Task-based statistical image reconstruction for high-quality cone-beam CT.
Hao Dang,J Webster Stayman,Jennifer Xu,Wojciech Zbijewski,Alejandro Sisniega,Michael Mow,Xiaohui Wang,David H Foos,Nafi Aygun,Vassilis E Koliatsos,Jeffrey H Siewerdsen
DOI: https://doi.org/10.1088/1361-6560/aa90fd
IF: 3.5
2017-01-01
Physics in Medicine and Biology
Abstract:Task-based analysis of medical imaging performance underlies many ongoing efforts in the development of new imaging systems. In statistical image reconstruction, regularization is often formulated in terms to encourage smoothness and/or sharpness (e.g. a linear, quadratic, or Huber penalty) but without explicit formulation of the task. We propose an alternative regularization approach in which a spatially varying penalty is determined that maximizes task-based imaging performance at every location in a 3D image. We apply the method to model-based image reconstruction (MBIR-viz., penalized weighted least-squares, PWLS) in cone-beam CT (CBCT) of the head, focusing on the task of detecting a small, low-contrast intracranial hemorrhage (ICH), and we test the performance of the algorithm in the context of a recently developed CBCT prototype for point-of-care imaging of brain injury. Theoretical predictions of local spatial resolution and noise are computed via an optimization by which regularization (specifically, the quadratic penalty strength) is allowed to vary throughout the image to maximize local task-based detectability index (d'). Simulation studies and testbench experiments were performed using an anthropomorphic head phantom. Three PWLS implementations were tested: conventional (constant) penalty; a certainty-based penalty derived to enforce constant point-spread function, PSF; and the task-based penalty derived to maximize local detectability at each location. Conventional (constant) regularization exhibited a fairly strong degree of spatial variation in d', and the certainty-based method achieved uniform PSF, but each exhibited a reduction in detectability compared to the task-based method, which improved detectability up to similar to 15%. The improvement was strongest in areas of high attenuation (skull base), where the conventional and certainty-based methods tended to over-smooth the data. The task-driven reconstruction method presents a promising regularization method in MBIR by explicitly incorporating task-based imaging performance as the objective. The results demonstrate improved ICH conspicuity and support the development of high-quality CBCT systems.