Challenges in CMR image acquisition:
Here, three major obstacles to be overcome include time inefficiency, operator dependency, and suboptimal image contrast. Limited time efficiency makes CMR examinations excessively long and costly, which translates into restricted information and patient discomfort due to repetitive breath-holds. Operator dependency, owing to CMR sequence selection, complex image planning, and patient interaction, affects the exam quality, precision, repeatability, accuracy, and the general diagnostic yield. Finally, suboptimal image contrast, particularly at the blood-scar interface, leads to ambiguity and makes the analysis of CMR images the most arduous step of CMR interpretation and the one associated with the longest learning curve.
Time inefficiency and Operator dependency: Push-button free-running CMR sequences help address both shortcomings. As they allow for ‘non-stop’ continuous data collection, examination time is abbreviated, meticulous scan planning no longer necessary, patient comfort enhanced, and time efficiency substantially increased. These techniques were initially built for coronary MR angiography and functional imaging. Combined with engineered magnetization pre-pulses (e.g. saturation, inversion), this milestone technology enables quantitative (i.e. objective interpretation by clinicians) assessment of the myocardial tissue.
Suboptimal image contrast: Bright-blood late gadolinium-enhanced (LGE) imaging is the current gold standard to assess myocardial injuries in patients. On LGE images, scar presence and distribution are the cornerstone of the etiological diagnosis of structural heart diseases, and the transmurality of scar (i.e. depth) is employed to assess myocardial viability in ischemic patients and predict the benefit of subsequent revascularization. Moreover, the burden and heterogeneity of scar on LGE images was shown to be a powerful predictor of ventricular arrhythmias, with the potential to significantly improve the primary prevention of sudden cardiac death with implantable cardioverter defibrillators. However, these clinical applications with direct impact on patient trajectories are all impaired by the poor contrast at the blood-scar interface, resulting in limited sensitivity to small scars, and limited robustness of scar quantification, as this heavily relies on accurate delineation of the endocardial border.
The articles published on the topic of image acquisition propose several solutions for:
- Limited time efficiency: Free-breathing 3D whole-heart imaging with 100% scan efficiency and advanced variable-density undersampling.
- Suboptimal image contrast: Black-blood scar imaging.
- Contrast-agent-free imaging: T1-rho mapping.
Challenges in CMR image analysis:
AI-based segmentation techniques are currently regarded as the gold standard for automated CMR analysis as they help support time efficiency, enable earlier detection of diseases, and improve diagnostic accuracy. In anatomical and functional imaging, convolutional neural networks (CNNs) enable highly accurate partition of the image into several meaningful areas based on which clinical CMR indices can be extracted, such as left and right ventricular volumes, ejection fraction, wall thickness, and myocardial mass.
In tissue characterization, existing state-of-the-art approaches for myocardial scar segmentation from bright-blood LGE (PSIR) images can be broadly categorized into automated (e.g. using clustering techniques, such as fuzzy c-means or super-pixel segmentation) or semi-automated (such as full-width at half-maximum) approaches, both requiring prior manual delineation of the left ventricular wall. Scar and wall segmentations are subsequently used to derive quantitative metrics (scar burden and transmurality) that may be directly employed to optimize patient management (decision to perform defibrillator implantation or revascularization). Although promising, these technologies require heavy operator expertise, are limited to 2-dimensional imaging, and the quantifications remain poorly reproducible, as the border between the scar and the blood pool remains poorly defined due to the lack of contrast on bright-blood images.
The articles published on the topic of image analysis propose several solutions for:
- MR operator dependency: fully automated TI-scout analysis.
- Radiologist dependency: fully automated T1 mapping analysis.
