Task-based statistical image reconstruction for high-quality cone-beam CT

Hao Dang, Joseph Webster Stayman, Jennifer Xu, Wojciech Zbijewski, Alejandro Sisniega Crespo, Michael Mow, Xiaohui Wang, David H. Foos, Nafi Aygun, Vassilis El Koliatsos, Jeff Siewerdsen

Research output: Contribution to journalArticle

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 (). Simulation studies and test-bench 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 , 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 ∼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.

Original languageEnglish (US)
Pages (from-to)8693-8719
Number of pages27
JournalPhysics in Medicine and Biology
Volume62
Issue number22
DOIs
StatePublished - Nov 1 2017

Fingerprint

Computer-Assisted Image Processing
Cone-Beam Computed Tomography
Intracranial Hemorrhages
Least-Squares Analysis
Point-of-Care Systems
Head
Skull Base
Diagnostic Imaging
Brain Injuries
Noise

Keywords

  • cone-beam computed tomography
  • image quality
  • imaging task
  • intracranial hemorrhage
  • regularization
  • statistical image reconstruction

ASJC Scopus subject areas

  • Radiological and Ultrasound Technology
  • Radiology Nuclear Medicine and imaging

Cite this

Task-based statistical image reconstruction for high-quality cone-beam CT. / Dang, Hao; Stayman, Joseph Webster; Xu, Jennifer; Zbijewski, Wojciech; Sisniega Crespo, Alejandro; Mow, Michael; Wang, Xiaohui; Foos, David H.; Aygun, Nafi; Koliatsos, Vassilis El; Siewerdsen, Jeff.

In: Physics in Medicine and Biology, Vol. 62, No. 22, 01.11.2017, p. 8693-8719.

Research output: Contribution to journalArticle

@article{d49c1908f2f04703be47909c05bf5eca,
title = "Task-based statistical image reconstruction for high-quality cone-beam CT",
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 (). Simulation studies and test-bench 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 , 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 ∼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.",
keywords = "cone-beam computed tomography, image quality, imaging task, intracranial hemorrhage, regularization, statistical image reconstruction",
author = "Hao Dang and Stayman, {Joseph Webster} and Jennifer Xu and Wojciech Zbijewski and {Sisniega Crespo}, Alejandro and Michael Mow and Xiaohui Wang and Foos, {David H.} and Nafi Aygun and Koliatsos, {Vassilis El} and Jeff Siewerdsen",
year = "2017",
month = "11",
day = "1",
doi = "10.1088/1361-6560/aa90fd",
language = "English (US)",
volume = "62",
pages = "8693--8719",
journal = "Physics in Medicine and Biology",
issn = "0031-9155",
publisher = "IOP Publishing Ltd.",
number = "22",

}

TY - JOUR

T1 - Task-based statistical image reconstruction for high-quality cone-beam CT

AU - Dang, Hao

AU - Stayman, Joseph Webster

AU - Xu, Jennifer

AU - Zbijewski, Wojciech

AU - Sisniega Crespo, Alejandro

AU - Mow, Michael

AU - Wang, Xiaohui

AU - Foos, David H.

AU - Aygun, Nafi

AU - Koliatsos, Vassilis El

AU - Siewerdsen, Jeff

PY - 2017/11/1

Y1 - 2017/11/1

N2 - 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 (). Simulation studies and test-bench 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 , 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 ∼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.

AB - 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 (). Simulation studies and test-bench 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 , 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 ∼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.

KW - cone-beam computed tomography

KW - image quality

KW - imaging task

KW - intracranial hemorrhage

KW - regularization

KW - statistical image reconstruction

UR - http://www.scopus.com/inward/record.url?scp=85034234802&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85034234802&partnerID=8YFLogxK

U2 - 10.1088/1361-6560/aa90fd

DO - 10.1088/1361-6560/aa90fd

M3 - Article

C2 - 28976368

AN - SCOPUS:85034234802

VL - 62

SP - 8693

EP - 8719

JO - Physics in Medicine and Biology

JF - Physics in Medicine and Biology

SN - 0031-9155

IS - 22

ER -