An Adaptive Filter to Approximate the Bayesian Strategy for Sonographic Beamforming

A first-principles task-based approach to the design of medical ultrasonic imaging systems for breast lesion discrimination is described. This study explores a new approximation to the ideal Bayesian observer strategy that allows for object heterogeneity. The new method, called iterative Wiener filtering, is implemented using echo data simulations and a phantom study. We studied five lesion features closely associated with visual discrimination for clinical diagnosis. A series of human observer measurements for the same image data allowed us to quantitatively compare alternative beamforming strategies through measurements of visual discrimination efficiency. Employing the Smith-Wagner model observer, we were able to breakdown efficiency estimates and identify the processing stage at which performance losses occur. The methods were implemented using a commercial scanner and a cyst phantom to explore development of spatial filters for systems with shift-variant impulse response functions. Overall we found that significant improvements were realized over standard B-mode images using a delay-and-sum beamformer but at the cost of higher complexity and computational load.