TL;DR
This paper introduces a novel method for automatically detecting motion artifacts in MRI scans using decision forests. By simulating realistic motion artifacts in k-space and training a machine learning model, they demonstrate the potential for real-time feedback during scans, improving diagnostic quality and reducing the need for rescans.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) scans, while providing excellent soft tissue contrast, often require patients to remain still for extended periods. This can be challenging, especially for patients, leading to motion artifacts like ghosting, blurring, and smearing. These artifacts degrade image quality, potentially compromising diagnostic accuracy and necessitating repeat scans. Current methods for assessing image quality are often limited to specific body regions or require manual inspection. This paper addresses the need for an automated, robust, and generalizable method to detect these motion artifacts, enabling real-time feedback for clinicians and researchers.
Methods
The core of this work lies in a supervised machine learning approach using decision forests. To train and evaluate their method, the authors developed a strategy to synthetically introduce motion artifacts directly in k-space. This involves:
- Simulating Motion: They simulated two types of motion:
- Bulk Head Motion: Rigid motion (rotation and translation) applied to head MRI scans.
- Respiratory Motion: Periodic motion simulating breathing cycles applied to cardiac MRI scans.
- K-space Manipulation: Instead of simulating the entire MRI acquisition process, they directly altered the k-space data. This involved replacing portions of the original k-space data with data from a motion-transformed volume or from different points in a simulated breathing cycle. This approach is claimed to be more efficient and directly targets the source of motion artifacts.
- Feature Extraction: Various intensity-based features were extracted from the reconstructed images to serve as input for the decision forest classifier. These included:
- Box Features: Mean intensity and variance within random patches.
- Line Features: Standard deviation and intensity difference along random lines.
- Histogram Features: Mean, variance, skewness, kurtosis, entropy, and energy from image histograms.
- Texture Features: Haralick features to capture texture characteristics.
- Decision Forest Classification: A decision forest classifier was trained to distinguish between artifact-free and artifact-corrupted images. Nested cross-validation was used for robust evaluation.
A key novelty is the direct manipulation of k-space data to simulate artifacts, which is presented as a more direct way to generate realistic artifacts compared to full physics simulation.
Datasets
- Bulk Head Motion: T2-weighted MR scans from the IXI dataset, comprising 578 volumes of healthy subjects.
- Respiratory Motion: 3D scans acquired at Hammersmith Hospital as part of the UK Digital Heart project.
The authors generated synthetic datasets by introducing motion artifacts into these real scans.
Key Insights and Ideas
- K-space manipulation is effective for simulating realistic motion artifacts: Directly altering k-space data, rather than simulating the entire acquisition, proved sufficient to generate recognizable ghosting and blurring artifacts.
- Sampling pattern significantly impacts artifact detectability: Cartesian sampling is highly sensitive to motion, especially when it affects central k-space data. Radial and spiral sampling, however, are more robust, averaging out inconsistencies due to their acquisition strategy.
- Machine learning can effectively learn motion artifact characteristics: Decision forests, when trained on appropriate features, can achieve high accuracy in detecting motion artifacts across different sampling patterns.
- Feature choice is crucial and depends on the sampling pattern: Histogram-based features performed well for Cartesian and radial sampling in respiratory motion detection, while texture features excelled for spiral sampling.
- Real-time feedback is a tangible application: The ability to automatically detect motion artifacts opens doors for real-time feedback to clinicians, potentially improving scan quality and patient care.
Key Results
- Bulk Motion:
- Classification accuracy for Cartesian sampling was highly dependent on the fraction and offset of k-space affected, reaching up to 100% when central k-space data was altered.
- Radial and spiral sampling showed much higher robustness, with detection rates generally above 95% unless only a very small number of spokes/spirals were affected. The time point of motion within the acquisition had less impact on radial/spiral compared to Cartesian.
- Respiratory Motion:
- For Cartesian acquisition, histogram features achieved high accuracy (95.9% - 97.3%).
- For radial acquisition, detection rates were lower (up to 68.8% with combined features), attributed to the sampling pattern suppressing artifacts.
- For spiral acquisition, texture features performed best (88.3% - 93.8%), demonstrating robustness to motion.
Limitations and Future Work
- Limitations:
- The study primarily used synthetically generated artifacts. While the authors argue for their realism, evaluation on real, annotated datasets would be beneficial.
- The performance of the classifier can be influenced by the specific characteristics of the scanner and acquisition parameters, which might require careful training for different systems.
- The authors acknowledge that comparing their results to previous studies is difficult due to variations in datasets and methodologies.
- Future Work:
- Evaluating the developed machine learning approach on real, annotated data to validate the findings.
- Exploring features invariant to image scaling, translation, and rotation (e.g., SIFT) to handle recordings from different orientations.
- Investigating how ghosting artifacts might affect the performance of such invariant features.
Summary updated