sequences as either containing motion artefacts or being of good quality.
• Image preprocessing includes intensity normalization and motion-based region-of-interest (ROI) extraction to 80×80 pixels, avoiding image resizing.
• A novel k-space based data augmentation technique corrupts k-space data of good-quality images to simulate realistic motion-related artefacts, such as blurring.
• The k-space corruption process preserves phase and magnitude information, generating artefacts that closely resemble clinical motion artefacts, as shown in Fig. 1c.
• Additional data augmentation via translational shifts further improves model robustness and generalization.
• A 10-fold stratified cross-validation strategy is used to evaluate performance, with metrics including accuracy, precision, recall, and F1-score.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively detect motion artefacts in cardiac MR images from a large, real-world dataset like the UK Biobank?
• RQ2 How does k-space based data augmentation compare to traditional data augmentation in improving detection performance on highly imbalanced CMR datasets?
• RQ3 To what extent does the proposed method outperform state-of-the-art image quality assessment techniques in terms of recall and F1-score?
• RQ4 Can the model achieve high inference speed while maintaining high detection accuracy for clinical deployment?
• RQ5 Does the use of k-space corruption generate more realistic artefacts than synthetic or patch-based methods?

Key findings

• The proposed k-space based data augmentation significantly improves recall from 0.466 (no augmentation) to 0.642, demonstrating its effectiveness in addressing class imbalance.
• With both k-space and translational data augmentation, the model achieves a F1-score of 0.722, outperforming all other tested methods including Adaboost and traditional image quality metrics.
• The 3D CNN model achieves a precision of 0.809 and recall of 0.652, indicating strong performance in identifying true positive artefact cases.
• The method detects motion artefacts in less than 1ms per scan, enabling real-time or high-throughput processing in large-scale studies.
• The k-space corruption approach generates highly realistic artefacts that closely mimic clinical motion-related blurring, as validated by visual inspection and comparison with real artefacts.
• The model outperforms not only classical machine learning techniques but also established image quality metrics like NIQE and Variance of Laplacians in both recall and F1-score.