Towards dynamic cardiac scenes interpretation based on spatial-temporal knowledge
Introduction
Motion interpretation is basically a visual understanding process. Depending on the context, specialists focus on both qualitative evaluations and quantitative measurements of dynamic objects. These approaches should be combined because they do not deliver alone, all the required elements to obtain a proper dynamic scene interpretation. Qualitative results give a deficient precision and vary from one specialist to another, depending on experience. On the other hand, numerical results depend on the applied algorithms, and can not directly give meaningful information, unless a post-processing is carried out. Disciplines like medical diagnosis [7], [12], [30], [51] and robot vision [13], [23], [24], [38] are privileged application fields, where efficient techniques for representing objects and events, as well as for reasoning about objects interaction are required. In these domains, the understanding of temporal image sequences aims to produce quantitative and qualitative evaluation of either objects or their dynamic behaviors, for example in regard to a particular organ physio-pathological motion, or the guidance of an autonomous vehicle.
Particularly, in medical diagnosis, computer assisted processing of an acquired image sequence raw data, is most likely restricted to the extraction of structures present in the images, their description and eventually their recognition by comparison with an anatomical model. Any analysis beyond those numerical facts, must use knowledge in order to pursue a computer aided scene analysis. The degree of recognition and interpretation of the analyzed scene depends entirely, on the abstraction level used to describe the extracted information, as well as on knowledge modeling. Cardiac movement analysis is a complex task carried out by a highly qualified specialist and is mainly concerned with pathological motion interpretation, related to coronaries circulation deficiencies or myocardial anomalies [25], [37]. Available imaging systems do not permit the acquisition of full volume real time spatial and temporal information about the cardiac muscle behavior during one or several heartbeats. Consequently, only some key-aspects revealing a trouble cardiac function are evaluated.
Quantitative measurements (left ventricular global performance, ventricular volume) and qualitative evaluations (motion phases, cavity morphology and infarction degree) are conventionally obtained from left ventricle images acquired in DSA (Digital Subtraction Angiography) imaging. A wide variety of methods have been intended to aid clinicians in this task [18], [26], [27], [47], seeking to produce quantitative evaluation of the left ventricular wall dynamic behavior. Nevertheless, none of them has gained general acceptance because of drawbacks produced by the model, its reference systems or the description parameters. Moreover, automatic cardiac motion interpretation as an aid for clinical decisions has mainly been approached using two-dimensional images, applying some of the methods mentioned above, to compare left ventricle contours at end-systole and end-diastole. A complete automatic interpretation should allow the labeling of anatomic structures, the detection and characterization of morphological and functional lesions, eventually leading to improved diagnostic support [9], [45]. Until now, research done in this field has demonstrated how difficult it is to integrate spatial and temporal information provided by images sequences [8].
Quantitative images evaluation followed by a computer aided scene interpretation, are seldom approached as being aspects of the same problem. It is necessary to complete them through progressive transformations of entities, from image analysis to understanding. Despite its importance, cardiac motion computer aided dynamic scene interpretation, has not been widely studied. Our work addresses cardiac scene dynamic interpretation, based on quantitative evaluation followed by computer aided scene interpretation. In addition, an alternative to 2D ventricular motion, 3D arterial motion, is studied. Other than the extensive use of coronary arteries observation to detect the position and degree of ischemic lesions [25], complementary analysis focused on potential pathological motion detection can be carried out by accessing 3D coronary arteries kinetic information. Since the arteries centerlines are in general spatially distributed over the cardiac muscle surface, their displacements are the same as those of the associated epicardium [31]. Therefore, it is possible to examine in detail the motion of the related anatomical regions along each artery centerline. Besides, ventricular and arterial angiography are routinely performed, under the same acquisition conditions and during the same examination.
The proposed experimental methodology originality resides in the use of calculated 3D arterial movement from bi-plane DSA image sequences, combined with spatial-temporal knowledge in a scene interpretation system, to perform cardiac motion interpretation. This work main contribution is to propose a general methodology to approach in an integrated manner, quantitative images evaluation followed by a knowledge-based scene interpretation. Spatial knowledge is constructed upon progressive descriptive objects which include anatomic segments, detected homogeneous segments and anatomic regions. Temporal knowledge is based on instant and interval representations, which are appropriate for identifying events. On top of this representation, mechanisms of spatial and temporal reasoning that implement deduction with symbolic constraints resolution are applied. They allow to obtain high level interpretation labels, like motion tendencies of contraction or expansion, which can concern local myocardial regions or global muscle behaviors. The rest of the paper is organized as follows. Section 2 presents a summary of previous works concerning dynamic scene interpretation. Section 3 introduces the idea of a knowledge-based approach in the context of cardiac motion interpretation. Section 4 describes how significant facts are obtained and represented. Section 5 formulates the numeric to symbolic transformation process. Section 6 presents the main ideas of the dynamic scene interpretation methodology. The complexity of envisioning interpretation is outlined in Section 7. Concluding remarks are presented in Section 8.
Section snippets
Related research
Most image understanding reported works are focused on low level image segmentation and some on medium level analysis, yet few include either the temporal dimension or high level analysis. Nazif and Levine [35] developed a rule-based system that segments natural scene images, to automatically understand their content. The knowledge of the system is organized into modules which make it possible to adjust the processing strategy. Low level processing like contour detection, segment shape, length
Towards a knowledge-based architecture
The main objective of computer aided dynamic scene interpretation in the medical context, is to be used as a diagnostic support tool. It concerns the generation and management of complex and large amounts of information, in order to allow anatomical structures labeling, as well as the detection and characterization of morphological and functional lesions. The two last ones are directly related to heart motion local and global kinetic analysis. Examining how clinicians proceed to understand
Obtaining significant facts
A knowledge-based architecture relies on facts databases upon which reasoning schemes are applied. Databases are generated after a quantitative and symbolic analysis of the available low level information, obtained from initial image processing. Facts databases contain the history of spatial-temporal events, that have been detected on an image sequence. Even if results are presented as clearly designed contours or physiological parameters, alone, they lack of the required structure to pursue
Numeric to symbolic transformation
Ventricular motion analysis is commonly defined by means of high level terms like akinesis or hypokinesis, to describe for example the contraction characteristics of a ventricular contour segment. Whereas coronary arteries motion analysis is not a clinical procedure, our purpose is to improve computer aided interpretation quality, providing an experimental set of qualitative labels. Moreover, information abstraction is increased, producing a completely new scope of analysis, which involves
Dynamic scene interpretation
After events’ detection and representation, the aim is to obtain interpretations from a complex set of related facts stored as symbols. A rule-based system, combined with spatial and temporal knowledge, is used for that purpose [40]. Rules are organized in several levels depending on complexity of the requested interpretation. Tracking of temporal and spatial events is achieved in this way, performing interpretations along arteries’ centerlines, arteries’ anatomical regions or irrigated cardiac
Envisioning interpretation
For the trained specialist it is not always possible to clearly identify the variety and simultaneity of cardiac muscle displacements in 3D. We propose multiple description and interpretation possibilities, only limited by user demands. Examples of possible applications are suggested. As described above, dynamic scene interpretation starts with the description of homogeneous segments detected along arteries reconstructed centerlines, and goes up to dynamic cycles identification. Despite the
Discussion and conclusions
The proposed methodology to perform coronary artery centerlines motion interpretation, originally approaches in an integrated manner, quantitative images evaluation followed by a knowledge-based scene interpretation. Applied knowledge is mainly geometrical, kinetic, temporal and anatomic. Displacement data is analyzed by means of a carefully selected group of dynamic features, which are intended to detect significant events, related to anatomical regions through time. These events are
Acknowledgements
This work was partly supported by Inter-American Development Bank/CONICIT project E-08 ‘Sistemas Inteligentes en Cardiologı́a’, the Programme de Coopération Post-Gradué between France and Venezuela (PCP-Biomédica) and ARECOM.
References (53)
Towards a general theory of action and time
Artif. Intell.
(1984)- et al.
An artificial intelligence approach for medical picture analysis
Pattern Recognit. Lett.
(1988) - et al.
A knowledge-based approach to integration of image processing procedures
CVGIP Image Understanding
(1993) - et al.
Artificial vision approach to the understanding of heart motion
J. Biomed. Eng.
(1992) - et al.
NeoGanesh: a working system for the automated control of assisted ventilation in ICUs
Artif. Intel. Med.
(1997) - et al.
Abstract temporal diagnosis in medical domains
Artif. Intel. Med.
(1997) - et al.
Assessment of regional myocardial performance from biplane coronary cineangiograms
Am. J. Cardiol.
(1971) - et al.
Understanding coronary artery movement: a knowledge-based approach
Artif. Intell. Med.
(1998) - et al.
Three-dimensional motion and reconstruction of coronary arteries from biplane cineangiography
Image Vision Comput.
(1994) - et al.
Knowledge-based analysis of carotid angiograms
Image Vision Comput.
(1985)
Automatic interpretation of medical image sequences
Pattern Recognit. Lett.
A knowledge-based system for the three-dimensional reconstruction of the cerebral blood vessels from a pair of stereoscopic angiograms
A framework for automatic analysis of the dynamic behavior of coronary angiograms
Int. J. Card. Imag.
Sur l'analyse de mouvement en imagerie médicale
Innov. Technol. Biol. Méd.
Computer vision approaches for three-dimensional reconstruction of coronary arteries: review and prospects
Crit. Rev. Biomed. Eng.
3-D Cardiac Modeling: a computer vision approach in spatio-temporal image sequences
J. Biol. Systems
Reconstruction of 3D left ventricular motion from plannar tagged cardiac MR images: an estimation theoretic approach
IEEE Trans. Med. Imag.
An integrated spatio-temporal approach to automatic visual guidance of autonomous vehicles
IEEE Trans. Syst. Man. Cybern.
Knowledge directed left ventricular boundary detection in equilibrium radionuclide angiography
IEEE Trans. Med. Imag.
Quantitative left ventricular wall motion analysis: a comparison of area, chord and radial methods
Circulation
Quelques propositions pour la réalisation d'un système expert de segmentation d'images
Traitement Signal
A knowledge-based approach for 3-D reconstruction and labeling of vascular networks from biplane angiographic projections
IEEE Trans. Med. Imag.
Cited by (5)
Motion estimation of 3D coronary vessel skeletons from X-ray angiographic sequences
2011, Computerized Medical Imaging and GraphicsCitation Excerpt :Since problems to be solved for both steps are to globally minimize a predefined function (i.e., energy function for sequential reconstruction and matching error function for motion estimation), both of the optimization problems can be solved with dynamic programming, an efficient optimization tool, to obtain global optimums with low computation cost and numerical stability. Compared with the method in [20,22–27], not only possible errors introduced by calibration parameters of the imaging system are avoided, but also estimation procedures are simplified. And application of DP algorithm ensures low computation cost.
Knowledge discovery approach to automated cardiac SPECT diagnosis
2001, Artificial Intelligence in MedicineAnalysis of coronary arterial dynamics from X-ray angiographic sequences
2009, ISCID 2009 - 2009 International Symposium on Computational Intelligence and DesignFunctional bull's eye based on three dimensional velocity information to quantify left ventricle motion in gated-SPECT
2008, Proceedings of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS'08 - "Personalized Healthcare through Technology"