A mathematical model approach towards combining information from multiple image projections of the same patient
The purpose of this study was to, i) use a mathematical observer model to combine information obtained from multiple angular projections of the same breast to determine the overall detectability of a simulated lesion in a multi-projection breast imaging system and, ii) determine the optimum acquisition parameters of such a system. Multi-projection imaging is similar to tomosynthesis, except that the raw projection images are directly analyzed instead of reconstructing those images, thereby avoiding reconstruction artifacts. 25 angular projections of each breast from 82 human subjects in our tomosynthesis clinical trials were supplemented with projections from a simulated 3 mm 3D lesion. The lesion was assumed to be embedded in the compressed breast at a distance of 3 cm from the detector. The contrast of the lesion was determined taking into account the energy spectrum of the x-ray beam, properties of the digital detector, scatter fraction, and compressed breast thickness. A linear Hotelling observer with Laguerre-Gauss channels (LG CHO) was applied to each image. Detectability was analyzed in terms of ROC curves and the area under ROC curves (AUC). Three different methods were used to integrate ROCs from multiple (correlated) views to obtain one combined ROC as an overall metric of detectability. Specifically, 1) ROCs from different projections were simply averaged; 2) the test statistics from different projections were averaged; and 3) a Bayesian decision fusion rule was used. Finally, the number of angular projections, angular span and the acquisition dose level were optimized for highest AUC of the combined ROC as a parameter to maximize the performance of the system. It was found that the Bayesian decision fusion technique performs better than the other two techniques and likely offers the best approximation of the diagnostic process. Furthermore, if the total dose level is held constant at 1/25 th of the standard dual-view mammographic screening dose, the highest detectability performance is observed when considering only two projections spread along an angular span of 11.4°.