Knowledge-guided boundary determination in low-contrast imagery: An application to medical images.

Abstract

Findings an accurate and complete description of left ventricular (LV) boundaries from a motion sequence of X-ray images is important in both qualitative and quantitative analysis of cardiac function. Due to the complications of non-rigid motion in a low-contrast X-ray image-sequence, the construction of a complete boundary is a compelling problem in Computer Vision. Expert knowledge in the form of models of the heart boundary and motion must be applied to guide the system. In addition, data-directed and goal-guided approaches must be combined using opportunistic problem-solving for control. We employ a blackboard architecture as the fundamental framework of our system. The blackboard is a powerful and modular architecture due to the partitioning of the domain knowledge into a set of functionally independent knowledge sources (KS's), the uniformity of the knowledge interaction, and the choice of domain dependent control strategies. This architecture allows us to build our system to handle such a knowledge intensive task. We show how edge points are extracted and linked in our system to form boundary fragments, how adjacent or overlapping fragments are collected and connected, how boundary fragments are tracked, how they are matched against heart models, how experimental models of the LV boundaries are derived, how uncertainties are dealt with using the correct importance and the confidence measures, how these modules (KS's) interact, and finally how KS activations are organized by the control structure. The system was evaluated based on the performance of each KS, individually, as well as its interaction with other KS's. The system was run on many image sequences and the results (discrete boundaries) were compared with manual tracings, confirming the correctness of the boundaries obtained by this approach. The boundary points are then interpolated to obtain a continuous representation. Our interpolation technique fits a cubic spline to the boundary edge points using the edge positions and tangent slopes derived from edge orientations, and computes tangent magnitudes by a minimization based on the second derivatives. It also handles unreliable edge orientations. We show how our technique does not suffer from the anomalies present in other methods. Overall, this system results in robust image interpretation which is required in medicine and many other applications of Computer Vision.