Novel metrics for assessment of image quality for non-linear computed tomography reconstruction algorithms

Abstract

Background and Purpose: Iterative Reconstruction (IR) algorithms in Computed Tomography (CT) are non-linear, resulting in object-dependent noise and spatial resolution. Traditional image quality metrics measured in mainly uniform phantoms are therefore less relevant in predicting clinical performance. The Structural Similarity metric, SSIM, was assessed in this research project as a candidate for measurement of relative image quality for IR algorithms. Methods: SSIM was evaluated in images of a traditional phantom alongside Contrast-to-noise ratio (CNR) and a task-based contrast signal detectability index d’NPWE for a range of acquisition doses, reconstruction kernels and IR strengths on two commercial CT scanners. Relationships between these metrics were assessed, as well as the potential dose reduction from maintaining constant metric values while employing higher IR strengths. SSIM was then evaluated in an anthropomorphic phantom containing realistic anatomical structures, and evaluated against subjective image scores of anatomical detail clarity from expert observers in order to evaluate the potential of SSIM as a surrogate for human observer studies of image quality. Results: SSIM correlates with traditional image quality metrics when evaluated in traditional phantoms. SSIM is correlated with subjective scores in CT images of the thorax, although this was less significant at higher IR strengths where noise texture and contrast-dependent spatial resolution show greater variation from zero strength IR implementation. SSIM improves with acquisition dose, and with CNR and d’NPWE image quality metrics, reaching a maximum plateau where image quality matches the reference image set. Conclusions: SSIM is a viable image quality metric for routine performance testing of CT scanners, and as an indicator of clinical observer image quality in realistic phantoms. However, observer scoring studies remain an important tool in image quality assessment despite cost and use of time resources.

Date of Award	31 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Marianne Aznar (Supervisor)