In recent years, cardiovascular Magnetic Resonance Imaging (MRI) has become generally accepted as a diagnostic imaging modality to assess cardiovascular disease. This relatively harmless technique provides the radiologist and cardiologist with a fully three-dimensional and dynamic set of images of the beating heart (multiple frames per cardiac cycle). Apart from geometrical anatomical information, additional information about cardiac function can be collected during the same patient examination with specialized MR imaging techniques, involving for instance the state of the heart muscle (MR tagging or velocity encoded MR imaging), blood flow velocity (velocity encoded MR imaging) or the perfusion of the heart muscle ( dynamic perfusion studies).
Latest developments also indicate that the coronary arterial wall and plaque composition can be determined in vivo by MRI, which is of great clinical relevance in light of the recognition of the "vulnerable plaque".
Therefore MRI is increasingly referred to as the 'one-stop shop' for cardiac imaging (1): currently many clinical groups are pursuing a protocol, which allows the acquisition of all facets of cardiac function in one MR patient examination within one hour.
Because of the rapid progress in cardiovascular MR and CT acquisition techniques, a new problem has surfeced: the amount of image data involved in comprehensive patient study is massive (1500-5000), particulary with MRI stress studies including various stress levels (usually a baseline acquisition, 3-5 stress levels and a recovery acquisition). During each MRI acquisition typically 6-8 cross-sectional slices are acquired through the heart from base to apex, all with about 20 time phases over the cardiac cycle, plus a number of long-axis cross sections, adding up to about 200 images per cardiac acquisition. This all makes it a formidable task for the radiologist, cardiologist or technician to interpret the image data, let alone to analyze all these images with simple analytical tools. Moreover, manual analysis is subjective, and compromises the accuracy and reproducibility of quantitative measurements. In addition, in the current era of shortage of highly trained personnel, these lengthy and complicated procedures from the bottleneck for a widespread usage of these modern technologies, despite their tremendous potential. These factors have generated a great demand for tools that facilitate and further automate the accurate analysis of cardiac MR and CT patient examination: the radiologist/ cardiologist is eagerly awaiting this "single analysis button" which extracts relevant clinical information about the function of the heart and the status of the coronary arteries from these thousands of images, all within a matter of minutes.
Though a significant amount of techniques has been described to semi-automatically analyze cardiac function from MR images, many methods suffer from two practical limitations. First, although the amount of user interaction is reduced compared to manual contour drawing, the automated contour tracing is not yet sufficiently robust for use in routine clinical practice. Because of large variations in image quality and characteristics, a time-consuming manual correction of the automatically detected contours is often necessary, and for routine clinical use, it is absolutely necessary to further minimize the amount of user interaction. Second, most contour detection algorithms only operate on single 3D image sets, while there is a great need for the simultaneous analysis of multiple image sets (acquired during the same patient examination). From a clinical perspective, it would be of great practical value to analyze several image sets at the same time (for instance rest-stress pairs, short-axis and long-axis views, short-axis and perfusion pairs), because these image sets contains complementary diagnostic information. From a technical perspective, the fusion of complementary image information from several sources provides the opportunity to drastically improve the reliability of automated contour tracing algorithms. It would therefore be of great clinical and technical benefit to develop algorithms, which could be applied to analyze several MR image sets acquired during the same patient examination, at the same time.
The goal of the research proposed here is the development of highly robust automatic contour detection algorithms for the analysis of such 'one-stop-shop' cardiac MR patient examinations. To realize this, we will address two important, yet unresolved technical goals:
- the development of specialized contour detection techniques driven by different sources of complementary image information,
- The development of methods to integrate anatomical knowledge about cardiac shape, appearance and physiology into the automated analysis of cardiovascular MR images.
As a starting point, a number of model driven contour detection algorithms will be utilized, such as 3D Point Distribution Models, Active Appearance Models and CAD models. These methods demonstrated highly promising results in previous LKEB projects with respect to automated contour detection in cardiovascular images. However, futher fundamental research into these methods is required to meet the two previously described technical research goals.
In summary, the goal of this grant proposal is to develop algorithms to futher automate the quantification of cardiac function from 'one-stop-shop' cardiac MR patient examinations by developing novel methods to extract and combine information from all readily available image sets acquired during the same patient examination. This will have a significant impact on the efficience and the quality of the routine and clinical research MR procedures and will contribute to promoting the 'one-stop-shop' acquisition of cardiac MR images by enabling the analysis of all aspects of cardiac function from one single patient examination. The 'one-stop' cardiac patient examination demands an integrated, 'one-stop' post-processing approach.