An approach to system optimization for X-Ray photon-counting systems using performance on a detection/localization task

We address the problem of optimizing data acquisition for photon-counting CT. We formulate a task-driven approach using a clinically relevant task of detection and localization of a lesion in a search region. The appropriate scalar measure of task performance is ALROC, the area under the LROC curve. For hardware optimization, the observer performing the task should operate on the raw (sinogram) data so that the best possible information, independent of the parameters of any particular reconstruction algorithm, is collected. To carry out the task, we use the ideal observer (IO), a numerical observer that yields the maximal ALROC amongst all observers. This differs from the usual numerical observer that operates on the reconstructed data and is designed so that its performance tracks that of a human observer. So while the optimization is carried out in the sinogram domain, the task definition is in the object domain. In this work, we mathematically specialize a previously developed IO to the case of photon-counting transmission tomography. We applied the IO to a 2D simulation of CT with the task of detecting a 3mm lung nodule in a lung region of a 512×512 phantom. The application was to optimize the number of angular acquisitions given a fixed dose. A plot of ALROC vs. angle number peaks at 42 angles. We also plotted ALROC vs. dose (at 105 angles) to characterize the increase in task performance with count level.