Source-linked AI summary
Automatic CNN-based detection of cardiac MR motion artefacts using k-space data augmentation and curriculum learning
Ilkay Oksuz, Bram Ruijsink, Esther Puyol-Anton, James Clough, Gastao Cruz, Aurelien Bustin, Claudia Prieto, Rene Botnar, Daniel Rueckert, Julia A. Schnabel, Andrew P. King
TL;DR
Poor-quality medical images can undermine subsequent analysis, and manual motion-artefact review is time-consuming. The paper evaluates deep-learning detection of motion artefacts in cine CMR using synthetic k-space corruption and a severity-based curriculum. It reports improved classification performance with these strategies and better performance for curriculum learning than anti-curriculum or random sampling.
Problem
Manual quality assessment of motion artefacts in cine CMR is time-consuming, while poor-quality images can mislead subsequent analysis.
Source: §1 · pp. 3–4Method
The paper evaluates deep-learning motion-artefact detection using synthetic k-space corruption and training samples ordered by corruption severity.
Source: §1 · p. 4, §4.4 · p. 21Results
K-space corruption outperformed other tested augmentation strategies, and curriculum learning performed better than anti-curriculum and random sampling.
Source: §6 · p. 34Takeaways & Limitations
The authors conclude that k-space corruption and curriculum learning improve motion-artefact detection accuracy in their evaluated setting.
Source: §6 · p. 35The evaluation uses existing network architectures and loss functions; basal and apical slice quality and tailored architectures remain future work.
Source: §6 · p. 34
Abstract
from arXiv · showhide
Good quality of medical images is a prerequisite for the success of subsequent image analysis pipelines. Quality assessment of medical images is therefore an essential activity and for large population studies such as the UK Biobank (UKBB), manual identification of artefacts such as those caused by unanticipated motion 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) cine images. We compare two deep learning architectures to classify poor quality CMR images: 1) 3D spatio-temporal Convolutional Neural Networks (3D-CNN), 2) Long-term Recurrent Convolutional Network (LRCN). Though in real clinical setup motion artefacts are common, high-quality imaging of UKBB, which comprises cross-sectional population data of volunteers who do not necessarily have health problems creates a highly imbalanced classification problem. Due to the high number of good quality images compared to the relatively low number of images with motion artefacts, we propose a novel data augmentation scheme based on synthetic artefact creation in k-space. We also investigate a learning approach using a predetermined curriculum based on synthetic artefact severity. We evaluate our pipeline on a subset of the UK Biobank data set consisting of 3510 CMR images. The LRCN architecture outperformed the 3D-CNN architecture and was able to detect 2D+time short axis images with motion artefacts in less than 1ms with high recall. We compare our approach to a range of state-of-the-art quality assessment methods. The novel data augmentation and curriculum learning approaches both improved classification performance achieving overall area under the ROC curve of 0.89.
1. Introduction
Reliable image-quality assessment matters because poor-quality CMR can mislead downstream analysis, while visual review is time-consuming and variable. The paper proposes automated motion-artefact detection for short-axis cine CMR, using synthetic k-space corruptions and a severity-based curriculum.
1. Introduction
- Poor-quality CMR can mislead analysis, and visual quality assessment is time-consuming and prone to inter- and intra-rater variability.The introduction notes that CMR supports cardiac-function assessment, including volumes, ejection fractions, and myocardial wall motion.Source: §1 · p. 3
- Motion artefacts, including mistriggering, arrhythmia, and breathing artefacts, can blur CMR images and make subsequent processing difficult.The paper notes these artefacts are more common in clinical acquisitions and can lead to erroneous quantification of myocardial wall motion.Source: §1 · pp. 3–4
- Automatic quality assessment is needed for large imaging studies such as UK Biobank, whose data are intended to maximise research value.The introduction says UK Biobank will eventually include CMR images from 100,000 subjects.Source: §1 · p. 3
- The paper proposes fully automated motion-artefact detection in short-axis cine CMR, with synthetic k-space artefacts at varying severity and a severity-based curriculum.The approach extends prior work on synthetic mistriggering artefacts to include breathing artefacts and multiple corruption levels.Source: §1 · p. 4
2. Related Works
The review surveys image quality assessment, class-imbalance methods, and curriculum learning before presenting the paper’s contributions to cine CMR motion-artefact detection.
2.1. Image quality assessment
- Medical image-quality work spans no-reference metrics and CNN approaches, while prior CMR studies address missing slices, coverage, motion, contrast, and segmentation quality.Earlier MR quality measures used simplistic distortions or site-specific thresholds; one motion-detection study used only artificially corrupted data and targeted breathing artefacts.Source: §2.1 · pp. 6–8
2.2. Data imbalance
- Imbalanced class distributions can bias classifiers toward the majority class, leaving minority-class features underlearned and minority samples misclassified.Source: §2.2 · pp. 8–9
2.2.1. Sampling techniques
- Random oversampling duplicates minority samples and may overfit, whereas random undersampling removes majority samples and may leave too little training data.Source: §2.2.1 · p. 9
2.2.2. Cost sensitive learning
- Cost-sensitive learning assigns different misclassification costs to classes, but requires appropriate costs that may be difficult to specify.Source: §2.2.2 · pp. 9–10
2.2.3. Data imbalance problem for neural networks
- For neural networks, the paper relates its minority-class synthetic-data strategy to SMOTE, but generates examples using prior knowledge of cine MR acquisition.Source: §2.2.3 · p. 10
2.4. Contributions
- The authors describe this as the first thorough analysis of machine-learning methods for automatic cine CMR motion-artefact detection on a large in-vivo database.Source: §2.4 · p. 12
- The proposed augmentation simulates motion artefacts through k-space corruption at varying severity, extending prior work to breathing and mistriggering artefacts.Source: §2.4 · pp. 12–13
- The proposed curriculum trains models on synthetic samples of increasing difficulty, using artefact severity to structure training.Source: §2.4 · pp. 12–13
3. Materials
The study evaluates short-axis cine CMR data from 3,510 UK Biobank subjects, with expert-verified binary quality labels.
3. Materials
- The analysis used 50-frame bSSFP sequences from short-axis stacks, with 1.8 × 1.8 mm2 in-plane resolution and 8.0 mm slice thickness.Stacks typically contained approximately 10 slices with a 2 mm gap.Source: §3 · p. 13
- The subset excluded other image-quality problems such as missing axial slices, and an expert cardiologist visually verified the data and validated binary labels.Source: §3 · pp. 13–14
- 3,360 of 3,510 acquisitions were good quality and 150 had motion artefacts, including mistriggering, breathing, arrhythmia, and mixed cases.Source: §3 · p. 13
4. Methods
The methods prepare consistent cardiac regions, compare two sequence-classification architectures, and address class imbalance with synthetic k-space artefacts and severity-based curriculum learning.
4.1. Preprocessing
- Fourier-domain cardiac activity and Hough-circle voting locate the left-ventricle centre, from which 80 × 80 regions are extracted for processing.The preprocessing was reported to correctly identify the myocardial cavity in all cases and was validated using myocardial masks.Source: §4.1 · pp. 14–15, Figure 2 · p. 15
4.2. Deep learning models
- The 3D-CNN processes time as a third dimension, while the LRCN extracts frame features and uses an LSTM for sequence classification.The 3D-CNN takes 80 × 80 crops with 50 time frames; the LRCN feature extractor has six convolutional and three pooling layers.Source: §4.2 · pp. 16–17, Figure 3 · p. 16, Figure 4 · p. 18
4.3. Balancing the classes
- To address class imbalance, the method generates synthetic mistriggering and breathing artefacts at varying severity using the cine MR acquisition process.The k-space strategy is intended to augment the low-quality training examples with realistic motion artefacts.Source: §4.3 · pp. 17–18
4.3.1. Mistriggering artefacts
- Mistriggering corruption replaces Cartesian k-space lines with lines from other cardiac phases, with the replacement rate controlling artefact severity.The corrupted training examples are generated from original good-quality images, reflecting the misallocation of k-space lines associated with mistriggering.Source: §4.3.1 · pp. 18–19, Figure 5 · p. 19
4.3.2. Breathing artefacts
- Breathing artefacts are simulated by translating frames in a sinusoidal pattern before combining their k-space representations and reconstructing images.The simulation uses four breathing cycles over 256 phase-encoding steps.Source: §4.3.2 · pp. 19–20, Figure 6 · p. 20
4.4. Curriculum learning
- Training begins with heavily corrupted, easier images and gradually adds less-corrupted, harder images according to corruption-based difficulty levels.The baby-step strategy retains previously introduced samples in the training pool as new levels are added.Source: §4.4 · pp. 21–22, Figure 7 · p. 22
4.5. Loss functions and optimisation
- Stochastic gradient descent minimizes binary cross-entropy, and training stops after 100 validation epochs without sufficient improvement.Sufficient improvement is defined as a relative increase of at least 0.5%; training used batches of 50 sequences, a 0.90 momentum, and a 0.0001 learning rate.Source: §4.5 · pp. 23–24
5. Experiments and results
The experiments compare deep-learning models and alternative classifiers, test augmentation strategies on imbalanced in-vivo data, and evaluate curriculum learning. Curriculum learning significantly outperformed random and anti-curriculum sampling, while LRCN showed high recall for in-vivo motion artefacts.
5.1. Evaluation metrics and methods of comparison
- Evaluation used repeated stratified 10-fold cross-validation and compared classifiers with NIQE and Variance of Laplacians baselines.The comparison also included K-nearest neighbours, SVMs, decision trees, random forests, Adaboost and Naive Bayesian.Source: §5.1 · pp. 25–26
5.2. Synthetic data
- On balanced synthetic data, LRCN and 3D-CNN had the highest accuracy, recall and balanced accuracy for both breathing and mistriggering artefacts.The authors attribute generally high performance to the low complexity of distinguishing originals from synthetically corrupted versions and to balanced data.Source: §5.2 · pp. 26–27, Table 1 · p. 27
5.3. Augmentation technique analysis
- Six in-vivo training configurations tested translation, synthetic mistriggering and breathing augmentation, their combination, and cost-sensitive learning.The cost-sensitive setup used weighted binary cross-entropy with weights determined by the 150:3360 class ratio; translation augmentation was applied to both classes.Source: §5.3 · pp. 27–28, Table 3 · p. 30
- LRCN identified in-vivo motion artefacts with high recall; accuracy exceeded 0.9 for all algorithms under severe class imbalance and was omitted.The in-vivo evaluation used 150 artefact and 3360 good-quality images.Source: §5.3 · pp. 27, 29
5.4. Influence of curriculum learning
- The curriculum used ten corruption levels, beginning with highly corrupted easy samples and gradually adding less corrupted, harder samples.The comparison included a reversed hard-to-easy anti-curriculum and randomly ordered control-curriculum samples.Source: §5.4 · pp. 29, 31
- Curriculum learning significantly outperformed random sampling and anti-curriculum learning (pvalue < 0.05) in DeLong’s test and enabled detection of borderline cases of motion artefacts with great success compared to control-curriculum.Source: §5.4 · p. 31
6. Discussion and conclusion
The study presents k-space corruption augmentation and curriculum learning for motion-artefact detection, reporting improved classification and better performance than alternative sampling strategies. Its scope reflects the controlled UK Biobank data and leaves several model and image-quality questions for future work.
6. Discussion and conclusion
- Future work includes novel loss functions and architectures, as well as assessing basal and apical slice quality.The study deliberately used existing architectures and loss functions to focus on augmentation and curriculum learning.Source: §6 · p. 34
- The UK Biobank is a controlled study with few motion-artefact images, unlike clinical acquisitions where such artefacts are more likely.The authors note that classes would remain imbalanced in clinical acquisitions.Source: §6 · p. 35
- K-space corruption augmentation and curriculum learning improved motion-artefact detection, with curriculum learning outperforming anti-curriculum and random sampling.The curriculum strategy also improved classification on borderline cases compared with control-curriculum.Source: Figure 9 · p. 32, Figure 10 · p. 33, §6 · pp. 34–35
- Automated quality control is presented as a way to help ensure the accuracy of subsequent analysis pipelines.The conclusion frames this within growing clinical imaging use and large population imaging cohorts.Source: §6 · p. 35
Data Access Statement
The imaging data were provided by the UK Biobank Resource under Application Number 17806.
Data Access Statement
- Researchers may apply to access the UK Biobank data resource for health-related research in the public interest.The dataset is freely available upon request.Source: Data Access Statement · p. 36